#gpu

OpenAI and Broadcom build a chip to rival Nvidia’s Blackwell

OpenAI và Broadcom hợp tác phát triển chip AI tùy chỉnh Jalapeño nhằm cạnh tranh với Nvidia Blackwell và Google TPU, nhắm vào workloads inference. Chip này đã được thử nghiệm với mô hình GPT-5.3-Codex-Spark và dự kiến triển khai vào cuối năm 2025, trong khi tình trạng thiếu hụt HBM đang ảnh hưởng đến biên lợi nhuận của Broadcom.

Lập trình viên nên đọc bài này để hiểu cách các công ty lớn như OpenAI và Broadcom hợp tác phát triển chip AI chuyên dụng, giúp tối ưu hóa hiệu suất cho các mô hình lớn như GPT-5.3, ảnh hưởng trực tiếp đến hiệu năng và chi phí của các ứng dụng AI trong tương lai.

Qt Canvas Painter: Accelerated performance using paths

Qt Canvas Painter giới thiệu QCanvasPath và nhóm path nhằm tối ưu hiệu suất render 2D, tiết kiệm ~60% bộ nhớ nhờ định dạng SoA thân thiện cache, hỗ trợ tái sử dụng giữa các frame và tương thích API với QCanvasPainter. Nhóm path cho phép GPU lưu trữ đỉnh (vertex) tĩnh, giúp xử lý biến đổi, màu sắc hay gradient mà không cần cập nhật buffer, đồng thời chứng minh khả năng render tiến triển 60fps trên thiết bị nhúng với tải CPU/GPU thấp hơn. Tính năng này sẽ trở thành module chính thức trong Qt 6.12.

Lập trình viên muốn tối ưu hiệu suất rendering 2D trên Qt nên đọc bài này để khám phá cách sử dụng QCanvasPath và path groups để giảm chi phí CPU/GPU, giảm bộ nhớ và cải thiện trải nghiệm 60FPS trên thiết bị embedded.

The AI memory crisis just hit DDR2, a standard from 2003, with 60% price hikes

Giá DDR2 tăng 55-60% trong Q2/2026 do thiếu hụt DRAM trầm trọng khi các nhà sản xuất chuyển sản lượng wafer sang sản xuất HBM cho AI, dự báo tiếp tục tăng 35-40% trong Q3. Các hãng phần cứng buộc phải hạ cấp từ DDR4 xuống DDR3, rồi DDR3 xuống DDR2 để đảm bảo nguồn cung, ảnh hưởng đến hệ thống nhúng, thiết bị công nghiệp và mạng. Nguồn cung DDR2 không thể cải thiện trước năm 2027-2028 khi các nhà máy mới của SK Hynix và Micron đi vào hoạt động.

Lập trình viên nên đọc bài này vì nó giúp bạn hiểu rõ về những thách thức về nguồn cung cấp bộ nhớ DDR2, ảnh hưởng đến thiết kế phần cứng cho các ứng dụng embedded và hệ thống AI, từ đó có thể tối ưu hóa kiến trúc phần mềm và dự đoán chi phí phát triển sản phẩm trong tương lai.

Nvidia accidentally made the RTX 50 series feel like a beta test for the RTX 60 series

The RTX 50 series launched with headline features like Multi Frame Generation, Ray Reconstruction, and Neural Texture Compression that were either unfinished or lacked broad software adoption. Months after launch, major fixes and updates are still arriving, and the most compelling exclusive features primarily benefit 4K gaming — a niche most PC gamers don't occupy. RTX 40-series owners already receive the biggest DLSS 4.5 image quality improvements, leaving the 50 series in an awkward middle ground. The author argues the generation feels like a transitional stepping stone, with the upcoming RTX 60 series (Rubin) positioned to be the hardware that fully realizes Nvidia's long-term rendering ambitions.

Sail raises $80M to make AI agents cheaper to run

Sail Research has raised $80M in combined seed and Series A funding at a $450M valuation to reduce the cost of running long-horizon AI agents. Founded by ex-Apple and ex-NVIDIA engineers, the company claims its inference engine delivers up to 10x lower cost per token than rivals by optimizing for throughput rather than latency. Sail also offers 'Sailboxes,' sandboxed environments that only charge for active agent compute time. The startup topped the BrowseComp-Plus benchmark at 90.72% accuracy and already has paying customers. The round was led by Sequoia and Kleiner Perkins, with notable angels from Anthropic, OpenAI, and Alphabet.

Unconventional AI releases its first model, built on an oscillator architecture that could cut AI power use by 1000x

Unconventional AI, founded by former Databricks AI chief Naveen Rao, has released Un-0, an image generation model running on a software simulation of a novel oscillator-based computing architecture. Unlike conventional digital logic using transistors, the approach uses coupled ring oscillators to encode and process information through physics, with Rao claiming it could reduce AI power consumption by 1000x. The model performs comparably to established diffusion models like Stable Diffusion, but currently runs only on a simulated version of hardware that doesn't yet exist. The company, backed by $475M in seed funding at a $4.5B valuation from Lightspeed, a16z, and Sequoia, plans to release chip schematics soon and eventually operate as a compute provider. The 1000x efficiency claim remains theoretical until physical hardware is built and validated.

The Download: Europe’s heat wave hits the grid, and IBM’s chip targets Moore’s Law

MIT Technology Review's daily briefing covers several major stories: IBM has unveiled a prototype chip with ~100 billion transistors on a fingernail-sized area, doubling transistor density from 2021 and potentially extending Moore's Law by building transistors upward rather than shrinking them. Europe is experiencing a record heat wave straining power grids, with some plants offline during historically planned maintenance windows. Anthropic has accused Alibaba of illicitly extracting Claude's capabilities via model distillation, calling it the largest known distillation attack. OpenAI and Broadcom revealed their first jointly designed AI chip called Jalapeño. Additional items cover the EU joining a US pact against Chinese AI, data center-driven inflation in memory chips, and companies struggling to control AI token spending costs.

IBM claims major chip breakthrough

IBM has announced the world's first sub-1nm chip technology, dubbed the 0.7nm (7-angstrom) node. The breakthrough uses a new transistor design called 'nanostack,' which stacks transistors in three dimensions rather than laying them flat. The result is nearly 100 billion transistors on a fingernail-sized chip — roughly twice the density of IBM's 2nm chip from 2021 — delivering up to 50% higher performance or 70% greater energy efficiency. IBM says commercial production could begin within five years, though no manufacturing partner has been announced yet. The news comes as chipmakers race to meet the growing compute demands of AI workloads, with Intel also recently advancing its 1.8nm 18A process.

AMD Contributes ONNX Runtime Backend To FFmpeg DNN Filter

An AMD engineer has contributed an ONNX Runtime backend to FFmpeg's DNN (Deep Neural Network) processing filter. The addition enables inferencing across multiple GPU and NPU platforms, including NVIDIA CUDA, Windows DirectML for all major GPU vendors, and AMD Ryzen AI NPU support via the ONNX Runtime VitisAI execution provider. This marks AMD's effort to make the Ryzen AI NPU useful within FFmpeg workflows.

The memory chip crunch is paying off for this U.S. company

Micron, the largest U.S. memory chip maker, reported record Q3 earnings with revenue quadrupling to $41.45 billion and profit surging from $1.88 billion to $28.2 billion year-over-year. The results were driven by surging AI demand for memory chips amid a global shortage predicted to last through 2027. Micron also announced a strategic deal to supply Anthropic with memory and storage chips and participated in Anthropic's Series H funding round. The company forecasts Q4 revenue of $49–$51 billion.

Micron’s revenue quadrupled as AI memory demand pushes gross margins above 81 percent

Micron Technology reported fiscal Q3 revenue of nearly $42bn, quadrupling year-over-year, driven by surging demand for high-bandwidth memory (HBM) used in AI accelerators. Gross margins exceeded 81%, up from 27% a year ago, and adjusted EPS came in at over $25 versus analyst expectations of $21. CEO Sanjay Mehrotra noted the company can only fulfill 50–67% of HBM demand, with all 2026 supply already sold out under multi-year contracts and $22bn in customer prepayments collected. Next-gen HBM4 chips are ramping twice as fast as HBM3E, with revenue already exceeding $1bn. Q4 guidance of ~$50bn far exceeded analyst estimates of ~$44bn. Micron's market cap crossed $1 trillion in May. Risks include being the smallest of three HBM suppliers, cyclical memory pricing, and dependence on continued hyperscaler capex.

Nvidia CEO Jensen Huang says smuggled data centres are a dead end and national security comes first

Nvidia CEO Jensen Huang told shareholders that national security takes priority over commercial opportunities, and that data centres built with smuggled Nvidia chips are unworkable because Nvidia won't provide support or repairs. He addressed the growing chip smuggling problem amid high-profile enforcement actions, including the Supermicro case involving $2.5B in servers routed to China. Grey-market prices for Nvidia's B300 servers in China have reached ~$1M, nearly double the US list price. Huang also highlighted Nvidia's financial performance — $216B in revenue and $96B in free cash flow for fiscal 2026 — and noted that AI ROI questions have been answered, citing tripled GitHub pull requests from AI-generated code.

Runpod hits $1bn valuation on $100M raise

Runpod, a cloud startup renting GPU compute to AI developers, has raised $100M led by Summit Partners at a $1bn valuation — a tenfold increase from its ~$100M seed valuation in 2024. The company reports $240M in annualized revenue, over 1 million developers on the platform, and 20 billion inference requests processed. Unlike rivals focused solely on inference, Runpod offers the full ML lifecycle: experimentation, training, fine-tuning, and deployment. It uses an asset-light model, renting capacity rather than owning data centers, and supports both Nvidia and AMD chips. The company says it turned down buyout offers exceeding $500M to remain independent. Key risks include reliance on third-party hardware, margin pressure when chips are scarce, and well-funded competitors like CoreWeave.

The integrated graphics in AMD's latest APU now trade blows with an RTX 3060, and I'm not sure my dedicated card has a job anymore

AMD's Radeon 8060S integrated GPU inside the Ryzen AI Max+ 395 now matches the RTX 3060 12GB in gaming benchmarks, a milestone for APU graphics. The key breakthrough is memory bandwidth: Strix Halo uses a wide LPDDR5X interface with a large shared memory pool, solving the bottleneck that historically crippled integrated graphics. However, the 8060S only ships in premium laptops and boutique mini PCs costing far more than a budget desktop build with a used discrete GPU. AMD's entry-level desktop APUs like the Radeon 860M remain far behind. The real takeaway isn't a change in current buying advice — used discrete cards still offer better value per dollar — but a shift in APU trajectory: once bandwidth constraints are engineered away at the top, that capability tends to trickle down, potentially threatening entry-level discrete GPUs within a few years.

Accelerating BEV Pooling on NVIDIA GPUs for Physical AI Applications

BEVPoolV3 is a new CUDA kernel optimization for bird's-eye-view (BEV) pooling used in autonomous vehicles and robotics. The post walks through a practical GPU optimization workflow: classify whether the working set fits in L2 cache, remove redundant scatter traffic via a five-array INT32 scatter map, implement interval-owned scatter-reduce to avoid atomics, and validate with NVIDIA Nsight Compute. On RTX PRO 6000 Blackwell Max-Q (large L2), BEVPoolV3 FP8 achieves up to 42x speedup over the V2 baseline. On RTX A6000 (small L2, DRAM-bound), the adapted FP16 path reaches 19x speedup. The post also explains why FP8 outperforms NVFP4 for L2-resident scatter-reduce workloads, and how the same methodology applies to sparse embeddings, voxelization, and other irregular memory-bound kernels.

Qualcomm in talks to design custom chips for ByteDance

Qualcomm is in talks to provide custom chip-design services to ByteDance, TikTok's parent company. The potential deal would see Qualcomm design video processing units and inference ASICs for ByteDance, partly leveraging technology from its $2.4bn AlphaWave Semi acquisition. For Qualcomm, it represents a strategic pivot away from a shrinking smartphone market toward the booming custom ASIC business dominated by Broadcom and Marvell. For ByteDance, it fills a gap left by US export restrictions that block access to Nvidia's most powerful accelerators. No deal has been signed, and the outcome remains uncertain amid escalating US-China chip export tensions. ByteDance is simultaneously exploring chips from multiple domestic Chinese suppliers including Biren, MetaX, and Moore Threads, keeping its options deliberately diversified.

OpenAI's Jalapeño chip: a way out from Nvidia

OpenAI unveiled Jalapeño, its first custom AI chip, designed in partnership with Broadcom for inference workloads. Built in a record nine-month tape-out cycle — partly using OpenAI's own AI models — the chip targets efficiency gains in serving user queries rather than training. OpenAI is already running early samples on Codex workloads, with commercial deployment at Microsoft and other partners expected by end of 2026 and full volume in 2027. The move is framed as a bid for full-stack control and reduced dependence on Nvidia, joining Google, Amazon, and Microsoft who already run custom silicon alongside Nvidia GPUs. Broadcom, which also backs Google's TPUs and has a compute pact with Anthropic, is emerging as the key partner in the industry's post-Nvidia chip diversification. Caveats remain: Jalapeño covers only inference, Nvidia still dominates training, and performance claims are self-reported pending an independent technical report.

New AMD Linux Patches Expose Gamma 2.4 + Gamma 2.6 Curves

New patches for the AMDGPU Linux kernel graphics driver expose Gamma 2.4 and Gamma 2.6 curve support, alongside their inverse curves. Currently the driver only exposes the standard Gamma 2.2 curve. Gamma 2.4 is widely used in movie and TV industries for higher contrast, while Gamma 2.6 is the official standard for digital cinema and theatrical production. The patches originated from interest expressed at the Display Next Hackfest 2026 in France and are expected to land in the Linux v7.3 kernel later this year.

Apple suppliers race to Hong Kong: $4bn in a week

Two of Apple's largest Chinese suppliers, Luxshare Precision and Lingyi iTech, are racing to list on the Hong Kong stock exchange in the same week, together raising roughly $4bn. Both companies are using the capital to pivot from smartphone component manufacturing toward AI hardware and humanoid robotics. Lingyi raised $1.1bn at the top of its price range and is targeting 500,000 humanoid robots annually by 2030, while Luxshare is gauging interest for a ~$3bn listing. The rush is partly driven by a June filing deadline that lets companies avoid refiling financials. Hong Kong IPO proceeds are on track for a six-year high of $43bn in 2026. Analysts note caution: Hong Kong shares priced at steep discounts to mainland listings, China's humanoid robot sector is crowded with over 150 competitors, and demand at scale remains unproven.

AMD FSR Upscaling 4.1 RDNA 3 Support Now Available in FSR SDK 2.3 Update

AMD has released FSR SDK 2.3 ("Redstone"), expanding ML-based FSR Upscaling 4.1 support to RDNA 3 architecture (Radeon RX 7000 Series) discrete GPUs, previously exclusive to RDNA 4. The update also includes FSR Frame Generation 4.0.1 with improved motion vector processing, and FSR Ray Regeneration 1.2 with new denoising options for ambient/specular occlusion and a checkerboard rendering mode (RDNA 4 only). Analytical fallback modes remain available for older AMD hardware. The SDK is available on GitHub.

ByteDance seeks a record $20bn offshore loan to fund its AI build-out

ByteDance is in early talks with banks for a roughly $20bn offshore loan, which would nearly double its 2024 record of $9.5bn. The funds are intended to support an estimated $22.7bn capital expenditure plan for 2026, primarily directed at AI infrastructure. The company faces unique procurement challenges due to US export controls blocking access to Nvidia's top accelerators, pushing it to develop custom CPUs on Arm and RISC-V tracks, partner with Qualcomm for inference chips, and source from smaller Chinese chip suppliers. Simultaneously, ByteDance remains Microsoft's largest AI customer, spending over $1bn annually on OpenAI models via Azure. The loan would help finance both its domestic hardware buildout and its continued reliance on Western AI models.

What's New in WebGPU (Chrome 149-150)

Chrome 149-150 bổ sung hai tính năng WebGPU quan trọng: Immediates (push constants/root constants) cho phép truyền dữ liệu nhỏ, thay đổi thường xuyên trực tiếp vào shader qua WGSL <immediate> và API setImmediates(), tối ưu hiệu suất bằng cách bỏ qua tạo buffer GPU. Đồng thời, quy tắc validation chặt chẽ hơn với transient attachments, yêu cầu viewFormats rỗng khi tạo texture tạm thời và hạn chế thay đổi usage flags trong createView().

Lập trình viên nên đọc bài này để cập nhật cách tối ưu hóa hiệu suất rendering bằng Immediates trong WebGPU, giúp giảm chi phí CPU-GPU và xử lý dữ liệu thay đổi thường xuyên trên mỗi vẽ hình, đồng thời nắm rõ các quy tắc mới về transient attachments để tránh lỗi và bảo đảm tính ổn định trong ứng dụng.

South Korea is in talks with Samsung and SK Hynix over a second chip cluster

South Korea's government is in active talks with Samsung and SK Hynix about building a second large-scale chip manufacturing cluster, driven by explosive AI-fueled demand for memory chips. A presidential adviser indicated that AI orders could accelerate construction timelines by over a decade, potentially bringing new capacity online by 2034–2035. The key challenges are finding a suitable site with adequate land, power, and water, and aligning with President Lee Jae-myung's regional development goals. No financial figures have been officially confirmed, though Korean media reports suggest investments could reach hundreds of trillions of won. The chip boom is also raising domestic policy debates about wealth distribution, housing inflation risks, and labor relations.

NVIDIA and AWS Collaborate to Bring AI to Production at Scale

NVIDIA and AWS have announced several joint infrastructure advancements for enterprise AI at scale. New Amazon EC2 G7 instances powered by NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs deliver up to 4.6x AI inference performance and 2.1x graphics performance over G6 instances, with support for up to 8 GPUs and 700 Gbps networking. Amazon OpenSearch Serverless now uses NVIDIA cuVS for GPU-accelerated vector indexing by default, enabling up to 10x faster vector indexing at a quarter of the CPU-only cost, making billion-scale vector databases buildable in under an hour. Additionally, AWS has achieved NVIDIA Exemplar Cloud status for GB300 training workloads, certifying that AWS meets NVIDIA's rigorous performance benchmarks for large-scale AI training.

Buying an AMD GPU instead of Nvidia? — You're sacrificing more than just ray tracing performance

AMD's GPU market share has collapsed from 36% in 2018 to just 5% by end of 2025. Beyond ray tracing, AMD users face real compromises: Nvidia's DLSS suite (including Multi Frame Generation, Ray Reconstruction, and Dynamic Frame Generation) still leads FSR in image quality, game support (97% vs 72%), and backward compatibility. For AI workloads, CUDA remains the dominant platform with better day-one library support over AMD's ROCm. Creative professionals using Adobe Premiere, After Effects, or Blender also benefit from Nvidia's CUDA acceleration and NVENC encoder quality. AMD has closed gaps significantly, especially with RDNA 4 and FSR 4.1, but Nvidia retains a genuine lead across gaming, AI, and creative software ecosystems.

SageMaker Notebook Instances now support G6e instance types

AWS announces general availability of Amazon EC2 G6e instances on SageMaker notebook instances. Powered by up to 8 NVIDIA L40S Tensor Core GPUs with 48 GB memory per GPU and AMD EPYC processors, G6e instances offer up to 2.5x better performance than G5 instances. They support interactive model training use cases including generative AI fine-tuning and deployment of LLMs with up to 13B parameters, as well as diffusion models for image, video, and audio generation. Available in multiple AWS regions including US East, US West, Asia Pacific, Middle East, and Europe.

Groq’s $650M raise rebuilds it after Nvidia

Groq has closed a $650M funding round led by Disruptive and Infinitum to rebuild itself as an AI inference cloud company, six months after Nvidia licensed its chip technology and hired away its founder Jonathan Ross. The company now operates 13 data centres globally, serves over five million developers, and plans to quadruple capacity to 200 megawatts by end of 2027. A new leadership team has been installed, including CEO Adam Winter and incoming CTO Sinclair Schuller. The core bet is that purpose-built inference chips will outperform general-purpose GPUs on speed and cost as AI workloads shift from training to inference — though Groq now shares its chip IP with Nvidia, the competitor it originally set out to displace.

Nearfield Instruments raises record $380M chip round

Nearfield Instruments, a Rotterdam-based startup that inspects chips at the atomic scale using atomic force microscopes, has closed a $380M Series D at a $1.6B valuation — the largest deep-tech funding round in Dutch history. The round was led by Fidelity and includes sovereign wealth funds from Singapore (Temasek) and Qatar (QIA), signaling strategic interest in chip inspection infrastructure. Nearfield's QUADRA platform performs non-destructive metrology on advanced chip structures like High-NA EUV and gate-all-around transistors, addressing a growing bottleneck as AI demands ever more complex chips. The company sits within a broader Dutch semiconductor cluster alongside ASML and other spinouts, reflecting Europe's push for semiconductor sovereignty. The global chip metrology market was valued at $10.3B in 2025 and is projected to reach $15B by 2031.

NVIDIA Powers Over 400 of the World’s 500 Fastest Supercomputers

NVIDIA now powers over 400 of the TOP500 fastest supercomputers (81%), with nearly 90% of newly added systems built on NVIDIA technology. NVIDIA GPUs accelerate a record 238 systems and NVIDIA networking connects a record 376. On the Green500 efficiency list, the top 8 systems all run NVIDIA GPUs, led by KAIROS at France's University of Toulouse using a Grace Hopper Superchip at 73.3 gigaflops per watt. Grace CPU adoption reached 26 systems, and new Blackwell architecture (B200/GB200) systems are entering rankings globally. Europe has a record 35 NVIDIA AI HPC supercomputers in development.

Optimize AI Project Cloud Costs: 7 Strategies That Actually Work in 2026

AI cloud costs are volatile and hard to attribute, with cloud waste rising to 29% in 2026 driven by AI workloads. Seven strategies are outlined to control spending: (1) allocate every dollar of AI spend to a team or feature using virtual tagging, (2) right-size GPU and model-serving infrastructure based on actual utilization, (3) match the model tier to the task complexity to avoid overpaying for LLM API calls, (4) forecast AI spend and set budget thresholds with alerts, (5) detect cost anomalies in real time before they compound, (6) optimize token usage through prompt compression, caching, batching, and output limits, and (7) apply GPU commitments, spot instances, and autoscaling to reduce compute costs by 30–90%. The post also covers what to look for in an AI cost optimization platform and common mistakes that inflate AI project costs.