AMD MI355X Beats NVIDIA B200 on Kimi K2.5 Inference via Community Hackathon

AMD's MI355X GPU has outperformed NVIDIA's B200 on Kimi K2.5 inference via vLLM, a result attributed to community-developed kernel optimizations from a $1.1 million AMD-sponsored hackathon run by GPU_MODE [7][5]. The winning Readonflow Team targeted W4A4 MoE kernels, Top-K kernels, tensor-parallel all-reduce kernels, and the MLA decode metadata planner, achieving over 4x end-to-end MI355X throughput gains [2][5]. Their work has been merged into AMD's AITER kernel library and upstreamed to the vLLM open-source framework [8]. However, AMD still trails NVIDIA on disaggregated vLLM inference for Kimi models — the benchmark win applies only to non-disaggregated configurations [10].

This is one of the first documented cases of AMD hardware beating NVIDIA's current-generation B200 on a major LLM inference workload using upstream, community-contributed software — not proprietary AMD tooling. The upstreaming of these kernels to vLLM means the gains are broadly available, not confined to AMD's own stack. The result tests whether structured community investment (prize-pool hackathons) can close ROCm's long-standing software gap with CUDA faster than traditional R&D.

In late July and early August 2026, AMD and the GPU_MODE community announced that the AMD MI355X had surpassed NVIDIA's B200 on Kimi K2.5 inference throughput using vLLM — a result produced entirely through community-written, open-source kernels contributed via a structured hackathon. AMD funded the event with a $1.1 million prize pool; GPU_MODE co-organizers Mark Saroufim and Anush Elangovan ran it [1]. The winning team, Readonflow, focused on four kernel areas critical to Kimi K2.5's mixture-of-experts architecture: W4A4 MoE kernels, Top-K selection kernels, tensor-parallel all-reduce kernels, and the MLA decode metadata planner [2][3].

The reported performance improvement shifted between the initial July 30 announcement and the fuller August 1 thread. SemiAnalysis initially cited an "over 2x" end-to-end gain [4], then revised to "over 4x" in the August 1 thread [5]. SemiAnalysis's InferenceX benchmarking platform separately published a figure of 7.7x throughput improvement and up to 15x interactivity gains over a 25-day optimization arc [6], likely reflecting a longer comparison window or different baseline. The headline competitive claim — MI355X vLLM beating B200 vLLM on Kimi K2.5 — is consistent across all versions of the reporting [7].

The results have been upstreamed in two places: into the main branch of AMD's AITER kernel library (AMD's official ROCm kernel collection) and into the vLLM open-source inference framework itself [8]. The dual upstreaming is significant because it means any operator running vLLM on MI355X hardware can access these gains without additional effort, and the optimizations enter the community baseline rather than remaining proprietary. SemiAnalysis framed this as a meaningful step toward ROCm–CUDA software parity [9].

One material caveat accompanies the headline result: for disaggregated vLLM inference on Kimi models — a configuration increasingly used in production for prefill/decode separation — AMD still trails NVIDIA [10]. SemiAnalysis flagged this explicitly and indicated they expect further improvements, along with a forthcoming evaluation on Kimi K3. Kimi K2.5 shares its model architecture with xAI Cursor Composer 2.5 [7], meaning the benchmark has some generalizability beyond a single deployment, but the scope of the win remains bounded to aggregated (non-disaggregated) inference on this architectural family.