GPU computing

Developer lyogavin's AirLLM project claims to enable inference of a 2.8-trillion-parameter large language model on a single 4GB GPU, a striking efficiency claim given the hardware typically required for models of that scale. The open-source repository, written largely as Jupyter Notebooks, has drawn attention on GitHub and is being discussed as a way to run frontier-scale models cheaply on consumer machines.

Apple researchers have detailed 'Minions', a virtual GPU cloud built by pooling idle Apple Silicon Macs into a distributed computing cluster. The approach turns spare capacity in existing Mac hardware into a shared resource for machine learning workloads, instead of relying solely on dedicated server GPUs. Readers are debating how practical the setup is for real AI training and inference compared with traditional data-centre hardware.

The AnyPS5 project has reportedly achieved 100% translation of PS5 GPU shader instructions to PC, according to OC3D coverage. If accurate, the breakthrough would mark a major step toward running PS5 games directly on PC hardware without official ports. Interest is rising around what this means for console emulation and Sony's platform exclusivity.

Bolt Graphics is highlighting hardware-accelerated physics modeling running on its Zeus test chip, according to HPCwire. The company says the demonstration shows its silicon handling physics computations directly in hardware rather than in software alone, an approach aimed at high-performance computing and simulation workloads. Details on performance figures and availability remain limited so far.

GPU cloud provider RunPod has named a new chief financial officer whose stated priorities are scaling the company's finance operations and supporting its AI infrastructure buildout. The appointment, reported by CFO Dive, signals the fast-growing AI cloud firm is professionalizing its financial leadership as demand for compute capacity keeps rising.

Intel has confirmed that its Crescent Island AI inference accelerator will ship commercially as the GPU D-Series family. The chips are advertised with 1.3 PFLOPS of FP8 compute and up to 480 GB of LPDDR5X memory delivering around 1.5 TB/s of bandwidth, positioning them for memory-heavy AI inference workloads rather than training. The specs put Intel in direct competition with Nvidia's inference offerings, and tech watchers are weighing whether the memory-first design gives Intel a real opening in the inference market.

NVIDIA is relying on PTFE, a low-loss fluoropolymer better known as Teflon, as a key material in the cabling and interconnects of its next-generation AI rack systems. The material helps preserve signal integrity at the extremely high data rates required to link thousands of GPUs inside a single rack-scale computing cluster.

A French-language joke is circulating comparing tech billionaires playing a game of drawing the short straw, where whoever has the smallest GPU is the one stuck with the task. The quip riffs on the AI race among wealthy tech figures, mocking how compute power, measured in GPUs, has become the scorecard separating winners from losers.

NVIDIA is running its GTC event in Washington DC, the company's conference focused on AI, chips and high-performance computing. The gathering typically features announcements on new GPU architectures, software platforms and partnerships, and draws policymakers given the strategic importance of semiconductors to US technology policy.

A new video asks what would happen if the tech industry stopped relying on GPUs for its computing needs. The question lands amid the AI boom, where graphics chips have become the most sought-after hardware in the world and a major economic driver. It invites debate over whether alternatives like CPUs, ASICs or other architectures could realistically take on the workloads GPUs now dominate.

Microsoft is developing the Maia 300, an in-house AI chip that analysts read as an attempt to reduce its reliance on the massive rented compute arrangements it helped fuel. The company was among the biggest backers of the datacenter buildout that made cloud GPU access the industry standard, and now appears to be positioning its own silicon as a path to lower costs and more control over AI infrastructure. The move highlights growing tension between hyperscalers and their compute suppliers.

Chinese AI chipmaker Iluvatar CoreX has unveiled its first supernode system, the Tiangai 300 IXL144, packing 144 GPUs into a single cabinet, which the company calls a new industrial record. The system was launched at the China International Industry Fair in Shanghai and is designed to support large-scale AI computing workloads.

CoreWeave, the AI cloud infrastructure company that went public on Nasdaq under ticker CRWV, is drawing attention from investors following its stock price movements. The company, which provides GPU-based cloud computing services to AI developers, has become one of the most closely watched new listings as markets weigh the outlook for artificial intelligence spending.

Nvidia has reportedly stopped production of its RTX 5090, the most powerful consumer gaming GPU on the market, according to a report cited by VGC. The company is said to be shifting manufacturing capacity toward RTX Pro GPUs destined for AI data centers, where demand for compute hardware has surged. Gamers and analysts are debating whether top-end gaming cards are being squeezed out by the AI boom, and what it means for prices and availability.

CoreWeave, the AI-focused cloud computing company, is being asked whether it can seriously compete with Amazon Web Services, the long-standing leader in cloud infrastructure. The question follows CoreWeave's rapid growth built on supplying GPU capacity for artificial intelligence workloads, a niche where demand has surged. Commenters are weighing whether that AI momentum can translate into broader cloud market share against Amazon's scale.

Reports circulating in Russian media claim SpaceX is preparing a deal worth around $40 billion to purchase GPUs from NVIDIA. If confirmed, the purchase would be one of the largest AI hardware acquisitions ever, underscoring the intense competition for computing power among major tech and space companies. Details of the reported deal remain sparse, and neither SpaceX nor NVIDIA has publicly confirmed the figures.

Three AI fronts accelerated at once this week. Anthropic open-sourced a small model, Haiku 5.5, while Moonshot released 1.56 TB of weights for its Kimi K3 model. SpaceX is reportedly raising $40 billion to buy NVIDIA GPUs, and Microsoft and NVIDIA launched AI PCs with integrated local inference. DeepMind also unveiled Gemini Robotics 2. Commenters note models are getting cheaper, compute scarcer, and PCs are starting to take over part of inference work, forcing engineering teams to reassess cost and deployment boundaries.

Law firm Pillsbury Winthrop Shaw Pittman has published an analysis of GPU financing, examining the legal risks and opportunities as graphics processing units become a distinct asset class. The piece covers how lenders, investors and borrowers can structure deals around compute hardware, a market growing quickly on the back of artificial intelligence demand, while flagging collateral, valuation and regulatory questions that remain unsettled.

Cedana has published a reference guide estimating how many GPUs are needed to serve different volumes of AI tokens, from one million up to billions or trillions. The piece offers engineers and planners a rough way to translate token demand into compute capacity, a recurring question as companies scale AI workloads and budget for infrastructure.

The Wall Street Journal issued a correction to its October 6 report that AI-computing startup Lambda is raising $4 billion in what would be its final funding round before a planned initial public offering. Lambda, which provides GPU cloud computing for AI workloads, is among the infrastructure firms riding demand for artificial intelligence compute capacity.

Security researchers have described PoeLLM, malware that scans for AI servers left exposed online and hijacks their computing power to mine cryptocurrency. The campaign exploits misconfigured machine-learning infrastructure, a growing target as companies deploy GPU-heavy AI workloads. Commenters are warning organisations to lock down their AI endpoints and monitor for unusual resource usage.

Architect has launched Liquid Inference, a platform that runs a real-time auction for large language model inference capacity, allowing buyers and sellers to trade compute as demand fluctuates. The move aims to make AI inference cheaper and more efficient by pricing GPU resources dynamically rather than through fixed contracts, adding to growing efforts to build open markets for AI compute.