AMD MI355X runs 32 Kimi-K3 agents per node vs 20 on B300, Signal65 PINNACLE tests show
ryanshrout · x · 2026-09-20
Signal65's Ryan Shrout announced that the AMD Instinct MI355X has joined the PINNACLE GPU benchmark results, alongside three new models: Kimi-K3, GLM-5.2, and an MXFP8 build of MiniMax-M3.
- Kimi-K3 (2.78T parameters, the largest open-weight model), served from MXFP4 weights, is the one model where MI355X clearly beats the B300: one node sustains 32 concurrent enterprise agents vs 20 on B300 (1.6x), and from 16 agents up the MI355X is also faster per agent, with first token still in fractions of a second.
- The MI355X sustains up to 6.5x the agents of the previous-gen MI300X.
- Methodology: a unified PINNACLE agent harness, fully loaded 8-GPU nodes, 90K-token agentic workloads at rising concurrency; service level means every agent decodes ≥10 tokens/s and 95% get first token within 60 seconds.
- The results now cover 12 models across 4 GPU platforms with cross-generation views; all serving configs are still being tuned, so numbers are point-in-time.
Related event: Signal65 Benchmarks Show AMD MI355X Doubles MI300X in Concurrent Agents(3 posts)→
More from Infra
- Sam Altman? No—Bezos Calls for Nuclear-Powered Offshore Data Center Platforms — beffjezos · 2026-11-04
- SpecPilot agent boosts speculative decoding from 4.46× to 10.87×, no retraining — pmttyji · 2026-11-04
- NEURA launches Neuron AI, but Europe's Physical AI test is whether the stack matures together — ingliguori · 2026-11-04
- S3-based databases cache in memory and NVMe, making disk vs S3 latency gap smaller than expected — mgill25 · 2026-11-04
- gpu-unlock project claims up to 2x FP16 tensor throughput for RTX 20/30/40 cards — paranoidray · 2026-11-04
- These 10 GPU optimization papers are why your model runs fast today — NandoDF · 2026-11-04