NVIDIA
14 articles in the last 30 days.
-
Nvidia used 20,000 training examples to fix its AI's mistakes. Its coding score jumped.
Nvidia used Factory's Droid coding agent to generate about 20,000 post-training examples for its open 550B Nemotron 3 Ultra model. It targets tool calls, planning and error recovery, and uses PivotRL.
-
Meta TLX, a Native GPU Compiler Extension Supporting Advanced Hardware Capabilities
Meta's Triton Low-Level Extensions give engineers explicit control over warp specialization, async memory and cluster execution on modern GPUs. Production results include 20–25% MFU on Blackwell GPUs and a 4x kernel speedup on H100.
-
Why your fine-tuning job takes longer than the FLOPs say
On 272 production SFT jobs on H200 GPUs, a FLOP-based runtime estimate explained little (R² 0.31). A CatBoost model using only pre-dispatch information reached R² 0.98 with a 1.71 h mean absolute error.
-
Agentic inference capacity grew 4.6× in a year, versus 2.6× for compute capacity
Epoch estimates that agentic inference capacity of AI chips sold to date grew 4.6× from Q2 2025 to Q2 2026, against 2.6× for peak compute. Blackwell chips account for 74% of the added capacity.
-
Critics balk at Trump's push to return Jensen Huang as Nvidia CEO
Tech Oversight Project billboards in Washington urged Congress to probe Trump administration ties with Jensen Huang. Huang said he would be honored to return, but Nvidia's board says it is not under consideration.