TechPrufer
Get started
Lambda
Lambda logoLambda · Feature trail

MLPerf Training Benchmark Performance

3 posts · Jun 2026 – Sep 2026

What this trail covers

This capability tracks Lambda's performance in MLPerf benchmarks, showcasing advancements in LLM and MoE training and inference speeds on cutting-edge NVIDIA hardware. It highlights software optimizations and system design impacts on training and inference efficiency, providing a baseline for real-world AI training and inference workloads. The v6.1 submission demonstrates significant speed improvements for GPT-OSS 120B on 4 NVIDIA Blackwell Ultra GPUs and introduces the first agentic inference workflow. The EGInterpolator research, accepted at ICLR 2026, demonstrates a novel approach to AI-driven molecular dynamics for drug discovery by first learning molecular structures from abundant static data before fine-tuning on scarce dynamic trajectory data. This method significantly improves the accuracy of molecular motion prediction and reduces reliance on expensive simulations, showcasing the power of GPU-native AI workloads for scientific computing.

All Lambda trails
  1. From structures to dynamics: scaling AI for molecular dynamics in drug discoveryThis post details the application of Lambda's GPU infrastructure and AI capabilities to accelerate molecular dynamics simulations for drug discovery. It introduces EGInterpolator, a two-stage AI model that first learns molecular structures from abundant static conformer data and then fine-tunes on scarce molecular dynamics trajectory data to predict molecular motion. This approach significantly improves the accuracy of predicting bond angles, bond lengths, and torsional motion compared to previous methods, reducing the Jensen-Shannon Divergence from reference MD simulations. The research highlights the data-efficient and scalable path toward AI-driven molecular dynamics, demonstrating the shift of molecular simulation towards GPU-native AI workloads.
  2. MLPerf Inference v6.1: pioneering agent, VLM benchmarksThis post details Lambda's MLPerf Inference v6.1 results, highlighting performance gains on NVIDIA Blackwell Ultra GPUs and HGX B200 systems. Key contributions include: 1) Leading Offline throughput on GPT-OSS 120B with 4 NVIDIA Blackwell Ultra GPUs, showing an 8.85% improvement over v6.0 due to stack optimizations. 2) The first agentic inference workload on datacenter hardware in MLPerf, using Kimi K2.6 (over a trillion parameters) on an NVIDIA HGX B200 system, demonstrating its capability with a novel open-division submission. 3) First benchmarking of Vision-Language Models (VLMs) on Lambda infrastructure with Qwen3-VL-235B-A22B-Instruct, showing significant speedups over v6.0.
  3. MLPerf Training v6.0: Lambda delivers fastest LLM training on NVIDIA GB300 NVL72 and fastest MoE training on NVIDIA HGX B200MLPerf Training v6.0 results are presented, showing Lambda's fastest LLM training on NVIDIA GB300 NVL72 (18.7% improvement over v5.1 for Llama 3.1 8B) and fastest MoE training on NVIDIA HGX B200 (GPT-OSS-20B). The post details system specifications for both GB300 NVL72 and HGX B200, and discusses the impact of hardware innovations (Blackwell Ultra) and software optimizations (tuned NeMo stack) on training speed and efficiency, particularly for MoE architectures.