ROCm 10.0: AMD's Open AI Stack Now Reaches 90% of H100 Performance

AMD’s ROCm just crossed its first decade. With ROCm 10.0, the open-source AI stack is no longer just a CUDA alternative for cost-conscious builders. On standard LLM inference workloads, AMD’s MI355X now reaches 90-95% of H100 throughput via vLLM, while AMD hardware costs 15-40% less.

What Happened

ROCm 10.0 ships with three new components under the ROCm.AI umbrella. The first is a unified CLI with commands like rocm serve for inference and rocm examine for diagnostics, without manually configuring backend libraries. The second is AMD Skills: AMD-validated knowledge integrated directly into Claude, Cursor, and Codex through an Agent Skills format, covering GPU workflows on MI300X through MI355X hardware and CPU inference via EPYC with ZenDNN. The third is Hyperloom, an autonomous optimization system that automates profiling, analysis, planning, and validation cycles that previously required weeks of manual work.

The entire release runs on TheRock, AMD’s automated open-source build system that reached production in ROCm 7.14, enabling unified compilation across AMD’s full GPU portfolio on both Windows and Linux.

Why It Matters

Three metrics define the platform position in 2026. First: MI355X reaches 90-95% of H100 throughput for LLM inference with vLLM or SGLang, according to AMD’s own testing. Second: AMD hardware costs 15-40% less than equivalent NVIDIA configurations. Third: Hyperloom cuts what AMD describes as “weeks of manual optimization work down to hours” by automating the full profiling and tuning cycle.

For agentic AI infrastructure, the CLI-first design matters as much as the raw performance numbers. rocm serve is a one-command inference deployment path; AMD Skills in Cursor and Codex means the toolchain knowledge is embedded where engineers already work. The historical barrier to ROCm was not performance, but setup friction, expertise requirements, and patchy framework support. This release addresses all three.

What This Means for Builders

vLLM and SGLang have official ROCm support with validated containers. PyTorch runs natively. Hyperloom’s autonomous optimization removes the expertise barrier that previously made AMD hardware a reasonable choice only for teams with deep GPU engineering experience. RCCL advanced from 2.28.3 to 2.30.4, closing communication library gaps with NVIDIA’s equivalents.

Teams evaluating GPU infrastructure for agentic workloads have a defensible open-source choice that was not available 18 months ago.

Source: AMD ROCm Blog

Content created with AI assistance and reviewed for accuracy.