{"resourceId":"nvidia-devblog-tensor-core-emulation-ml-interatomic-potentials-epfl-20261029","versions":[{"version":"external-dfac8c4ff9400a7ccc4669beccdf436d264ceecca0c75c6a490dfe8081a24a94","resource":{"id":"nvidia-devblog-tensor-core-emulation-ml-interatomic-potentials-epfl-20261029","title":"NVIDIA and EPFL: cuBLAS FP32 emulation on Tensor Cores speeds machine-learning interatomic potentials by up to 1.7x on Blackwell and 1.5x on A100 with one environment variable","organization":"NVIDIA with EPFL (Michele Ceriotti)","sector":"Scientific computing (vendor technical blog)","geography":"Global (Switzerland and United States contributors)","publishedAt":"2026-10-29","publicationDate":"2026-10-29","eventDate":null,"sourceName":"NVIDIA Technical Blog","sourceLabel":"Unlocking Tensor Core Performance for Machine Learning Interatomic Potentials","sourceUrl":"https://developer.nvidia.com/blog/unlocking-tensor-core-performance-for-machine-learning-interatomic-potentials/","evidenceClass":"vendor-claim","outcomeClass":"emerging","topics":["infrastructure","developers-agents"],"finding":"NVIDIA's Oct 29 technical post, co-authored with EPFL's Michele Ceriotti, reports that cuBLAS FP32 emulation can run single-precision matrix math on Tensor Cores without code changes, enabled by setting CUBLAS_EMULATE_SINGLE_PRECISION=1. On Blackwell, BF16x9 emulation sped large GEMMs in the PET model by 2.2-2.5x for a 17,280-atom system and the full molecular dynamics step by up to 1.7x. On A100, TF32x3 sped GEMMs by 1.7-1.9x and the MD step by up to 1.5x. The authors say end-to-end gains depend on system size and GEMM share, and TF32x3 should be validated against accuracy requirements.","sledRelevance":"University and national-lab-affiliated materials science groups on existing A100 and Blackwell clusters may get faster simulations without new hardware, extending the value of public research computing investments.","evidence":"Vendor technical blog with co-author benchmarks on specific models and systems; no independent replication.","architectureImplications":"Precision emulation is a library-level setting, so cluster teams can offer it through environment modules without changing user code.","governanceImplications":"Research groups should document numerical settings for reproducibility.","securityPrivacyImplications":"No direct security implications identified.","caveats":"Vendor-authored with benchmarks on a single model family; gains vary with system size; TF32x3 trades some accuracy and must be validated; combined gains with other optimizations were not measured.","streamIds":["nvidia"],"roles":{"sales":"For research computing leaders, this is a low-cost way to show more value from GPUs they already own. A single environment variable gave up to 1.7x faster simulation steps on Blackwell and 1.5x on A100 in the vendor and EPFL tests. Offer to help benchmark user codes on the campus cluster. Avoid promising these figures for other models, since they come from one model family and depend on workload.","engineering":"Engineering should benchmark CUBLAS_EMULATE_SINGLE_PRECISION=1 on representative user codes, recording speedup and accuracy against FP32 baselines. Confirm cuBLAS and driver versions support the emulation on each GPU generation. For TF32x3 on Ampere, check force errors and trajectory stability as the authors advise. Publish tested settings through environment modules and document where emulation should not be used. Retest after cuBLAS or driver upgrades.","delivery":"Delivery should run a short benchmarking campaign with two or three research groups, report results to users, and publish guidance on when to enable emulation. Add numerical settings to job documentation for reproducibility. Track adoption and queue impact. Coordinate with the utilization study so that software gains feed into capacity planning before new purchases. Keep FP32 runs available for comparison."},"roleDetails":{"sales":{"customerProblem":"Research groups need faster simulations but capacity budgets are limited.","stakeholders":["Research computing director","Materials science faculty"],"discoveryQuestions":["Which user codes are dominated by FP32 GEMMs?","What GPU generations are in the cluster?"],"valueHypothesis":"Library-level emulation may increase throughput on existing GPUs.","potentialEngagement":"Benchmarking support for user codes with emulation enabled.","unsupportedClaims":["Gains for models beyond those tested","Accuracy equivalence without validation"]},"engineering":{"fit":"Clusters running FP32-heavy scientific ML codes on Ampere or Blackwell GPUs.","architectureIntegration":"Environment-variable setting in cuBLAS, delivered via environment modules.","prerequisites":["Supported cuBLAS/driver versions","FP32 baseline for comparison"],"constraints":["Gains depend on system size and GEMM share","TF32x3 accuracy trade-off"],"security":"No direct security implications.","proofOfValue":["Benchmark speedup and accuracy on representative codes","Validate trajectory stability for TF32x3"]},"delivery":{"work":"Benchmark emulation with research groups and publish guidance.","dependencies":["Research groups","Cluster software stack"],"ownership":"Research computing user support.","skillsAdoption":"Guidance on numerical validation for researchers.","governanceCheckpoints":["Benchmarks reviewed with researchers","Guidance published"],"acceptanceCriteria":["Proposed: measured speedup with accuracy within user-defined tolerance","Proposed: documented settings in job scripts"],"risks":["Silent accuracy loss","Version incompatibilities"]}},"retrievedAt":"2026-10-30T03:02:32Z","enrichedAt":"2026-10-30T03:10:14Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Researchers need guidance on when emulation is numerically appropriate.","procurementImplications":"May improve throughput of already purchased GPUs, which can inform capacity planning.","operatingModelImplications":"Cluster support teams can publish tested settings and validation guidance for common codes.","sourceVerification":{"openedUrl":"https://developer.nvidia.com/blog/unlocking-tensor-core-performance-for-machine-learning-interatomic-potentials/","referenceExcerpt":"End-to-end gains depend on system size and the share of runtime spent in GEMMs.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}