首页 / Xinyuanyi Technology: KV-cache storage acceleration for LLM inference
English overview
Xinyuanyi Technology: KV-cache storage acceleration for LLM inference
Xinyuanyi Technology makes FX all-flash NVMe-oF storage that holds LLM KV cache outside GPU memory, and supplies AMD Instinct MI308X 8-GPU servers. On a 480B model, measured throughput rose 29–40% against a local NVMe drive (reports R2 and R3).
What was measured?
On eight AMD Instinct MI308X cards running Qwen3-Coder-480B-FP8 under long-context cold recovery, keeping KV cache on an FX100 array instead of a local NVMe drive raised throughput by 29–40% and cut median time to first token by 26–32% (report R2).
| Concurrency | FX100 TTFT p50 | Local NVMe TTFT p50 | FX100 throughput | Local NVMe throughput |
|---|---|---|---|---|
| 8 | 7.53 s | 10.17 s | 56.6 tok/s | 43.9 tok/s |
| 16 | 11.85 s | 17.31 s | 74.9 tok/s | 53.6 tok/s |
| 32 | 26.35 s | 35.73 s | 71.6 tok/s | 53.9 tok/s |
Other measured results: a parallel-read patch for LMCache cut cold-read TTFT from 37.97 s to 9.30 s (R1); checkpoint saves went from 178 s to 94 s (R1); on Huawei Ascend 910B, model loading was 6.2–9.3× faster than an NFS baseline (R9).
What does Xinyuanyi sell?
Two product lines. The FX series is all-flash NVMe-oF storage in four models, from the FX100 at 100 Gb per port to the FX400 planned for volume production in late 2026. The second line is AMD Instinct MI308X 8-GPU servers, available to inspect and test.
| Model | Status | Bus | Port | IOPS (vendor spec) |
|---|---|---|---|---|
| FX100 | Shipping | PCIe 3.0 | 100 Gb | 16M |
| FX200 | Shipping | PCIe 4.0 | 200 Gb | 32M |
| FX300 | Shipping | PCIe 5.0 | 400 Gb | 60M |
| FX400 | Test units since August 2026; volume production planned for late 2026 | PCIe 6.0 | 400 Gb | 140M |
As of 2026-09-27, 228 MI308X 8-GPU servers (1,824 GPUs, 192 GB HBM each) are stated as available, with reconciliation against vouchers still in progress. Lots are 1, 4, 8 or 16 servers. No RMB price is published.
How does an engagement start?
With a joint test of about 10 weeks that passes four gates: delivery acceptance, single-node baseline, the main acceleration gate, and a 72-hour stability run. Each gate has a written pass criterion, and the engagement stops at the first gate that fails.
The main gate requires time to first token to drop by at least 25% and throughput gains to land inside the measured 29–40% band. The detailed pages are in Chinese: products, benchmarks, process.
Questions buyers ask
How do I contact Xinyuanyi?
Call 13161818898 (Wang Limin) or write to mingxin@agentmail.to. Email is received around the clock by an AI assistant and passed to staff. Signed reports can be requested the same way, and the original PDFs are hosted on lvsuan.top, the company's other site.
Are the FX200, FX300 and FX400 figures measured?
No. Our published measurements were all taken on the FX100. Figures for the FX200, FX300 and FX400 are vendor specifications, and the FX400 is not yet in volume production. We test other models with the customer's own workload during a joint test.