Section

Hardware

Silicon, devices, robots, and the machines that run the models.

334 stories · page 11 of 12

Hardware, stories 301 to 330

301Nvidia's 800-Volt Racks Aim to Push 157% More Power Through the Same WireTexas Instruments unveiled the reference chips for Nvidia's design at GTC in March, hitting 97.6% conversion efficiency as GPU power draw jumped 75% in one generation.Sep 14
302Micron Aims to Nearly Double HBM4 Wafer Capacity by the End of 2026The reported jump to 100,000 wafers a month, relayed by TrendForce and a separate German outlet, would still leave Micron near a fifth of Samsung and SK Hynix's combined output.Sep 14
303Stanford's 'Intelligence per Watt' Metric Says Local AI Handles 88.7% of QueriesA benchmark across 20-plus models and eight accelerators finds on-device efficiency improved 5.3 times since 2023, though cloud GPUs still beat laptops by at least 1.4 times per watt.Sep 14
304DeepSeek's V4.1-Flash Activates Just 8 of Its 552 Billion ParametersThe mixture-of-experts model's 1-million-token context window pushes even a heavily compressed community build to 170 gigabytes, more than a single workstation GPU holds.Sep 14
305Samsung, SK Hynix Reject KEPCO's $18 Billion Power PrepaymentKorea's state utility wanted five years of chip-cluster electricity bills paid upfront, and its own numbers show why it asked.Sep 14
306Samsung, SK Hynix Fall Below 10 Days of Memory SupplyKB Securities calls the setup for 2027 the tightest memory market in its history, and TrendForce data suggests HBM isn't even the most profitable wafer right now.Sep 14
307Firmus Seeks $5 Billion Australian IPO After OpenAI DealThe Nvidia-backed data center builder wants to list in Sydney six days after naming OpenAI as an anchor customer for two new sites in Malaysia.Sep 14
308AWS Cuts LLM Cold Starts From Minutes to SecondsSageMaker HyperPod's new model caching claims a 60% faster scale-out, though the underlying cold-start problem it targets has an academic fix that already claims far more.Sep 13
309The Egret GT E-Scooter's Firmware Updates Need No AuthenticationA developer reverse-engineered six chips inside the scooter and rewrote its firmware in Rust after finding the update process checks nothing at all.Sep 13
310A New CUDA-on-AMD Project Walks Into Nvidia's Legal MinefieldSpeedstu published a Windows compatibility layer routing CUDA calls through ZLUDA to AMD's HIP stack, the exact use Nvidia's own license has banned since 2021.Sep 13
311A Software Fix Nearly Triples Apple's M3 Neural Engine BandwidthAn independent developer traced a 1-mebibyte DMA quirk that had been quietly cutting the chip's memory bandwidth to a third of its rated ceiling since the M3 shipped.Sep 13
312OpenAI's Storage Layer Now Handles 70 Million Requests a SecondA two-engineer Rust rewrite done in 2026 made Habitat, the system underneath ChatGPT and Codex, six times more CPU-efficient than the Python service it replaced.Sep 13
313Inside the 140 Micro-Instructions Behind One 8087 Chip InstructionA reverse-engineering group called the Opcode Collective found Intel's 1980 FSCALE instruction needs three levels of subroutine calls just to multiply a number by a power of two.Sep 13
314ISTA Quantizer Squeezes Qwen3.8-27B Into 8.4 GigabytesGSQ-RCO's smallest file scores 10 points higher on AIME25 than Unsloth's same-size quantization of the same model, according to ISTA-DASLab's own benchmark table.Sep 13
315iPhone 18 Pro Max Splits Its Modem Chip by CountryApple's C2 modem replaces Qualcomm worldwide in the new Pro Max, except in the US, where MacRumors says the launch unit still runs a Qualcomm chip.Sep 13
316Nvidia's SEC Filing Doesn't Name Its $70 Billion in BetsNvidia disclosed $47.9 billion in non-marketable equity stakes without naming a single company; a Bank of America analyst says $45 billion of a related $70 billion figure went to OpenAI, Anthropic and Safe Superintelligence alone.Sep 12
317iPhone Duo's 100-Part Hinge Enters a Market Growing 0.5%Apple's first foldable ships with a 3D-printed titanium hinge and a $1,999 price, but TrendForce forecasts the entire foldable smartphone category growing barely at all in 2026.Sep 12
318DeepSeek's 'Flash' Model Needs 475GB Before Anyone Runs ItDeepSeek's V4.1-Flash activates just 8 billion parameters per token, but the 763-billion-parameter repository behind that number rules out every consumer machine on the market today.Sep 12
319AWS Puts Graviton5 Behind Lambda's Managed InstancesNine months after previewing the chip, AWS still frames every Graviton5 performance number as its own calculation, not an independently reproduced one.Sep 12
320ASML Locks In TSMC and Samsung for a 12-Inch Chipmaking MaskThe Sept. 8 photomask pact fixes a lithography bottleneck that won't reach production until 2030 at the earliest, while the memory shortage it's meant to eventually ease is already raising DRAM prices today.Sep 12
321Apple's Neural Engine Gets Its First Public Register MapIndependent researcher Eileen Yoon's account of the M1's accelerator shows Apple's 2020 marketing number holds only when the workload never leaves on-chip cache.Sep 12
322Amkor Triples Its Arizona Packaging Bet to $12 BillionLess than a year after breaking ground at $7 billion, Amkor's Sept. 8 Phase 2 announcement pushes the same site to six times its original price tag.Sep 12
323TSMC Posts Record $16.35 Billion August Revenue, Preps Up to 15% Price HikesRecord monthly revenue and fully booked advanced nodes are the same fact seen from two sides, and the side Nvidia and AMD are on is about to get more expensive.Sep 11
324What It Actually Takes to Run Qwen3.8-27B: 262K Context Meets Real VRAM MathThe 27-billion-parameter model fits on a single 24-gigabyte consumer GPU at 4-bit precision, but that number describes the weights alone, not what a long context or more than one user adds on top.Sep 11
325Hugging Face Ships 207 WebGPU Kernels to Turn Browser Tabs Into Inference HardwareThe kernels ran 2.57 times faster than ONNX Runtime Web on an Apple M4 GPU, and the point isn't speed for its own sake, it's that the browser's own GPU access is now fast enough to matter.Sep 11
326AWS Cuts SageMaker HyperPod Cold Starts by 60% With Local NVMe CachingModel caching keeps weights and container images on local disk instead of pulling them fresh on every scale-out, closing a gap that has made GPU autoscaling unreliable for inference.Sep 11
327Oracle Triples GPU Deliveries to 300,000, Backlog Grows FasterOracle's cloud backlog rose to $664 billion in the quarter ended Aug. 31, nearly seven times the $90 billion to $95 billion it plans to spend on capital projects this fiscal year.Sep 11
328EuroHPC Signs $450 Million Deal for AMD-Powered LUMI-AIThe €387.8 million contract promises a tenfold jump in AI capacity over the current LUMI supercomputer by 2027, though EuroHPC did not say in what unit that capacity is measured.Sep 11
329How CHERIoT Locks Down Chips With No Memory Manager at AllThe Microsoft-born project reaches memory safety on cheap embedded chips by putting bounds and permissions inside every pointer instead of relying on hardware most of them don't have.Sep 11
330System76's Thelio Mira AI Tops Out at 192GB of GPU Memory, $40,538The Linux workstation starts at $3,299 with a single 4GB card, and the jump to dual RTX Pro 6000 GPUs alone adds $37,239 to that base price.Sep 11