推論系統
Atlas Submits MLPerf Edge Agent Tests on NVIDIA and AMD Using the Same Inference Core; Results Remain Under Embargo
The Rust/CUDA inference engine Atlas has submitted MLPerf Inference 6.1 results using GB10 and AMD gfx1151 systems, testing multi-turn tool calling on a single accelerator. The submissions are public, but throughput and latency results remain under embargo, so no conclusions about cross-hardware performance can yet be drawn.
Atlas Inference published a build-status update on August 24, stating that it had submitted MLPerf Inference 6.1 closed-edge tests on NVIDIA GB10 and AMD gfx1151. Atlas is a single binary of approximately 75MB whose primary serving path uses Rust and custom GPU kernels without relying on Python or PyTorch; the AMD version handles CUDA interfaces through the SCALE compatibility layer. The project says both hardware targets use the same CUDA source code, but that does not mean their compiled outputs, numerical paths, or kernel tuning are identical. The Atlas public status page also clearly states that performance data has not yet been released.
What makes this submission notable is the workload, not the still-undisclosed ranking. MLCommons’ new edge-agentic specification measures a tool-using agent built around a quantized Qwen3.6-27B model, a 32K serving context, and a single concurrent request. Its accuracy gate is based on approximately 995 single-turn cases from BFCL v4, while the performance test replays 20 agentic coding trajectories totaling 1,007 turns. Metrics include time to first token (TTFT), time per output token, and per-turn end-to-end latency, making the benchmark more representative of local coding-agent workloads than traditional tests that measure only tokens/s. The MLCommons specification also requires both overall accuracy and category-balanced accuracy to reach at least 97% of the reference values.
The Atlas repository has published its model recipes, KV cache quantization modes, and regression gates. Every image must pass startup and output-coherence checks, as well as a performance-regression check that allows no more than a 10% decline relative to the existing baseline. The short-prompt speeds currently listed in the public repository cannot substitute for MLPerf results because the context lengths, accuracy requirements, and measurement methodologies differ. The next step for engineering teams is to determine whether the official results include power consumption, percentile latency, and reproducible configurations, and to verify whether the AMD path truly maintains the same model precision as GB10. Until the data embargo is lifted, completing submissions on both platforms is merely a verifiable engineering milestone—not proof of victory.