GoKawiil
Tech News
A technical blog post walks through building a custom attention kernel for Nvidia's Blackwell B200 GPU entirely from scratch, using CUDA and some PTX code. Through a 14-step progression of optimizations illustrated with 60 diagrams, the kernel reaches 94.4% of FlashAttention-4's performance on 4K, 8K, and 16K sequence shapes, and is demonstrated by plugging it into a video-generation model.
Nvidia is re-launching its DLSS 5 neural rendering technology after early demos, including a widely mocked look at Resident Evil Requiem's Grace Ashcroft, drew accusations of 'AI slop.' The company now frames DLSS 5 as a tool that preserves developers' artistic intent while giving them control, and hands-on testing suggests it functions as a real-time image filter that adjusts lighting, shadow and detail rather than altering shapes or inventing objects.
A writing contest is inviting entries that envision a future where GPU compute has scaled enough that every person on Earth could have access to the equivalent of a high-end AI chip like the B300, running frontier language models around the clock. The prompt notes today's AI access is bottlenecked by scarce GPU supply, limiting who can use top-tier models like Fable or Sol.
New pricing data tracked across 10 countries shows graphics card prices jumping sharply in the six weeks since the last survey, reversing a modest decline seen earlier in 2026. Nvidia's GeForce lineup appears to be absorbing the steepest increases, based on lowest verified in-stock prices from comparison sites in each region.
The piece explains how Google's Tensor Processing Units (TPUs), built for AI workloads in data centers, differ from GPUs and NPUs. TPUs are now replacing NPUs in Google's upcoming Pixel 11 phones, with Google claiming up to 3.5 times faster AI processing at up to 3.5 times less energy use. GPUs, meanwhile, remain versatile chips handling gaming, 3D rendering, AI training and crypto mining.
Following its latest earnings report, Nvidia is showing investors that its competitive moat isn't limited to GPU chips, where rivals like Amazon and Google are building in-house alternatives. The company's Vera Rubin architecture bundles GPUs with CPUs, inference accelerators, storage and networking racks, positioning Nvidia as the orchestrator of entire gigawatt-scale AI data centers rather than just a chip supplier.
AI cloud provider Lambda has secured $1 billion in private, short-dated debt arranged by JP Morgan Chase to purchase Nvidia AI chips it plans to lease to Microsoft, according to Bloomberg. This follows a $1 billion secured credit facility Lambda closed in May and a $926 million loan this week for Nvidia GB300 GPUs tied to a separate contract with Nvidia itself.
Amazon and Nvidia expanded their partnership, with Amazon agreeing to add 2 million more Nvidia GPUs—including Blackwell Ultra, Rubin, and Rubin Ultra chips—to AWS data centers by 2027-2028. This comes just five months after a prior deal for over 1 million GPUs, with Nvidia citing demand that 'exceeded expectations.' The deal also extends to broader integration of Nvidia's networking hardware, models, CPUs, and robotics platforms into AWS infrastructure.
Apple's Mac mini lineup includes both the newer M6 chip and the older M5 Pro, but the M5 Pro model is priced higher because it targets a different performance tier. The M5 Pro offers more CPU and GPU cores, higher RAM ceilings up to 64GB, and storage options up to 8TB, compared to the M6's maximum 12-core CPU, 32GB RAM, and 2TB storage.
Researchers demonstrated a new Rowhammer variant called GPUThor that defeats NVIDIA's ECC memory protections on Ampere-class workstation GPUs, including the RTX A4000, A4500, A5000, and A6000. By using a non-uniform hammering pattern tuned to avoid triggering GDDR6's Target Row Refresh defenses, the attack generates far more bit flips than earlier concepts like GPUHammer, enabling denial-of-service conditions and root-level privilege escalation.
Apple kept the M5 Max Mac Studio's starting price at $2,499, but choosing the 40-core GPU option—listed as a $300 upgrade—automatically bumps memory from 32GB to 48GB, adding another $300. That makes the real cost of the GPU upgrade $600, not the $300 shown next to it. John Gruber flagged the discrepancy while tallying configuration prices for the new lineup.
Apple has introduced the M5 Ultra, its most powerful chip yet, by fusing two M5 Max dies into a single quad-die processor connected at 4.4TB/s. It offers up to 36 CPU cores, an 80-core GPU option, and support for up to 512GB of unified memory with 1.2TB/s bandwidth, arriving in new Mac Studio models starting at $5,499.