BriefsNewsAvailable now
NVIDIA details Tensor Core FP32 methods for AI interatomic potentials
NVIDIA details how BF16x9 and TF32x3 Tensor Core methods accelerate machine learning interatomic potentials without model retraining.
Summary
NVIDIA outlines how Tensor Core FP32 methods like BF16x9 on Blackwell GPUs and TF32x3 on Ampere GPUs accelerate matrix multiplication for the Point Edge Transformer model. These techniques speed up molecular dynamics steps and structural relaxations while preserving accuracy and physical consistency.
What changed
- BF16x9 on Blackwell GPUs accelerates large GEMMs in the Point Edge Transformer by 2.2–2.5x for a 17,280-atom system and 1.2–1.4x for a 192-atom system.
- TF32x3 on Ampere GPUs speeds up major GEMMs by 1.7–1.9x and the full molecular dynamics step by up to 1.5x.
- TF32x3 preserves potential-energy smoothness and energy conservation in NVE trajectories, matching FP32 behavior.
- TF32x3 converges in 218–222 steps with a 1.5x per-step speedup in structural relaxation of a 539-atom Cantor-alloy slab.
Availability
Available now“Read the cuBLAS documentation to enable BF16x9 on Blackwell GPUs.”
Founder takeaway
Read the cuBLAS documentation to enable BF16x9 on Blackwell GPUs if you want to test these performance methods.
Source
Researched and drafted with AI from the official sources above, then checked automatically before publishing. Spotted a mistake? Tell us.
Related
More briefs
OpenRouter details AI performance and cost comparisons for Claude 5.5 models
OpenRouter compares Claude Opus 5.5, Sonnet 5.5, and Haiku 5.5 on coding and benchmark tests, showing similar reliability with wide cost differences.
OpenRouter1 min read
Pika introduces the Pika H3 Flex AI video generation model
Pika released Pika H3 Flex, a post-trained MiniMax H3 video model that supports multiple reference files and audio stems.
Pika1 min read
IBM releases Granite Switch 4.1.1 AI model patch update
IBM released Granite Switch 4.1.1, a patch update for its modular AI architecture that includes an updated multilingual speech model and decision adapter.
Hugging Face1 min read
One email, every Monday
Briefs are published here through the day. Monday Heist is the email: one useful AI workflow, the files to use it, and the catch. The first issue is being prepared; subscribers get it first. Free, and you can leave in one click.
One email. Every Monday.