dataset string | version string | export_date timestamp[s] | license string | publisher dict | platform dict | notes list | experiments list |
|---|---|---|---|---|---|---|---|
mingxin-kvcache-bench results | 1.0.0 | 2026-07-18T00:00:00 | CC-BY-4.0 | {
"name": "Mingxin Technology (铭信)",
"url": "https://mingxinstorage.xyz",
"category": "AI inference storage acceleration"
} | {
"gpu": "8 x AMD Instinct MI308X (192 GB HBM each, gfx942)",
"cpu": "2 x AMD EPYC 9654 (384 threads), ~1.5 TB RAM",
"gpu_stack": "ROCm 7.2",
"inference_engine": "vLLM 0.20.1+rocm721",
"kv_library": "LMCache (upstream mainline, source-built 2026-06-29)",
"device_under_test": "Mingxin FX100 all-flash NVMe-oF... | [
"All numbers below are measured results from signed/official test reports (R1-R5 report IDs match https://mingxinstorage.xyz/evidence).",
"480B numbers use the production deployment form (TP8 single instance or TP4x2 dual instance) under long-context cold-recovery load.",
"Numbers are platform- and condition-sp... | [
{
"id": "tp8_three_way_comparison",
"report": "R2",
"title": "480B TP8 single instance, long-context cold recovery, three-way comparison x concurrency sweep",
"metric_units": {
"ttft_p50": "s",
"throughput": "tok/s",
"ttft": null,
"read_bandwidth": null,
"save_time": nu... |
Mingxin KV-Cache Tiered-Storage Benchmark Results
Measured results for LLM KV Cache tiered-storage acceleration on 8× AMD Instinct MI308X (ROCm 7.2, vLLM 0.20.1+rocm721, LMCache upstream mainline), model Qwen3-Coder-480B-FP8, published by Mingxin Technology.
Device under test: Mingxin FX100 all-flash NVMe-oF array (4-disk RAID0, RoCEv2, 100 GbE) vs local NVMe (PCIe Gen4) vs recompute-only baseline.
Files
kvcache_bench_results.json— all experiments in structured form (platform, per-experiment rows, headline numbers, caveats)tp8_three_way_comparison.csv— 480B TP8 long-context cold-recovery three-way comparison across concurrency 8/16/32
Headline numbers (measured)
- Throughput +29% to +40%, TTFT −26% to −32% vs local NVMe (480B, production deployment form)
- 8.6×–20× faster TTFT vs recompute-without-external-KV
- LMCache parallel-read patch: cold-read TTFT 4.1× better (37.97 s → 9.30 s), disk bandwidth 5.3× (0.98 → 5.23 GB/s)
fs://shared pool: cross-instance KV hot-sharing with zero performance penalty (32/32 cross-read full hits)
All numbers are platform- and condition-specific measurements, not general claims.
Reproduction
Code, orchestration scripts, and the LMCache patch: https://github.com/mingxin-tech/mingxin-kvcache-bench
Full test reports (Chinese, signed/official): https://mingxinstorage.xyz/evidence
Citation
If you use this data, cite "Mingxin Technology, mingxin-kvcache-bench (2026)" and link the GitHub repository.
- Downloads last month
- 33