Heterogeneous Compute Center: The Combinatorial Logic of Domestic Storage + Imported GPU

Against the backdrop of global GPU supply chain volatility and the rapid evolution of the domestic AI computing ecosystem, technical decision-makers in compute centers face a core challenge: how to build compute infrastructure that balances performance and cost-effectiveness without relying on a single supplier? Based on the heterogeneous combination of Mingxin FX100 all-flash NVMe-oF array and AMD Instinct MI308X GPU, measured performance improvements range from 1.9x to 20x in key areas such as KV cache tiered acceleration, model loading, and training checkpoint saving. This solution has also demonstrated 6.2–9.3x model loading acceleration on the Ascend platform (vs. NFS). The core of this combinatorial logic lies in using the high bandwidth and low latency of domestic storage to mitigate the memory bandwidth and capacity bottlenecks of imported GPUs, while reserving interfaces for future domestic GPU integration.

Why is the Combination of Domestic Storage + Imported GPU Practically Relevant?

The technical roadmap selection for compute centers is never a single-dimensional competition. Imported GPUs (e.g., NVIDIA H100/B200, AMD MI308X) hold advantages in general-purpose computing and ecosystem maturity, but they suffer from long supply cycles, high costs, and geopolitical supply disruption risks. Domestic GPUs (e.g., Ascend 910B, Cambricon Siyuan) are still catching up in ecosystem compatibility and single-card performance, but offer supply chain autonomy and relatively lower costs. Domestic storage (e.g., Mingxin FX series) acts as an intermediary layer, providing near-local NVMe access latency via the NVMe-oF protocol while enabling elastic scaling of capacity and bandwidth, serving as a "performance bridge" connecting the two types of GPUs.

Taking the AMD ROCm platform as an example, the Mingxin FX100, in long-context inference for a 480B-parameter model, achieves a 29–40% throughput improvement (concurrency 8–16) and a 26–32% reduction in time-to-first-token (TTFT) (p50 from 10.17–35.73s to 7.53–26.35s) through KV cache tiered acceleration [source: measured, report R2/R3]. The key to this performance gain is that the domestic storage array, with its 100GbE RoCEv2 interface and 16M IOPS random read performance, allows GPUs to recover KV caches from remote storage at near-memory access speeds when video memory is insufficient, rather than relying on the slow I/O of disks or network file systems (NFS).

How Does Domestic Storage Improve Inference and Training Efficiency on the ROCm Platform?

The AMD ROCm ecosystem is gradually maturing for LLM inference and training, but its video memory capacity (MI308X single-card 192GB HBM) remains constrained for 480B-level models. Traditional solutions rely on CPU memory as a secondary cache, but memory bandwidth (~500 GB/s) is far lower than video memory bandwidth (~3.5 TB/s), leading to significant "video memory paging" latency during model switching or long-context inference.

Mingxin FX100's solution is: Offload KV caches from video memory to an all-flash NVMe-oF array, and achieve a 4.1x TTFT improvement via the LMCache parallel read patch (single card, concurrency 16, cold read: TTFT 37.97s → 9.30s, bandwidth 0.98 → 5.23 GB/s) [source: measured, report R1]. In training scenarios, checkpoint saving time is reduced from 178s to 94s (sustained write bandwidth +96%), meaning that in 8-card 32B LoRA training, each snapshot saves 84 seconds, significantly reducing GPU idle wait time [source: measured, report R1].

For the extreme case of no external memory recomputation—i.e., the model is loaded entirely from disk—the FX100 achieves an acceleration factor of 8.6–20x (recomputation baseline TTFT p50 149.5s vs. FX100 11.85s; throughput 4.1 vs. 74.9 tok/s) [source: measured, report R2]. This data indicates that even in the worst-case cold-start scenario, domestic storage can compress inference latency from minutes to seconds.

How Replicable is This Combination on Domestic GPU (Ascend) Platforms?

The fragmentation of the domestic GPU ecosystem is a major concern for technical decision-makers in compute centers. Test results of the Mingxin FX100 on the Huawei Atlas 910B platform provide empirical evidence for the cross-platform replicability of this combination: DeepSeek-32B model service loading time reduced from 691s to 112s (6.2x acceleration), and DeepSeek-70B from 1399s to 150s (9.3x acceleration) [source: measured, report R9 (Ascend platform)].

The logic behind this performance improvement is consistent with the ROCm platform: the Ascend 910B's video memory capacity (~96GB HBM) also faces bottlenecks with 70B-level models, and the low-latency NVMe-oF access of domestic storage enables model weights and KV caches to be loaded without relying on slow NFS or local hard drives. Notably, the baseline for this test was NFS (network file system), not local NVMe, so the acceleration factor reflects the actual benefit of migrating from inefficient network storage to an all-flash NVMe-oF solution. For compute centers already using NFS, this combination can be considered a "zero-code modification" storage layer upgrade.

Conclusion

The combination of domestic storage and imported GPUs is not a simple narrative of "domestic substitution," but a pragmatic choice based on measured performance and supply chain risk. Measured data from the Mingxin FX series on AMD ROCm and Ascend platforms shows that by accelerating KV caches and model loading with an NVMe-oF all-flash array, compute centers can significantly improve inference throughput and training efficiency without changing GPU selection. For technical decision-makers evaluating heterogeneous compute solutions, it is recommended to focus on storage layer performance bottlenecks: when GPU utilization falls below 60% due to I/O wait, the acceleration effect of domestic storage may exceed expectations. Mingxin offers an approximately 10-week gated joint testing process (G1 arrival acceptance/G2 single-unit baseline/G3 main gate: TTFT reduction ≥25%, throughput +29–40% measured in-band/G4 72h stability). Teams with testing needs are welcome to contact us for collaboration.

Key Q&A

Q: What are the core performance improvement figures for the combination of domestic storage (Mingxin FX100) and imported GPU (AMD MI308X)?
A: In long-context inference for a 480B model, KV cache tiered acceleration improves throughput by 29–40% (concurrency 8–16) and reduces time-to-first-token by 26–32% (p50 from 10.17–35.73s to 7.53–26.35s) [source: measured, report R2/R3]; training checkpoint saving time is reduced from 178s to 94s (+96% write bandwidth) [source: measured, report R1].

Q: Is this combination effective on domestic GPU (Ascend) platforms?
A: Yes. On the Huawei Atlas 910B platform, DeepSeek-32B model loading is accelerated by 6.2x (691s→112s), and DeepSeek-70B by 9.3x (1399s→150s), with NFS as the baseline [source: measured, report R9 (Ascend platform)].

Q: How does the Mingxin FX100 achieve these performance improvements?
A: Through an all-flash NVMe-oF array (single-port 100GbE, 16M IOPS) providing near-local NVMe access latency, combined with the LMCache parallel read patch, the loading speed of KV caches and model weights is increased to 5.23 GB/s (vs. baseline 0.98 GB/s), reducing GPU idle time caused by I/O wait.