News

Latest developments

What changed in local AI hardware, updated daily. GPUs, Macs, AI PCs and laptops, and the small boxes built to run models at home. Each item says what happened, why it matters if you're buying, and links to the original source.

Last updated RSS feed

Microsoft previews RDMA over USB4, letting several Windows PCs run one AI model together

Microsoft announced a developer preview that links two or more USB4-capable Windows PCs with plain USB4 cables so one model can be split across their combined memory and GPU compute. It uses the Windows NetworkDirect RDMA interface and an open-source llama.cpp-based testbed, and is coming soon to Windows Insiders in the Experimental channel. In Microsoft's post, three Ryzen AI Max mini PCs with 128GB each ran Qwen3-235B-A22B at 4-bit at about 18.5 tokens per second.

Why it matters

Models too big for one machine no longer need a server or special network cards: two or three unified-memory mini PCs can pool their memory over USB4. It is an early preview that needs test signing and a registry change, so buy for what one machine runs today, but fast USB4 or Thunderbolt ports and plenty of unified memory now add room to grow.

Source: Microsoft Tech Community (Microsoft USB Blog) techcommunity.microsoft.com
  • ai-pc
  • strix-halo
  • software
  • memory

Tom's Hardware: three used V100 cards run Qwen 3.6 35B at about 100 tokens per second

Tom's Hardware reports on a homelab builder's blog benchmarks of decommissioned NVIDIA Tesla V100 16GB cards on one dual-Xeon board with llama.cpp. Three V100s ran Qwen 3.6 35B at 8-bit at about 101 tokens per second at 4K context and about 73 at 128K. An M5 Max Mac Studio, using separate benchmark data, was slightly faster at 4K and 32K, but the V100s led at 64K and 128K.

Why it matters

Old data-center cards can still run mid-size MoE models fast if you can deal with heat, noise and power: V100s ship without fans, and the builder 3D-printed a blower duct for each card. More cards add memory for bigger models, not more speed, and the Mac and V100 numbers came from different runtimes.

Source: Tom's Hardware tomshardware.com
  • gpu
  • mac

JEDEC previews DDR6, with shorter modules and about four times the baseline bandwidth of DDR5

Tom's Hardware reports that JEDEC released a preview of the DDR6 memory standard. DDR6 starts at 12.8 GT/s, modules are 74mm long versus 133mm for DDR5, and each module splits into four 24-bit sub-channels, for 153.6 GB/s per module at the baseline speed against 38.4 GB/s for DDR5. JEDEC has set no launch date, and Tom's says many expect commercial DDR6 around 2029.

Why it matters

Memory bandwidth sets how fast a model generates tokens when it runs from system RAM, so DDR6 should give CPU and integrated-graphics AI machines a big lift. It is years away, so it is no reason to put off a machine you need now.

Source: Tom's Hardware tomshardware.com
  • memory

AMD and Unsloth bring Ryzen AI NPU support to Unsloth Desktop for local AI agents

AMD said Unsloth Desktop now supports XDNA 2 NPUs on supported Ryzen AI PCs, the first NPUs the app supports. Compatible apps send requests to the Unsloth Desktop OpenAI-compatible API, and Lemonade and FastFlowLM run the model on the NPU. AMD says more than 100 agentic apps and workflows work with it, including Open WebUI, Cline, LangChain and GitHub Copilot.

Why it matters

On a Ryzen AI laptop, an always-on assistant can now run on the low-power NPU and leave the CPU and GPU free for other work. The NPU is aimed at efficient, sustained inference; for large models, memory size still decides what you can run.

Source: AMD Newsroom newsroom.amd.com
  • ai-pc
  • laptop
  • software

RTX Spark compact desktops from ASUS, Dell and MSI go on sale with up to 128GB of unified memory

NVIDIA says RTX Spark mini PCs from ASUS, Dell and MSI are available now, with Acer, GIGABYTE, HP and the Microsoft Surface RTX Spark Dev Box coming later. They use the same platform as the RTX Spark laptops: a 20-core Grace CPU, a Blackwell RTX GPU with 6,144 CUDA cores and up to 128GB of unified memory, plus 10-gigabit Ethernet.

Why it matters

Until now the RTX Spark machines on sale were laptops. A desktop gives Windows and CUDA users a small always-on box with a 128GB memory pool, which NVIDIA says is enough for 120B-parameter models, so it now sits directly against a Mac Studio, a DGX Spark or a Strix Halo mini PC.

Source: NVIDIA Blog blogs.nvidia.com
  • ai-pc
  • rtx-spark
  • memory

Report: Apple is close to a multi-year DRAM supply deal with Samsung

MacRumors, citing Korea's Maeil Business Newspaper, reports that Apple is close to a four- to five-year DRAM supply agreement with Samsung starting in the first half of 2027. The deal would set a minimum supply volume, with terms tied to market conditions, and follows a similar reported agreement with SK hynix.

Why it matters

Unified memory is what decides which models a Mac can run, and the memory shortage has made high-memory configurations harder to get. A guaranteed volume of DRAM could help Apple keep 64GB-plus Macs in stock through 2027, though the report says nothing about what those configurations will sell for.

Source: MacRumors macrumors.com
  • mac
  • memory
  • rumor

NVIDIA's DGX Spark HPC Developer Kit links four Sparks in a switchless ring for 512GB of memory

At the SC26 supercomputing conference, NVIDIA introduced the DGX Spark HPC Developer Kit, a cluster of four DGX Spark systems wired directly to each other over ConnectX-7 with no network switch. Together they offer 512GB of combined memory and 20TB of NVMe storage, and NVIDIA says the kit is on sale now through its OEM partners and NVIDIA Marketplace, with an HPC SDK playbook coming soon.

Why it matters

NVIDIA says four linked Sparks can run models of up to 700B parameters, such as DeepSeek V4.1, in NVFP4 on a desk. It is aimed at researchers, but it shows the supported way to grow a single DGX Spark into a four-node cluster without buying a switch.

Source: NVIDIA Blog blogs.nvidia.com
  • dgx-spark
  • memory

ggml 1.0, the engine under llama.cpp, adds FP8 support and faster MoE paths on CUDA and Metal

The ggml tensor library that powers llama.cpp and many local AI apps reached version 1.0.0. The release adds FP8 support, fused mixture-of-experts work on CUDA and Metal, a Vulkan prefill FlashAttention path and a scheduler that skips unused inputs. It also removes deprecated APIs and bumps the RPC protocol to v9, which rejects older peers.

Why it matters

Faster MoE handling helps the large sparse models people now run on single GPUs, Macs and unified-memory PCs. If you split models across machines with llama.cpp RPC, update every node together, since v9 will not talk to older versions.

Source: ggml release notes (GitHub) github.com
  • software

TrendForce: some PC makers may use lower-spec memory and fewer CPU cores in 2027 models

TrendForce says laptop shipments fell 15.8% from the previous quarter in Q3 2026 as higher memory costs pushed retail prices up, leaving PC makers with growing piles of unsold stock. Some are now cutting memory orders for late 2026 and early 2027 and planning 2027 models with lower-spec memory chips and CPUs with fewer cores. It also expects PC DRAM contract prices to keep rising in Q4, by 13–18%, though more slowly than last quarter.

Why it matters

Running models locally depends on memory, so watch 2027 spec sheets closely: the same model name may come with less or slower RAM than before. If you need 32GB, 64GB or more, check the exact configuration before you buy instead of assuming next year's version will match.

Source: TrendForce trendforce.com
  • memory
  • laptop
  • ai-pc

Modder unlocks the full 80GB of memory on NVIDIA CMP-170HX mining cards with a voltage-glitch hack

Tom's Hardware reports that a modder named Kate used a small microcontroller board to glitch the voltage on CMP-170HX mining cards, getting write access to the GA100 chip's normally locked fuse controller. That let her switch on the HBM2e memory and shader modules disabled at the factory, reaching the full 80GB and PCIe 4.0 speed, beyond what the existing CMPunlocker software mod allows. The fuse changes are permanent.

Why it matters

It shows how much memory is locked away inside cheap used data-center silicon, which is why these cards keep coming up in home AI cluster builds. It is a soldering-iron, no-undo hardware mod, not a buying recommendation; for most people a 24GB to 32GB consumer card or a unified-memory machine is the safer path to big models.

Source: Tom's Hardware tomshardware.com
  • gpu
  • memory

llama.cpp build b12284 adds FP8 support to its CUDA backend for NVIDIA GPUs

llama.cpp's b12284 build, published November 11, adds OCP FP8 support to the CUDA backend, including FP8 matrix math through cuBLAS and FP8 handling for the KV cache. It builds on the FP8 tensor type added in v0.9.0 and the FP8 KV cache that arrived in b12274. The change notes include a test of a 27B model on a 32GB RTX PRO 4500 Blackwell card.

Why it matters

FP8 files need about half the memory of 16-bit ones while keeping more precision than 4-bit quantization, so NVIDIA GPU owners get a middle option between size and quality. As always, check that your GPU and CUDA version are covered before switching formats.

Source: llama.cpp release notes (GitHub) github.com
  • software
  • gpu
  • rtx

Kora, a new Linux desktop OS, will run AI models locally by default and split them across CPU, GPU and NPU

A team led by Cyanogen co-founder Kirt McMaster unveiled Kora, a Linux desktop OS that runs local models by default, falls back to the cloud when needed, and can spread work across a PC's CPU, GPU and NPU or other machines on the same network. Its inference engine is built on llama.cpp and other open runtimes and mainly uses GGUF models. Snapdragon X2 Elite NPUs are supported first, AMD and Intel NPU support is planned but not for language models, and Kora is due on PCs later in 2027.

Why it matters

Kora is still a 2027 product, but it shows where AI PCs are heading: the OS deciding which chip runs a model. Its makers recommend 64GB of memory for better models, and today only Qualcomm's NPU runs language models in it, so GPU and memory still matter most if you're buying now.

Source: Tom's Hardware tomshardware.com
  • software
  • ai-pc
  • memory

HP's ZGX Fury, a deskside GB300 workstation with 748GB of memory, is now listed for order in the US

HotHardware reports that HP's ZGX Fury G1n is now listed at HP and B&H. It's built on NVIDIA's GB300 Grace Blackwell Ultra superchip, pairing a 72-core Grace CPU and 496GB of LPDDR5X with a Blackwell Ultra GPU that has 252GB of HBM3e, for 748GB in total. The liquid-cooled tower ships with Ubuntu 24.04, can be rack-mounted, and has two 400Gbps ports for linking several units.

Why it matters

This is a machine for teams and companies, not a home desk, but it shows where local AI tops out: HP says it can run inference on models up to a trillion parameters at FP4 and fine-tune models over 100B parameters without the cloud. For personal use, 128GB unified-memory boxes remain the practical high end.

Source: HotHardware hothardware.com
  • workstation
  • memory
  • gpu

Report: NVIDIA drops its planned 24GB RTX 5080 and keeps the 16GB card in production

HotHardware relays a post from Board Channels, a Chinese supply-chain forum spotted by Gazlog, saying NVIDIA told its card partners it has cancelled a January 2027 launch of a 24GB RTX 5080 with no new date. The card would have swapped the current 2GB GDDR7 chips for 3GB chips, which the report says are too expensive right now, and the 16GB model is said to be coming back. NVIDIA never announced the 24GB card and has not confirmed the report.

Why it matters

If this holds, 16GB stays the ceiling for the RTX 5080, which is tight for 27B to 32B models with long context. Anyone waiting for a 24GB mid-high card should plan around today's options, such as a used 24GB card, an RTX 5090, or a unified-memory machine.

Source: HotHardware hothardware.com
  • gpu
  • rtx
  • memory
  • rumor

llama.cpp adds FP8 KV-cache support in build b12274

llama.cpp merged support for storing the KV cache, the memory a model uses to hold your conversation and documents, in the OCP FP8 format, with optional per-tensor scales. It shipped in build b12274 on November 10, and the project cites NVIDIA's Mistral-Medium-3.5-128B as a model that uses a scaled FP8 cache.

Why it matters

An 8-bit cache takes about half the memory of a 16-bit one, so long chats and big documents leave more room for the model itself on a GPU or unified-memory machine.

Source: llama.cpp release notes (GitHub) github.com
  • software
  • memory
  • gpu

Apple's Claris previews Harbor, an app that turns Macs into a private local AI server

Claris, an Apple company, has opened an early-access interest page for Harbor, an app that runs open-weight models on Apple silicon Macs and serves them to other apps. It can also link two or more Thunderbolt 5 Macs to pool their unified memory for bigger models. Harbor is in private preview, with no public release date yet.

Why it matters

Pooling memory across Macs has worked through open-source tools, but setup takes real effort. If Harbor makes it plug-and-play, a second Thunderbolt 5 Mac becomes a practical way to run models too big for one machine's unified memory.

Source: MacRumors macrumors.com
  • mac
  • memory
  • software

Tom's Hardware: share of 64GB PCs drops 18% in CPU-Z data as memory costs climb

Tom's Hardware reports that CPU-Z's user data shows 64GB systems down 18% year over year, now 7.3% of machines, while 32GB slipped 2.5% and 16GB and 24GB setups grew. DDR4 also gained share over DDR5 as buyers stick with older platforms. The data reflects CPU-Z's user base, not retail sales.

Why it matters

Running local models leans on lots of RAM, and the DRAM squeeze is pushing typical builds the other way. If you need 64GB or more for CPU offload or big models, plan for it early, or look at unified-memory machines where memory comes with the system.

Source: Tom's Hardware tomshardware.com
  • memory

Tom's Hardware: a lawyer runs DeepSeek V4.1 Flash on eight used 2018-era V100 GPUs to keep case files offline

Tom's Hardware covered a homelab build that uses eight used Tesla V100 32GB data center GPUs on two NVLink boards, for 256GB of VRAM, to run a mostly 2-bit DeepSeek V4.1 Flash locally. The builder reports up to 215 tokens per second on code and up to 118 on legal writing, with the cards power-capped to 150 W each. Tom's notes used V100 prices have more than doubled and that the R580 driver branch is the last to support these cards.

Why it matters

Old data center GPUs can be a way to get a lot of fast VRAM in one machine, but they bring real tradeoffs: server-style boards and cabling, heavy power draw, a driver branch that's ending, and used prices that are climbing as others catch on.

Source: Tom's Hardware tomshardware.com
  • gpu
  • memory

Ollama 0.40.16 cuts Qwen3.8 Flash Next memory use on Macs and adds CUDA 13.5 for Windows on Arm

Ollama's v0.40.16 release, published November 9, says Qwen3.8 Flash Next now uses about 27 GiB less memory on Apple silicon. It also adds CUDA 13.5 support for Windows on Arm, updates its bundled llama.cpp, and adds Ripple Preview, a 65B dense model built for agentic coding.

Why it matters

Saving about 27 GiB can be the difference between a model fitting on your Mac or not. CUDA 13.5 on Windows on Arm matters for Arm-based NVIDIA machines like RTX Spark laptops running Ollama on Windows.

Source: Ollama release notes (GitHub) github.com
  • software
  • mac
  • rtx-spark

Tom's Hardware: M5 Ultra Mac Studio beats one DGX Spark, but two Sparks win on the biggest models

In new local AI tests, Tom's Hardware found a 256GB M5 Ultra Mac Studio finished an agentic benchmark with Qwen 3.8 27B in about half the time of a single GB10 (DGX Spark-class) box, while using about 16% less energy. With large models like DeepSeek V4.1 Flash and GLM 5.3 Flash, a pair of clustered Sparks processed long prompts much faster, partly because faster inference engines and serving recipes exist for NVIDIA's platform.

Why it matters

Memory bandwidth alone doesn't decide speed: which inference engine supports your model on your hardware matters just as much. A Mac Studio is strong and efficient for mid-size models today, while a Spark cluster has broader software support for the largest open models. Tom's Hardware also notes Mac Studio lead times now stretch into February 2027.

Source: Tom's Hardware tomshardware.com
  • mac
  • dgx-spark
  • memory
  • software

GIGABYTE launches AI TOP AERO, a compact RTX Spark desktop with up to 128GB of unified memory

GIGABYTE announced AI TOP AERO, a small desktop built on NVIDIA's RTX Spark N1X Superchip, which pairs a 20-core Grace CPU with a Blackwell RTX GPU. It comes with up to 128GB of LPDDR5x unified memory and up to a 4TB PCIe Gen5 SSD, runs Windows 11, and supports tools like LM Studio and Ollama. GIGABYTE says it launches November 10, with configurations varying by region.

Why it matters

RTX Spark is no longer laptop-only: desktops like this give you the same 128GB unified memory pool for large local models in a box that sits on your desk. MSI also detailed a similar RTX Spark mini PC, the EdgeMesa N AI+, the same day.

Source: GIGABYTE via PR Newswire prnewswire.com
  • rtx-spark
  • ai-pc
  • memory

AMD's Ryzen AI Halo now comes with 192GB of unified memory and Microsoft's ready-to-code Windows setup

AMD says a new Ryzen AI Halo developer box with the Ryzen AI Max+ PRO 495 is available with Microsoft's Project Zenith preinstalled, a Windows 11 setup that ships with common developer tools and WSL already configured. The machine has up to 192GB of unified memory, up to 160GB of which the GPU can use, with 273 GB/s of bandwidth. AMD also includes guided playbooks for running local LLMs, ComfyUI image generation and LM Studio.

Why it matters

Up to 160GB of GPU-addressable memory lets a Strix Halo-class machine load much larger models than the 128GB version, though bandwidth stays at 273 GB/s, so big models will still generate tokens slowly. The preinstalled setup also cuts the time from unboxing to running a local model on Windows.

Source: AMD Blog amd.com
  • strix-halo
  • ai-pc
  • memory
  • software

RTX 5090 gets harder to find, as HP and Lenovo show no desktops with it in stock

TechRadar found no readily orderable RTX 5090 desktop configurations on HP's or Lenovo's US stores, and several major retailers list the standalone card as out of stock. The cards still listed are marked far above NVIDIA's launch level, while a few smaller system builders and Dell's Alienware line still offer RTX 5090 desktops.

Why it matters

The RTX 5090's 32GB of VRAM is the most you can get on a single consumer card, so its scarcity squeezes anyone planning a one-GPU local AI build around it. If you can't find one at a sane level, a 128GB unified-memory machine or a pair of smaller cards may get you more model capacity for the money right now.

Source: TechRadar techradar.com
  • gpu
  • rtx
  • availability

NVIDIA says RTX Spark doubles laptop performance per watt and keeps 98% of its speed on battery

At a press event in Seoul on Nov. 6, NVIDIA said RTX Spark delivers about twice the performance of a Core Ultra 9 laptop with a discrete RTX 5070 at 60 watts, based on 20 games and LLM prompt processing. It also said RTX Spark keeps about 98% of its plugged-in performance on battery. These are NVIDIA's own numbers, reported by The Elec.

Why it matters

If NVIDIA's battery claim holds up in independent tests, an RTX Spark laptop can run local models unplugged without a big slowdown, which most discrete-GPU laptops can't. Treat it as a vendor claim for now; independent testing has already found platform quirks, like a large slowdown on Linux when plugged in.

Source: The Elec thelec.net
  • rtx-spark
  • laptop
  • ai-pc

Free GPU Taskman tool shows which apps and AI models are using your VRAM

Microsoft VP Scott Hanselman released gputaskman, an open-source, MIT-licensed GPU task manager for Windows, Linux and macOS. Tom's Hardware reports it lists per-process GPU load and VRAM use, and for local LLM servers like llama-server it breaks out the model, its weights and its KV cache. It's labeled experimental, with portable Windows downloads of v0.3.0.

Why it matters

When a model won't fit, the first question is where your VRAM already went. A tool that shows the desktop, the browser and each model's weights and KV cache side by side makes it much easier to judge whether you need a bigger card or just need to close something.

Source: Tom's Hardware tomshardware.com
  • software
  • gpu
  • memory

ASUS announces XG Core, a pocket-size USB4 graphics dock with a 16GB Radeon RX 9060 XT

ASUS's XG Core packs a low-profile Radeon RX 9060 XT with 16GB of GDDR6 into a small external box that plugs into laptops, mini PCs and handhelds over USB4 or Thunderbolt 3. ASUS says it supports local image generation and LLM inference in apps like ComfyUI, LM Studio and Ollama. Pricing and release dates weren't announced.

Why it matters

It's a way to add 16GB of dedicated VRAM to a thin laptop or mini PC without replacing it, enough for many quantized 8B to 14B models. The USB4 link tops out at 40Gbps, so loading models is slower than in a desktop slot, but generation speed mostly depends on the card once the model is in VRAM.

Source: ASUS Pressroom press.asus.com
  • gpu
  • radeon
  • egpu
  • laptop

AMD confirms a Radeon driver bug that can keep writing to your SSD while the PC sits idle

Radeon owners report that AMD's External Events Utility, installed with the Adrenalin drivers, writes to the system drive nonstop, around 5MB/s at idle in one report. AMD has acknowledged the problem and says it hopes to fix it in the next driver release. Tom's Hardware notes that how much of this turns into real flash wear is still unknown.

Why it matters

Local AI boxes lean on big, fast SSDs to hold model files, so a background process quietly using up drive endurance is worth checking for. On a Radeon system, open Resource Monitor's Disk tab and look for steady writes from atieclxx.exe or atiesrxx.exe until AMD ships the fix.

Source: Tom's Hardware tomshardware.com
  • gpu
  • radeon
  • software
  • storage

XDA plans a local AI desktop around a 16GB Radeon instead of chasing a scarce high-end GeForce

XDA used its PC builder to assemble a local AI workstation aimed at models that fit a 16GB card, landing on an RX 9070 XT with 640 GB/s of memory bandwidth plus 32GB of system RAM. The write-up notes that ROCm, llama.cpp, Ollama, LM Studio, vLLM and ComfyUI list the card as supported, while many workflows still favor CUDA first, and that Vulkan on RDNA 4 often beats ROCm for token generation. It also shows how to swap in a used NVIDIA card or change parts inside the planner.

Why it matters

With top GeForce cards hard to find, a 16GB AMD card with strong bandwidth is a practical path for mid-size local models if you accept a CUDA-first software world. The piece is also a reminder that system RAM now dominates build cost, so set a hard budget before you start shopping.

Source: XDA xda-developers.com
  • gpu
  • software
  • build

The Verge: buy the new M6 MacBook Pro for everyday work, not for serious local AI

In its November 5 review, The Verge calls the 14-inch M6 MacBook Pro a solid chip bump of Apple's entry Pro, with the same chassis as recent years and enough power for everyday creative work. For local language models, high-res video, or 3D, the review says go straight to an M5 Pro or M5 Max with as much unified memory as you can get. Both the M6 model and the redesigned OLED M5 Pro and M5 Max machines become available November 6.

Why it matters

Launch day makes it easy to grab the newest badge. For local AI the memory ceiling still matters more: the M6 tops out at 32GB, while M5 Max still reaches 128GB and far higher bandwidth, so pick the chip for the models you want to run, not the launch date.

Source: The Verge theverge.com
  • mac
  • laptop
  • memory

RTX PRO 5500 Blackwell with 84GB GDDR7 shows up at a US retailer, still listed as coming soon by NVIDIA

VideoCardz reports a PNY RTX PRO 5500 Blackwell Workstation Edition listing at a US IT reseller, matching NVIDIA's own page for an 84GB GDDR7 ECC card with about 1,398 GB/s of bandwidth and Multi-Instance GPU into two 42GB slices. NVIDIA still marks the product as available through OEM and distribution partners and has not posted its own store price. Large enterprise orders reported earlier point to commercial deliveries in early 2027.

Why it matters

84GB on one discrete card is a rare step up from consumer 24–32GB GeForce cards for local models that no longer fit in mid-range VRAM. If you need CUDA workstation memory without a multi-GPU split, this SKU is worth watching once stock actually lands, and MIG lets two users or jobs share one card.

Source: VideoCardz videocardz.com
  • gpu
  • rtx
  • workstation
  • memory

CUDA Toolkit 13.5 brings Vera Rubin to general availability and better multi-workload GPU sharing

NVIDIA's CUDA 13.5 release notes and developer blog put Vera Rubin GPU support into general availability and extend Windows on Arm support to DGX Station for Windows. The toolkit adds Multi-Process Service v3 with weighted timeslicing and SM partition borrowing, multi-GPU CUDA Tile communication over NVLink, and CCCL's batched Top-K path aimed at inference. New platform features need an R620-branch driver, and the toolkit no longer bundles the driver.

Why it matters

If you run local inference or fine-tuning on NVIDIA hardware, this is the stack to target for Rubin cards and for sharing one GPU across several jobs more cleanly. Batched Top-K and multi-GPU Tile matter most once you outgrow a single card; everyday GeForce users mainly need a matching R620 driver when they upgrade tools that depend on 13.5.

Source: NVIDIA Technical Blog developer.nvidia.com
  • software
  • gpu
  • cuda

Team Group's CEO says the memory shortage will worsen through 2027 and not balance until late 2028

On the company's Nov. 2 earnings call, Team Group CEO Chen Ching-wen said the memory shortage should get worse from the end of 2026 through the end of 2027, with supply and demand not balancing until the second half of 2028 or later. He said spot prices run 30% to 50% above contract prices and that budget parts won't ease until 2029, per Tom's Hardware.

Why it matters

If a local AI build or upgrade needs RAM, don't count on prices falling soon. Buying the memory you need now beats waiting, and this is one company's forecast, not a guarantee.

Source: Tom's Hardware tomshardware.com
  • memory
  • dram
  • shortage
  • availability

Test finds an RTX Spark laptop runs AI nearly 79% slower on Linux when plugged in

MindStudio tested an MSI Prestige N16 Flip AI+ with NVIDIA's RTX Spark and measured 65.4 tokens per second on Windows with GPT-OSS-120B, but only 13.9 on Linux while plugged in. The GPU ran near 780 MHz instead of about 2.2 GHz, which MindStudio traced to missing Linux support for an NVIDIA power-management component. Linux on battery and Windows Subsystem for Linux both came close to Windows speeds.

Why it matters

If you want an RTX Spark laptop for Linux-based AI work, plan on Windows with WSL for now. This is one laptop, not yet independently replicated, and NVIDIA hasn't commented.

Source: Tom's Guide tomsguide.com
  • rtx-spark
  • laptop
  • linux
  • benchmark

BOSGAME opens pre-orders for a 192GB Ryzen AI Max+ PRO 495 mini workstation

BOSGAME's M5 Max pairs AMD's Ryzen AI Max+ PRO 495 with 192GB of soldered unified memory and a 2TB PCIe 4.0 SSD, with dual 10GbE, two USB4 ports and an OCuLink port for an external GPU. Pre-orders opened Nov. 3 as a limited first batch, and BOSGAME advertises support for models up to 320 billion parameters.

Why it matters

It is another way to get 192GB of memory in a small box for running large local models. BOSGAME lists the memory as LPDDR5 and hasn't confirmed its speed, which matters for token rates, so wait for real benchmarks.

Source: VideoCardz videocardz.com
  • amd
  • strix-halo
  • mini-pc
  • memory

AMD claims its Ryzen AI chips beat RTX Spark on CPU speed and gaming battery life

At its Advancing AI Seoul event on Nov. 4, AMD said its Ryzen AI part scored 14% higher than RTX Spark in a multi-core CPU test and lasted nearly 50% longer in a gaming battery test, according to The Elec. Partner Nota AI also showed Strix Halo optimizations that let the GPU and NPU share one model in memory and cut LLM inference time by up to 30%.

Why it matters

These are AMD's own numbers, AMD didn't name the chip it tested and it skipped graphics performance, so treat them as claims until independent reviews confirm them. The Nota AI work shows Strix Halo software still has room to improve.

Source: The Elec thelec.net
  • amd
  • strix-halo
  • rtx-spark
  • vendor-claims

PCWorld's RTX Spark tests: a 128GB laptop beats an RTX 5090 desktop on a 120B model that outgrows 32GB of VRAM

PCWorld ran a frozen gpt-oss 120B research workload on an ASUS ProArt P16 with 128GB of unified memory, about 96GB of it given to the GPU, and finished in 33 minutes versus almost 53 on a GeForce RTX 5090 desktop with 32GB of VRAM. PCWorld stresses the laptop's GPU is not faster and that models fitting in 32GB favor the 5090. It also reports heavy battery drain and loud fans under sustained AI loads.

Why it matters

Once a model outgrows VRAM, how much memory the GPU can reach matters more than raw GPU speed. If your models fit in 32GB, a fast discrete GPU still wins; if you want 100B-class models on a portable machine, unified memory is the point, with battery, noise and Windows-on-Arm software checks as the catch.

Source: PCWorld pcworld.com
  • rtx-spark
  • laptop
  • memory
  • gpu

llama.cpp v0.9.0 adds FP8 quantization, faster Gemma 4 small-model prompts and multi-node tensor parallelism

The v0.9.0 release notes list a new OCP FP8 E4M3 tensor type, a flash-attention change that makes the unified KV cache efficient for multi-sequence use, and large prompt-processing gains for Gemma 4 E2B and E4B. The bundled ggml update adds initial NCCL-based multi-node tensor parallelism and fused mixture-of-experts kernels for CUDA. New model support includes Mistral Large 4, and NVFP4 files now work for LFM3.

Why it matters

If you run llama.cpp directly, this is a worthwhile update for Gemma 4 small-model speed and for serving several chats at once. Multi-node support is early, but it points toward linking more than one machine for models too big for a single box. Check the release notes for your own backend before upgrading.

Source: llama.cpp Releases (GitHub) github.com
  • software
  • gpu
  • memory

Apple tells investors memory costs are "unprecedented" and will keep climbing beyond this quarter

On Apple's November 2 earnings call, John Ternus called current memory pricing unprecedented and said Apple expects even higher memory costs in the December quarter. He expects memory purchase prices to keep rising after that as the benefit of earlier-bought inventory fades. Tom's Hardware notes Apple has not announced another round of price changes.

Why it matters

Unified memory is what Macs offer for local AI, so memory costs land directly on the high-memory configurations people want. Apple's own earlier price changes were partly a response to this, and supplier warnings point the same way, so don't plan around Mac memory options getting cheaper or easier to find soon.

Source: Tom's Hardware tomshardware.com
  • mac
  • memory

A new USB boot tool brings the RTX 20, 30 and 40 AI throttle unlock to Windows

Tom's Hardware reports on an open-source tool that runs from a USB stick before Windows loads and removes NVIDIA's half-speed tensor limit on RTX 20, 30 and 40-series desktop and laptop cards, with up to 65% higher FP16 math on the developer's own RTX 4060 laptop. It has to be run again on every boot, RTX 50 cards aren't supported, and RTX 2080 and 2080 Ti cards fail in testing.

Why it matters

This follows the Linux-first unlocks we covered in October and makes them easier to try on Windows, but it's still an unofficial workaround that needs Secure Boot off or a self-signed boot file, and gains depend on power and thermal limits. Treat it as a bonus for a card you already own, not a reason to buy one.

Source: Tom's Hardware tomshardware.com
  • gpu
  • rtx
  • software

Ollama 0.40.14 stops overriding a model's own sampler settings on OpenAI-style APIs

Ollama's official release notes for 0.40.14 say its OpenAI-compatible endpoints no longer force a temperature and top_p of 1.0 when a client sends none, so the model's own packaged defaults apply, matching its native chat API. The release also adds faster top-k sampling on MLX for models with large vocabularies, which is the Apple silicon path.

Why it matters

If you point a coding tool or chat app at Ollama's OpenAI-style endpoint, replies may now behave the way the model's maker intended without any settings changes. Mac owners also get a small speed gain from the MLX sampling change, though this is a software fix and doesn't change what hardware you need.

Source: Ollama release notes github.com
  • software
  • mac

Creators report Amazon cancelling DGX Spark and GPU orders, one account closed

Tom's Hardware reports that AI creator Nick Saraev says Amazon closed his account after he tried to buy seven DGX Spark systems across four orders, with the closure email giving no reason. Epic CEO Tim Sweeney says his regular orders were cancelled just before arrival after he tried to buy two NVIDIA GPUs. Amazon hasn't explained itself, and the DGX Spark is still listed for sale there.

Why it matters

The reported cases involve multi-unit or repeat orders, not a single purchase, but they show a big retailer can cancel AI hardware orders without explanation. If you're buying several machines or GPUs, expect possible cancellations, buy in smaller batches, and keep a second seller in mind.

Source: Tom's Hardware tomshardware.com
  • dgx-spark
  • gpu
  • availability

XDA runs a 125B model on a 128GB RTX Spark laptop for almost six hours to build its own AI app

XDA ran Qwen 3.8 Flash Next, a 125B mixture-of-experts model, through llama.cpp on a 128GB RTX Spark laptop at around 45 to 50 tokens per second, and gave it one prompt to build Copilot+ style features. Working unattended for almost six hours, it produced a working app with screen search, live captions and file search. Long prompt reprocessing near a 262K context made some single replies take several minutes.

Why it matters

It's a real-world test of what 128GB of unified memory buys you on a laptop: a model this size can run long agent jobs locally. It also shows that prompt processing, not just generation speed, decides how long big-context agent work takes.

Source: XDA xda-developers.com
  • laptop
  • rtx-spark
  • software

PC memory contract prices hit a new record in October, TrendForce says

TrendForce data reported by Chosun Biz shows mainstream PC DRAM contract prices set another record in October, with DDR5 module prices for PC makers up by double digits from September. TrendForce expects Korean makers to produce more than 20% less PC DRAM next year as they shift lines to AI server memory, and still sees fourth-quarter PC DRAM contracts rising.

Why it matters

Memory is now the biggest swing factor in what a local AI machine with 32GB, 64GB or 128GB costs. With output for PCs set to shrink in 2027, high-memory configurations are unlikely to get easier to find soon.

Source: Chosun Biz biz.chosun.com
  • memory

NVIDIA's own store drops the RTX 5080 listing, weeks after the RTX 5090

Tom's Hardware reports that the GeForce RTX 5080 listing and its filters have disappeared from NVIDIA's online marketplace, following the RTX 5090. The RTX 5070 Ti and RTX 5070 are still listed, marked out of stock, and remaining RTX 5080 stock at US retailers is thinning. NVIDIA hasn't said whether the card is discontinued, and Tom's Hardware has asked for comment.

Why it matters

The RTX 5080's 16GB was already a mid-size option for local models, and now both top GeForce cards are hard to find. If you need more VRAM soon, a 32GB workstation card, a used card or a unified-memory machine may be the easier path than waiting for restocks.

Source: Tom's Hardware tomshardware.com
  • gpu
  • rtx
  • availability

Intel expects memory costs to slow PC sales through at least mid-2027

On its third-quarter call, Intel said memory prices and availability are cutting PC demand and that it expects the slowdown to last at least through the first half of 2027. It is also converting some Intel 18A capacity in Oregon, which mostly builds Panther Lake laptop chips, to make more Xeon server processors, and said demand for its low-cost Wildcat Lake chips is outrunning supply.

Why it matters

This is another big supplier saying PC memory stays tight into next year. Fewer laptop chips plus expensive memory means high-memory laptops for local AI are likely to stay scarce, so if a configuration with enough RAM fits your needs now, waiting is unlikely to make it easier to find.

Source: TechPowerUp techpowerup.com
  • memory
  • ai-pc
  • laptop
  • intel

Samsung expects tight memory supply through 2028, with its 2027 HBM already sold out

On its quarterly earnings call, Samsung's memory business said tight supply conditions will continue into 2028, driven by AI demand. The company also said its HBM supply for 2027 is already sold out.

Why it matters

This is the second big memory maker this week to say the squeeze will last for years. GPUs with lots of VRAM, high-capacity RAM kits and unified-memory machines are likely to stay hard to find, so waiting for memory to loosen up isn't a plan.

Source: The Verge theverge.com
  • memory

XDA finds RTX Spark laptops run the same chip as DGX Spark, but with faster memory

XDA compared an HP OmniBook Ultra 16 with RTX Spark against a DGX Spark and found the two report the same chip, CPU cores and GPU. The laptop runs its 128GB of LPDDR5X at 9,400 MT/s versus 8,533 MT/s on the DGX Spark, and measured about 18% more GPU read bandwidth. It also enables USB4 PCIe tunneling and an NPU that GB10 boards leave switched off.

Why it matters

Token speed on these machines is mostly limited by memory bandwidth, so the RTX Spark laptop can be a little quicker at generating text than a DGX Spark with the same memory. The DGX Spark still has ConnectX-7 clustering and Linux, where the GPU can use the whole memory pool, which matters more for serving and fine-tuning.

Source: XDA xda-developers.com
  • rtx-spark
  • dgx-spark
  • laptop
  • memory

AI demand is now causing a circuit board shortage for PC makers

Citing a Nikkei Asia report, Tom's Hardware says PC makers are struggling to secure printed circuit boards because AI hardware is soaking up capacity. Some PCB orders now take up to 30 months to fill, and a major supplier of board laminates says its order book is filling up to 2030.

Why it matters

Every GPU, motherboard and laptop needs these boards, so this adds another squeeze on top of the memory shortage. Expect PCs and graphics cards to stay tight on supply into next year.

Source: Tom's Hardware tomshardware.com
  • memory
  • gpu
  • ai-pc

Ollama 0.40.13 brings Liquid AI's LFM3 models to Windows and Linux

Ollama 0.40.13 lets Liquid AI's LFM3 models, including text-to-speech in lfm3-audio and the lfm3-omni model, run on Windows and Linux as well as Apple silicon. The previous release had them on Macs only.

Why it matters

LFM3 models are small enough to run on almost any recent PC, so you can try local speech and multimodal assistants on an existing GPU or AI laptop without buying anything new.

Source: Ollama release notes github.com
  • software

SK hynix expects the memory shortage to last through 2030

In its third-quarter results, SK hynix said demand for memory will keep outrunning supply through 2030, because leading-edge processes are hard and new fabs take years to build. It has long-term agreements covering more than half of its 2027 and 2028 production, mostly with large cloud customers, and expects HBM4 to become its main product this quarter.

Why it matters

With much of future DRAM output already committed to data centers, high-memory GPUs, RAM kits and unified-memory machines are likely to stay tight for a long time. If you've been waiting for memory to get easier to find before building a local AI box, don't count on it soon.

Source: SK hynix Newsroom news.skhynix.com
  • memory

RTX Spark laptops get Nemotron 3.5 Transcribe, LFM3 support and upcoming GPU sharing controls for agents

NVIDIA said its Nemotron 3.5 Transcribe open-weight speech model, which handles transcription, speaker attribution and translation, is available on RTX Spark laptops, along with day-zero support for Liquid AI's LFM3 models. It also listed upcoming controls in the NVIDIA App for splitting GPU resources between foreground apps and background agents.

Why it matters

Background agents and your everyday apps fight over the same unified memory and GPU on these laptops. Controls for splitting resources between them would make an RTX Spark machine more practical as an always-on local agent box, and LFM3 runs on Windows here even though Ollama's build is Mac-only for now.

Source: NVIDIA Blog blogs.nvidia.com
  • ai-pc
  • laptop
  • rtx-spark
  • software

Apple's redesigned MacBook Pro adds an OLED touchscreen and keeps M5 Pro and M5 Max with up to 128GB

Apple announced a redesigned 14- and 16-inch MacBook Pro with a tandem OLED display it calls PROLED XDR, plus touch input on a Mac for the first time. It uses the M5 Pro and M5 Max chips rather than M6, with up to 128GB of unified memory and up to 16TB of storage. Pre-orders opened October 28, and it arrives November 6.

Why it matters

This is now the MacBook Pro to look at for bigger local models: 128GB of unified memory and up to 614GB/s of bandwidth on M5 Max, versus a 32GB ceiling on the new M6 14-inch. The chips are the same as the previous M5 Pro and M5 Max models, so the new design and screen don't make models run faster.

Source: Apple Newsroom apple.com
  • mac
  • laptop
  • memory

Only the 24GB and 32GB M6 MacBook Pro configurations get the faster 170GB/s memory

Apple's spec page for the new 14-inch MacBook Pro with M6 lists the base 16GB configuration at 153GB/s of memory bandwidth, the same as the M5 model. Only the 24GB and 32GB options get the full 170GB/s, MacRumors reports.

Why it matters

Memory bandwidth sets how fast a Mac generates tokens, and capacity sets which models fit. If you're buying the M6 MacBook Pro for local AI, skip the 16GB base model: 24GB or 32GB gets you both more room and the faster memory.

Source: MacRumors macrumors.com
  • mac
  • laptop
  • memory

llama.cpp v0.8.0 adds LFM3, DeepSeek V4.1 Flash and a cache for MoE experts kept in system RAM

The llama.cpp v0.8.0 release adds Liquid AI's LFM3 text, vision, audio and omni models, DeepSeek V4.1 Flash, Flint and several other architectures. It also adds a cache for mixture-of-experts layers held in system memory, new realtime-audio and text-to-speech server endpoints, and CUDA, SYCL and OpenCL performance work.

Why it matters

Many PC owners run large mixture-of-experts models by keeping some experts in system RAM when they don't fit in VRAM, and this release targets exactly that setup. It also means LFM3 now runs on Windows and Linux through llama.cpp directly, ahead of Ollama's PC builds.

Source: llama.cpp Releases (GitHub) github.com
  • software
  • gpu
  • memory

Ollama 0.40.12 brings Liquid AI's LFM3 small models to Macs first

The final Ollama 0.40.12 release adds LFM3, Liquid AI's family of small models for local agents and tool calling. It includes a 1.7B text model with a 256K-token context, plus vision, audio (speech recognition, translation and text-to-speech in 18 languages) and omni versions. Liquid's d1 decision models also arrive in a 2B size. In this release both run on Apple hardware only, and Ollama says Linux and Windows support is coming soon.

Why it matters

Models this small run comfortably on 16GB machines, so even an entry-level Mac can handle local agents, voice and structured-output tasks without a cloud service. The release also ships the faster Apple-silicon prompt processing first seen in Monday's pre-release. PC users will need to wait for the Linux and Windows builds.

Source: Ollama Releases (GitHub) github.com
  • software
  • ollama
  • mac

Apple's 14-inch MacBook Pro moves to M6, with up to 32GB of unified memory and 170GB/s

Apple announced a 14-inch MacBook Pro with the new M6 chip: a 12-core CPU, a 12-core GPU with Neural Accelerators in each core, and a dual 16-core Neural Engine. It starts at 16GB of unified memory, can be configured up to 32GB, and has up to 170GB/s of memory bandwidth. Pre-orders open October 28 and it arrives November 6. Apple's redesigned 14- and 16-inch MacBook Pro, announced alongside it, uses M5 Pro and M5 Max instead.

Why it matters

Apple claims up to 25% faster AI workloads and up to 1.3x faster time-to-first-token for LLMs compared with M5. Even so, 32GB is the ceiling, which suits small and mid-size models but not 70B-class models. If you need more memory, look at the redesigned MacBook Pro with M5 Pro or M5 Max (up to 128GB) or a desktop Mac.

Source: Apple Newsroom apple.com
  • mac
  • laptop
  • memory

Apple's new iMac gets M6 and 170GB/s of memory bandwidth for on-device AI

Apple refreshed the 24-inch iMac with the M6 chip, with up to a 12-core CPU and GPU, Neural Accelerators in each GPU core, a dual 16-core Neural Engine and up to 170GB/s of memory bandwidth. Apple's spec page lists 16GB of unified memory standard, with up to 32GB on the 12-core model. Pre-orders open October 28, with availability from November 6.

Why it matters

The iMac is now a reasonable all-in-one for small local models and coding assistants, with more bandwidth than the M4 model's 120GB/s. Its 32GB limit makes it a starter machine for local AI, not a replacement for a Mac mini or Mac Studio with more memory.

Source: Apple Newsroom apple.com
  • mac
  • memory

Palit denies reports that it closed the Galax GPU factory over NVIDIA chip shortages

Palit issued a statement saying claims that it shut the Galax factory because it couldn't get enough NVIDIA GPU chips are inaccurate. It says Galax and Palit share factories, production continues as normal, and Galax RTX 50-series cards remain on sale through existing distributors and retailers. The rumor came from a Spanish outlet's report.

Why it matters

Galax cards stay on the market for now, but the episode shows how tight supply of high-end GeForce cards like the RTX 5090 remains. If you plan a local AI build around a top-end card, expect limited stock and buy when you find the model you need.

Source: Tom's Hardware tomshardware.com
  • gpu
  • rtx
  • nvidia
  • availability

XDA tests NVIDIA's RTX 5070 claim for RTX Spark and finds it holds only with frame generation

XDA put a 64GB RTX Spark laptop against a desktop RTX 5070 with the same 6,144 CUDA cores. In GPU-limited tests, four of five games ran at 51 to 62% of the desktop's frame rate, and the laptop reached about 80 to 84% of laptop RTX 5070 results from PC Gamer's tests. CPU-limited games under Windows on Arm emulation varied widely by engine.

Why it matters

The test also highlights memory bandwidth: 300GB/s on RTX Spark against 672GB/s on the desktop card. Bandwidth sets how fast a model writes tokens, so RTX Spark wins on how big a model fits, while a discrete GPU wins on speed for models that fit in its VRAM.

Source: XDA Developers xda-developers.com
  • laptop
  • rtx-spark
  • review
  • gpu

Ollama 0.40.11 adds speech output on Macs and small decision models

Ollama 0.40.11, released late Friday, can now speak text or a chat model's reply aloud on macOS as it generates, and apps can stream that audio through an OpenAI-compatible endpoint. It also adds d3, a family of 2B to 27B models that classify, answer yes or no, or score options. A 0.40.12 pre-release posted Monday updates MLX with faster prompt processing for Gemma, Flint and Laguna models.

Why it matters

Voice output on the same Mac that runs your model means a fully local voice assistant no longer needs a second app or a cloud service. The small d3 sizes run comfortably on 16GB machines.

Source: Ollama on GitHub github.com
  • software
  • ollama
  • mac

ASUS's 32GB Arc Pro B70 goes on sale in Europe, with no US date yet

ASUS has started selling its UGen Intel Arc Pro B70 in Germany, Austria and Switzerland, Digital Trends reports. The card pairs 32GB of ECC GDDR6 with 32 Xe2 cores, uses a dual-slot blower cooler and is rated at 310W, up from 230W on Intel's own version. ASUS hasn't said when it will reach the US.

Why it matters

It's another way to get 32GB of VRAM on one card while NVIDIA's high-memory cards are scarce. Digital Trends notes it sits in the same price tier as an RTX 5090, so it only makes sense if Intel's software stack runs your models well.

Source: Digital Trends digitaltrends.com
  • gpu
  • intel
  • arc
  • workstation

Unofficial firmware unlock lifts a half-rate AI limit on some RTX 20, 30 and 40 cards

Tom's Hardware reports that hobbyist projects use a flaw in NVIDIA's signed GPU firmware to remove a limit that runs some tensor math at half rate on GeForce RTX 20, 30 and 40 cards. Reported matrix-math gains range from about 12% to 94% depending on card and power limit, and the RTX 4090's peak tensor rate doubles on paper. Support is mostly Linux, many cards are untested, RTX 50 cards aren't covered, and the developers warn it can hang a system or damage hardware.

Why it matters

The gains land in prompt processing and image generation, not in chatbot reply speed, which still depends on memory bandwidth. Don't pay extra for a used RTX 3090 or 4090 because of this; VRAM is still what decides which models a card can run.

Source: Tom's Hardware tomshardware.com
  • gpu
  • rtx
  • nvidia
  • software

HP OmniBook Ultra 16 review: 32GB of memory holds RTX Spark back for local AI

How-To Geek tested the 32GB HP OmniBook Ultra 16 with NVIDIA's RTX Spark chip and found it strong for gaming and creative work. With only about 20GB usable for models, it couldn't reliably run a 35B mixture-of-experts model with the 64K context a local agent needed. Smaller models ran well, with Qwen 3.5 9B at about 46 tokens per second and Qwen 3.5 27B at about 15.

Why it matters

On RTX Spark, as on Macs, memory is fixed at purchase and decides which models fit. If running local agents or 30B-class models is part of the plan, the reviewer's advice is to skip the 24GB and 32GB versions and get 64GB or 128GB.

Source: How-To Geek howtogeek.com
  • laptop
  • rtx-spark
  • review
  • memory

Ollama 0.40.9 brings Flint to Windows and Linux

Ollama's 0.40.9 release runs the Recursive Flint coding model on Linux and Windows, not just macOS. It also makes streaming OpenAI Responses and Anthropic Messages requests report generation failures instead of ending silently after partial output.

Why it matters

Windows and Linux users can now run Flint with one Ollama command instead of setting up llama.cpp by hand.

Source: Ollama on GitHub github.com
  • software
  • ollama
  • gpu

LattePanda Sigma 2 puts Intel Core Ultra X7 and up to 64GB of memory on a single board

LattePanda launched the Sigma 2, an x86 single-board computer built on Intel's Core Ultra X7 358H with Arc B390 graphics and up to 64GB of LPDDR5X memory. The company says up to 59GB can be given to the GPU for AI work, and that it runs the Qwen3.6-35B-A3B model at around 30 tokens per second with OpenVINO.

Why it matters

It's a sign that 35B-class mixture-of-experts models now fit on small, low-power Intel hardware, not just big GPUs. Its memory bandwidth is about half of a Strix Halo or DGX Spark, so treat it as an embedded or always-on box rather than a main AI workstation.

Source: PR Newswire (LattePanda) prnewswire.com
  • ai-pc
  • memory
  • intel

Jensen Huang says he sees no memory oversupply coming by 2029

Speaking to CNBC at GTC Berlin, NVIDIA's CEO said he doesn't expect a memory glut by 2029, pushing back on a week of memory-stock selling driven by oversupply worries. He said NVIDIA will keep using large amounts of both HBM and LPDDR memory in data centers.

Why it matters

Data centers competing for the same memory chips is a big reason high-RAM machines and big-VRAM GPUs are scarce and marked up right now. If you need a high-memory machine for local models, don't plan around memory getting cheap soon.

Source: CNBC cnbc.com
  • memory
  • nvidia

NVIDIA shows the full 753B GLM 5.3 running locally on an eight-node DGX Spark cluster

NVIDIA's technical blog published its first benchmarks for officially supported eight-node DGX Spark clusters, which pool 1TB of unified memory from eight 128GB units. NVIDIA says the full GLM 5.3 (753B) model runs across the cluster, and that eight-node support in its Sync setup app is coming by the end of the year.

Why it matters

Clustering is now a real path to frontier-size open models at home or in a small office, but scaling isn't linear: in NVIDIA's own test, a 27B model decoded about 4.4 times faster on eight nodes than on one, not eight times. For most buyers, one or two Sparks still covers the models people actually run day to day.

Source: NVIDIA Technical Blog developer.nvidia.com
  • dgx-spark
  • memory
  • nvidia

Ollama and llama.cpp add Recursive Flint, with Ollama running it on Macs first

Ollama 0.40.8 added Recursive Flint, the open 40B mixture-of-experts coding model released this week, and says it currently runs on macOS only, with other GPUs to follow in a later release. llama.cpp added Flint support in build b11803 the same day, with prebuilt CUDA, ROCm, Vulkan and Metal builds.

Why it matters

Mac owners with enough unified memory can try Flint today with one Ollama command. On Windows and Linux GPUs, llama.cpp is the way to run it until Ollama's broader GPU support lands.

Source: Ollama on GitHub github.com
  • software
  • mac
  • ollama
  • llama-cpp

IBM's Granite Speech 5.0 brings a 2B transcription model that runs in llama.cpp and on Apple Silicon

IBM released granite-speech-5.0-2b, an open Apache 2.0 speech model for transcription and translation in six languages, plus faster updates to its smaller Turbo CTC and multi-speaker models. The 2B model works with llama.cpp, vLLM and mlx-audio, and IBM publishes an official GGUF build.

Why it matters

Accurate meeting and voice transcription no longer needs a cloud service or a big GPU. A 2B model is small enough for most recent laptops, Macs and entry-level graphics cards.

Source: Hugging Face Blog (IBM Granite) huggingface.co
  • software
  • speech
  • mac

NVIDIA's DGX Station is coming to Windows in December

At GTC Berlin, NVIDIA said a Windows version of its DGX Station desktop will ship from partners including ASUS, Dell, GIGABYTE, HP, MSI and Supermicro in December. The Linux model, with 748GB of coherent memory, is shipping now, and the Windows version pairs local models and agents with Windows engineering apps from companies like Siemens, Dassault Systèmes and MathWorks.

Why it matters

Until now the biggest NVIDIA desk-side AI box was Linux-only. Teams whose design and engineering tools live on Windows can soon run large local models on the same machine instead of keeping a separate Linux system.

Source: NVIDIA Blog blogs.nvidia.com
  • dgx-station
  • workstation
  • nvidia

AMD and Intel merge their CPU AI libraries into one PyTorch path

AMD is folding its ZenDNN library into the open-source oneDNN library that Intel started, so PyTorch gets one backend tuned for both companies' x86 chips. AMD also joined the UXL Foundation, which publishes oneDNN, as a steering member.

Why it matters

Running models on the CPU is the fallback when a model doesn't fit in VRAM. A single shared backend should make that fallback better tuned on both AMD and Intel desktops without extra setup.

Source: Tom's Hardware tomshardware.com
  • software
  • cpu
  • amd
  • intel

Recursive Flint, an open 40B coding model, is tuned for RTX Spark and RTX cards

Recursive Superintelligence released Recursive Flint, a 40-billion-parameter mixture-of-experts coding model with 2.5 billion active parameters. NVIDIA says it tops 220 tokens per second on a 64GB RTX Spark, and the open weights are on Hugging Face under the Apache 2.0 license, runnable on RTX Spark, DGX Spark, GeForce RTX and RTX PRO hardware.

Why it matters

A fast coding model with only 2.5 billion active parameters makes a 64GB machine a practical local coding assistant, not just a test bench. The Apache 2.0 license also means you can use it in commercial work without restrictive terms.

Source: NVIDIA Blog blogs.nvidia.com
  • software
  • rtx-spark
  • nvidia
  • laptop

NVIDIA brings faster speculative decoding to llama.cpp and vLLM, plus memory-saving FP4 checkpoints

NVIDIA says new multi-token-prediction speculative decoding in llama.cpp makes agent tasks with Qwen 3.8 Flash Next up to 1.4x faster on an RTX PRO 6000 Blackwell, and vLLM changes speed up Qwen3.8 27B by up to 1.3x on DGX Spark. It also released FP4 model checkpoints built to leave memory free for context, starting with Qwen3.8-Flash-Next on a single 128GB DGX Spark or RTX PRO 6000.

Why it matters

These are free software gains on hardware you may already own, and they show up in the tools most local setups use. The FP4 checkpoints matter most for big-memory machines, where fitting the model and still having room for long context is the real limit.

Source: NVIDIA Blog blogs.nvidia.com
  • software
  • nvidia
  • dgx-spark
  • gpu

FLUX.3 Klein image model runs on 8GB GeForce cards with NVIDIA's optimized builds

Black Forest Labs' FLUX.3 Klein is a 5.6-billion-parameter open-weights model for making and editing images. NVIDIA says its optimized builds run on GeForce RTX GPUs with as little as 8GB of VRAM and make a 1-megapixel image in about four seconds on an RTX 5090.

Why it matters

Current-generation image generation no longer needs a 16GB card to get started, so an entry-level 8GB GPU can handle it. A faster card with more memory still pays off in images per minute and room for LoRAs.

Source: NVIDIA Blog blogs.nvidia.com
  • gpu
  • software
  • rtx
  • nvidia

DGX Spark 64GB arrives October 23, and clusters now scale to eight units

NVIDIA says a 64GB version of DGX Spark starts rolling out from partners on Friday, October 23, keeping the same GB10 chip and software as the 128GB model. NVIDIA also says DGX Spark now supports manual clusters of up to eight units, and its Sync app gets a Model Launcher later this month that downloads a model suited to your setup and connects it to OpenCode.

Why it matters

64GB is a lower-capacity way into NVIDIA's Linux desktop AI box, with room for models like Qwen3.8 27B or the 24GB Nemotron 3.5 Bolt, but not the 100B-plus models the 128GB unit is bought for. If you want headroom for larger models or long contexts, the 128GB model or a second unit is still the path.

Source: NVIDIA Blog blogs.nvidia.com
  • dgx-spark
  • memory
  • nvidia

Analyst: about 80% of Samsung's 2027 memory output is already spoken for

The Korea Herald reports a KB Securities estimate that Samsung has locked up roughly 80% of its planned 2027 memory production through long-term contracts, including all of its expected high-bandwidth memory. Micron said last month that more than 75% of its 2027 output is committed and expects tight supply through 2028.

Why it matters

Most new memory is going to data centers under multiyear deals, so RAM kits and high-memory PCs are unlikely to get easier to find next year. If you need 64GB or more for local models, buy from a retailer or the brand when a fair listing shows up rather than waiting for the shortage to end.

Source: The Korea Herald koreaherald.com
  • memory
  • availability

NVIDIA's Nemotron 3.5 Bolt packs a 74B agent model into 24GB for local PCs

NVIDIA released new open Nemotron 3.5 models, including Bolt 74B-A9B, a mixture-of-experts model with 9 billion active parameters that NVIDIA squeezed into 24GB with 2-bit quantization. NVIDIA built it for local agents on RTX Spark machines and says it runs in llama.cpp, vLLM, Ollama, Unsloth and LM Studio. A smaller Bolt E4B for conversational agents is due later this month.

Why it matters

A capable agent model that fits in 24GB puts it within reach of 24GB and 32GB graphics cards, not only 128GB unified-memory machines. Keep in mind that 2-bit quantization trades some quality for size, so leave headroom for context if you plan long agent sessions.

Source: TechCrunch techcrunch.com
  • software
  • nvidia
  • rtx-spark
  • gpu

Apple sets an October 28 "Pro and tell" event, with new MacBook Pros expected

Apple has announced a media event in Los Angeles on October 28. Apple hasn't said what it will show, but 9to5Mac, citing Bloomberg's Mark Gurman, reports the headliner should be touchscreen OLED MacBook Pro models running the current M5 Pro and M5 Max chips, plus an entry-level MacBook Pro and an iMac with the M6. The product details are still reports, not announcements.

Why it matters

If the reports hold, the new MacBook Pro changes the design but not the chips, so memory options for running models should stay where they are. Mac buyers waiting for more local-AI headroom should watch the memory configurations Apple lists on the day, not the screen.

Source: 9to5Mac 9to5mac.com
  • mac
  • laptop
  • rumor

NVIDIA says the RTX 5090 is not discontinued, but supply stays tight

NVIDIA GeForce chief Jeff Fisher said the RTX 5090 has not reached end of life and that shipments to partners continue, responding to reports it had been quietly dropped. He acknowledged that supply and production capacity are constrained, and HotHardware notes the card is rarely in stock at its usual retailers.

Why it matters

The 5090's 32GB of VRAM is why many local-AI builders want it, and it is still in production, but buyers should expect thin stock and heavily marked-up reseller listings. Check that a listing is sold by the retailer or the brand before buying.

Source: HotHardware hothardware.com
  • gpu
  • rtx
  • nvidia
  • availability

Report: Apple's redesigned MacBook Pro may keep M5 Pro and M5 Max chips

9to5Mac, citing Bloomberg's Mark Gurman, reports that Apple's first touchscreen OLED MacBook Pro could be announced in late October and ship as early as November 6. The report says it would use the current M5 Pro and M5 Max chips rather than new M6 versions. Apple has not announced the machine, so treat this as a rumor.

Why it matters

If the report holds, the next MacBook Pro would bring a new design but likely the same chip options, so unified memory choices for running models would not change. Mac buyers waiting only for more AI performance may not gain much by holding off.

Source: 9to5Mac 9to5mac.com
  • mac
  • laptop
  • rumor

ASUS launches its first Intel Arc card, a 32GB Arc Pro B70 built for local AI

ASUS introduced the UGen Arc Pro B70, a dual-slot blower card with 32GB of ECC GDDR6 and 608 GB/s of memory bandwidth on Intel's largest Battlemage chip. ASUS pitches it for local inference, with up to four cards for 128GB of combined VRAM and support for Ollama, llama.cpp, vLLM and PyTorch. ASUS has not confirmed when it will reach stores.

Why it matters

32GB on one card holds mid-size models that don't fit in 16GB or 24GB, and the blower design suits stacking several cards in one box. A big-name board partner backing Arc Pro also widens the choices beyond NVIDIA for VRAM-heavy builds.

Source: VideoCardz videocardz.com
  • gpu
  • intel
  • arc
  • workstation

NVIDIA RTX Spark laptops go on sale with up to 128GB of unified memory

NVIDIA says RTX Spark laptops are now available from ASUS, Dell, HP, Lenovo, Microsoft and MSI, with more systems to follow. The chip pairs a Blackwell RTX GPU with a Grace CPU and shares up to 128GB of memory between them, rated at up to 1 petaflop of FP4 AI compute.

Why it matters

This is the first time a Windows laptop can give a GPU access to 128GB of memory, so big models that used to mean a Mac or a desktop workstation now fit in a laptop with CUDA support.

Source: NVIDIA Newsroom nvidianews.nvidia.com
  • laptop
  • rtx-spark
  • ai-pc
  • memory

First week with RTX Spark: strong GPU, slower CPU, and local AI that still feels early

The Verge tested four RTX Spark laptops from ASUS, HP, Lenovo and Microsoft for a week. Its writers found the GPU can match an M5 Max in some render tests, while Apple's chips stay well ahead on CPU work. They also found local agents ran hot, drained batteries quickly and still felt aimed at developers.

Why it matters

If local AI is the main reason to buy one, expect to work plugged in and to do some setup yourself. Buyers who mostly want CPU speed or long battery life may be better served by a MacBook Pro.

Source: The Verge theverge.com
  • laptop
  • rtx-spark
  • review
  • mac