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

In current compute center construction, the combination of domestic storage and imported GPUs is not a simple replacement strategy but a heterogeneous architecture validated through real-world testing that can significantly improve inference throughput, latency, and training efficiency. Based on multiple measured reports of Mingxin FX100 on the AMD MI308X platform, this article demonstrates the actual performance of this combination in key areas such as KV cache acceleration, model loading, and checkpoint saving, and explores its adaptation logic within the ROCm and Ascend ecosystems, providing quantifiable references for technical decision-makers.

How Domestic Storage Compensates for VRAM and I/O Bottlenecks of Imported GPUs

Imported GPUs (e.g., AMD Instinct MI308X) have inherent limitations in VRAM capacity and PCIe bandwidth, especially when handling long-context inference or large model training, where frequent VRAM swapping and cold-start loading become performance bottlenecks. Domestic storage accelerators (e.g., Mingxin FX100) directly compensate for these shortcomings through NVMe-oF protocol and all-flash arrays.

Taking KV cache acceleration as an example, on a 480B-parameter MoE model (Qwen3-Coder-480B-FP8), FX100 reduces the time-to-first-token (TTFT) without external recomputation from 149.5 seconds to 11.85 seconds, achieving a 12.6x speedup; throughput increases from 4.1 tok/s to 74.9 tok/s, a 18.3x improvement [source: measured, report R2]. This improvement stems from FX100 offloading KV cache from GPU VRAM to high-speed all-flash storage, and through parallel read patches, reducing single-card cold-read TTFT from 37.97 seconds to 9.30 seconds, with bandwidth increased by 5.3x [source: measured, report R1]. For compute centers, this means existing imported GPU clusters can handle longer contexts or larger models without upgrading VRAM.

Measured Gains in Inference and Training Scenarios: From Throughput to Latency

In inference scenarios, the combination of FX100 and AMD MI308X shows stable performance in KV tiered acceleration. In the production deployment of the 480B model, throughput increases by 29% at 8 concurrent requests and 40% at 16 concurrent requests, with a full-system (TP4×2) improvement of 35–36% [source: measured, reports R2/R3]. Time-to-first-token (TTFT) under TP8 with three concurrent requests decreases by 26–32%, with p50 dropping from 10.17–35.73 seconds to 7.53–26.35 seconds [source: measured, report R2]. These data indicate that domestic storage accelerators effectively relieve VRAM pressure on imported GPUs in inference workloads, particularly suitable for online services requiring low-latency responses.

In training scenarios, checkpoint saving is a critical factor affecting training efficiency. In 8-card 32B LoRA training, FX100 reduces full model snapshot save time from 178 seconds to 94 seconds, and sustained write bandwidth increases from 3.26 GB/s to 6.40 GB/s (+96%) [source: measured, report R1]. For large-scale distributed training, this acceleration significantly reduces training downtime and improves GPU utilization.

Adaptation Logic of Domestic Storage with ROCm and Ascend Ecosystems

The value of domestic storage accelerators lies not only in performance but also in cross-ecosystem compatibility. The measured platform for FX100 is based on ROCm 7.2 and vLLM 0.20.1, demonstrating deep adaptation with the AMD GPU ecosystem. On the Huawei Atlas 910B platform, FX100 reduces DeepSeek-32B service loading time from 691 seconds to 112 seconds (6.2x), and DeepSeek-70B from 1399 seconds to 150 seconds (9.3x) [source: measured, report R9 (Ascend platform)]. This cross-platform capability means compute centers can mix imported GPUs (e.g., AMD) and domestic GPUs (e.g., Ascend) on the same storage layer, reducing supply chain risks.

From an architectural perspective, domestic storage accelerators implement a shared storage pool via NVMe-oF protocol, supporting RoCEv2 networking, and can be integrated without modifying upper-layer inference frameworks (e.g., vLLM, LMCache). This "storage-as-acceleration" model avoids the complexity of optimizing storage stacks individually for each GPU ecosystem.

Balancing Cost and Supply Chain

In terms of cost, the reference price for a fully configured Mingxin FX100 is approximately ¥371,200 (about ¥2,014/TB), FX200 about ¥331,200 (about ¥1,797/TB), and FX300 about ¥924,000 (about ¥5,014/TB). Compared to upgrading imported GPU VRAM or adding more GPUs, domestic storage accelerators have lower unit costs and can reduce overall TCO through shared pooling. For example, in the 480B inference scenario, using FX100 increases single-machine throughput by 35–36%, effectively achieving equivalent compute expansion without adding GPUs.

On the supply chain side, domestic storage accelerators use all-flash memory and domestic controller chips, unaffected by export controls on imported GPUs. For data centers aiming for a higher proportion of domestic computing power, this combination can meet policy requirements without sacrificing performance.

Conclusion

The combination of domestic storage and imported GPUs, validated through measured data, demonstrates significant improvements in inference throughput, time-to-first-token, and training efficiency, while offering cross-ecosystem compatibility and cost advantages. Mingxin Technology's FX series products and gated joint testing service (approximately 10 weeks, with measured in-band TTFT reduction ≥25% and throughput +29–40%) provide compute center architects with a quantifiable heterogeneous upgrade path. For further details on joint testing or to obtain estimation models, please contact the Mingxin technical team.

Key Q&A from This Article

Q: What are the measured performance improvements of the domestic storage + imported GPU combination in inference scenarios?
A: On the 480B model, the FX100 and AMD MI308X combination achieves KV tiered acceleration with throughput improvements of 29–40% and time-to-first-token reductions of 26–32% (measured, reports R2/R3).

Q: What is the main benefit of this combination in training scenarios?
A: Checkpoint saving is accelerated by 1.9x, with sustained write bandwidth increased by 96% (measured, report R1), reducing training downtime and improving GPU utilization.

Q: Is the domestic storage accelerator compatible with domestic GPU ecosystems?
A: Yes. FX100 has been validated on the Huawei Atlas 910B platform, with model loading accelerated by 6.2–9.3x (measured, report R9), supporting both ROCm and Ascend ecosystems without modifying upper-layer frameworks.