Xinyuanyi
XinyuanyiArchitecture
中文

TECHNICAL BRIEF · VS NVIDIA CMX

Accelerate storage-computeA huge cache pool

Spilled KV cache is stored in G2.5, not recomputed. CMX and FX both fill this tier.

Compute rack: GPU nodes above, Xinyuanyi FX storage servers belowGPU compute nodesXinyuanyi FX storage
10–30%
time computing

Even if you deploy a million GPUs, they actually compute maybe 10% to 30% of the time.

Joungho Kim · KAIST · often called the “father of HBM” · Dong-A Ilbo interview · July 2026, as reported

Public trend: compute vs memory

  • Compute
  • Bandwidth
  • Capacity
×1×4×16201620182020202220242026×3.8
×20.4compute · 2.3y×5.4bandwidth · 4.1y×3.8gap / 10y

Both fill G2.5. Neither replaces HBM.

Bandwidth: HBM/DRAM are vendor peaks; G2.5 is R1 measured; local NVMe is PCIe Gen4 class; NFS is derived from R9 load time. GPU util 46.7%→62.8% is R1 (20 switches/hour).

KV cache spills down, no recompute RDMA-attached storage · prefetched back to HBM on hit

Same tier. Different binding.

5× and “token output doubles” use different denominators. Do not cross-compare.

NVIDIA CMX

Public / vendorReported: 2H 2026
Up to 5×vs “traditional storage” · baseline undisclosed

Xinyuanyi FX

SignedShipping
Token output doublesvs recompute · R2

NVIDIA figures are vendor-public with no disclosed baseline. Every Xinyuanyi figure carries a report ID (R1–R9) and is open to re-verification. Not a disparagement or substitution claim. Kim’s 10–30% is a public industry judgment, not our measurement; 46.7%→62.8% is R1. NVIDIA CMX product page · NVIDIA Technical Blog · TechBang on Joungho Kim / Dong-A Ilbo · Epoch AI · Trends in ML hardware (2023)

Checkable materials

The acceptance report is a free download. The two tables below are vendor-published specifications for comparing SKUs, not our measured throughput.

Multi-model token-supply capacity and acceptance report (external official V2.3)Full text is free. It is not one of the paid R1–R9 PDFs in the evidence library.

Vendor specification table comparing MI308X and H20
Vendor specification tables comparing MI308X, MI325X, MI350, and MI355 with the listed NVIDIA SKUs

Vendor-published specifications. Not our measured throughput, and not a commercial multiple versus NVIDIA.