GPU News & Updates
The latest news articles relevant to datacenter GPUs, systems and the things that will impact the engineers working to bring AI to life.
Skynet Ran at 60 Teraflops. That Is One Gaming GPU Now.
In Terminator 3 (2003), Skynet runs at "60 teraflops per second", which beat the fastest computer on Earth at the time. In 2026 that is one RTX 4090, or 1/18,000th of a GB300 NVL72 rack at FP4. Count in FP64, the precision a 2003 supercomputer meant, and the whole rack is only 1.7 Skynets. A Halloween look at why a TFLOPS figure without a precision is a plot device, not a spec.
NVIDIA Wants GPU Code to Care Less About the GPU
NVIDIA is bringing native Rust support to CUDA with two projects, cuda-oxide and cutile-rs. The bigger story may be CUDA Tile, a programming model where developers describe operations on tiles of data and the compiler decides how they map onto the GPU. The hardware differences between an A100, an H100 and a B200 are not going anywhere, but software may need to know less about them.
NVIDIA Quietly Lists the RTX PRO 5500 Blackwell: 84 GB, Built for the Rack
A new Blackwell workstation card appeared on NVIDIA's site with no press release: 84 GB of GDDR7 with ECC, 1,398 GB/s, up to 600 W and the RTX 5090's 21,760 CUDA cores, aimed at rack-mounted workstation deployment. NVIDIA has published no TFLOPS, price or date, and the 84 GB does not fit the listed 416-bit bus. We separate what is published from what is not, and the card is now in the database.
Flopper Now Tracks GPU Rental Prices Across 24 Cloud Providers
The spec database now carries price. 337 per-GPU hourly rates from 24 GPU clouds, normalized to USD per GPU-hour, refreshed every six hours, with price history back to July 2026. An on-demand H100 SXM5 ranges from $1.79 to $12.29 an hour depending on where you rent it. Here is what is covered, how the numbers are normalized, and where they show up.
OpenAI's First Chip, Jalapeño, Takes Aim at NVIDIA's Inference Margins
OpenAI unveiled Jalapeño, its first custom chip: an inference-only chip designed in-house and built with Broadcom, scaled over Ethernet rather than NVLink, with OpenAI's own models helping design it. OpenAI claims significantly better performance per watt than current hardware, but the chip is still in testing and it shared no TFLOPS, HBM capacity, power, or confirmed node. We separate what's confirmed from what's still marketing.
Huawei vs Nvidia: The 2026-2028 AI Accelerator Race
With Trump in Beijing and Jensen Huang pitching the H200 to Chinese buyers, we decode Huawei's Ascend 910 → 970 roadmap against Nvidia's Hopper-Blackwell-Rubin lineup. The BIS-TPP numbers most outlets cite aren't raw TFLOPS — once decoded, Huawei's 2028 flagship still sits ~8x behind Nvidia on FP4, but matches on memory capacity at the rack level.
Stay Updated
Get the latest GPU news and benchmarks delivered to your inbox