Huawei Ascend: CloudMatrix 384 Supernode and the CANN Full Stack

1. Introduction

1.1. Vendor Positioning

Huawei Ascend is a full-stack representative of domestic Chinese AI compute, and the platform with the richest deployment practices in this group. Its differentiated route can be summarized in two points:

1. Full-stack domestic localization: chips (Ascend 910 series) — interconnect (LingQu / UnifiedBus) — complete systems (Atlas) — software (CANN/MindSpore/MindIE) are all in-house developed, making it the only domestic vendor that comprehensively matches CUDA at the software-stack level;
2. Supernode differentiation: the CloudMatrix 384 supernode takes the route of "more cards + optical interconnect + global memory pool" to counter the NVIDIA NVL72, officially claiming to make up for the single-card gap at system-level scale.

Main narrative of this article: domestic full stack + supernode differentiation (PDC disaggregation, large-EP inference), with the highest deployment-case density in this group (the most A-grade sources).

1.2. Basic Information Card

1.3. Chip and Shipment Figures

Ascend 910C specifications: Chiplet dual-910B co-package, SMIC N+2 (second-generation 7nm) process, about 53 billion transistors; media figures put FP16 compute at roughly 640—800 TFLOPS, memory bandwidth around 3.2 TB/s, and 128 GB HBM2e; third-party evaluations put inference performance at about 60% of the H100 (not officially endorsed by Huawei, B/C-grade — no single-card parameter page for the 910C on the official site; see the information-gap statement for spec-table gaps).

Two inconsistent sets of shipment figures exist; per discipline they are listed side by side, with no single choice:

The statistical boundaries of the two sets (chip vs. accelerator card, shipment vs. delivery) differ and cannot be directly reconciled; citations must indicate the figures source.

1.4. Position within the AI Harness Framework

The Ascend stack supports the six Harness layers in a way similar to NVIDIA, but with one unique structure: the large-EP (large-scale expert parallelism) scheme pushes MoE-model expert scheduling down to the supernode communication layer, giving L3 (orchestration) hardware-level support on the "expert dimension" — this is, apart from PDC (Prefill/Decode/Compute disaggregation), the Ascend stack's most recognizable contribution to agentic workloads.

2. Glossary

3. Feature Description and Product Lines

3.1. Chips and Complete Systems

1. Ascend 910C: the current workhorse for scale deployment (see section 1.3 for specifications);
2. Atlas complete-system series: Atlas 900 supernode (910C cluster), Atlas 800 series servers (also integrating third-party accelerator cards such as Cambricon; see article 04);
3. Roadmap (2025-09 Huawei HUAWEI CONNECT, B-grade): through 2028, launch the Ascend 950 / 960 / 970 series with significantly higher interconnect bandwidth, compute, and memory capacity; subsequently the Atlas 950/960 SuperPoD (the figure of the Atlas 960 hosting 15488 Ascend cards is a third-party relay).

3.2. CloudMatrix 384 Supernode

Key specifications (HUAWEI CLOUD Ecosystem Conference 2025 figures, relayed via industry compilation, B-grade):

Supernode technical details (LingQu system, Baidu Baike cross-checked against Huawei's public materials, B/C-grade): a 15× communication-bandwidth improvement and sub-200 ns latency; 3168 optical fibers + 6912 400G optical modules supporting long-distance deployment beyond 2 m; a 48 TB global memory pool with unified memory addressing (direct Load/Store access between NPUs); the energy-efficiency ratio raised from the industry's ~30% to above 45%.

3.3. Software Stack: CANN / MindSpore / MindIE / MindSpeed

1. CANN: heterogeneous computing architecture (counterpart to CUDA); CANN 8.0 adds 200+ optimized operators, 80+ fused operators, and 100+ communication/matrix APIs; dual-track community/commercial editions (A-grade, Ascend Community);
2. MindSpore (昇思): 2.6.0 released at the 2025-04 Developer Day, with end-to-end support for training / fine-tuning / reinforcement learning / inference of MoE LLMs such as DeepSeek; since iterated to 2.9/2.10;
3. MindIE: inference engine with PD disaggregation, W8A8 quantization, and service-oriented deployment; MindIE Motor builds in large-EP deployment, DP load balancing, and minute-level fault recovery; interfaces with the vLLM/SGLang open-source ecosystem (see section 5.3);
4. MindSpeed: training acceleration library interfacing with Megatron-style APIs (cross-referenced with article 10 on the training-framework ecosystem).

3.4. Chip Roadmap

See item 3 of section 3.1. Key point: a three-year roadmap (950/960/970) published on an annual cadence — this is Huawei's first public chip plan benchmarked against NVIDIA's annual cadence, intended to stabilize domestic customer expectations.

4. Platform Architecture

4.1. CM384 vs. NVL72 Architecture Comparison

图 3-1｜CloudMatrix 384 与 GB200 NVL72 超节点架构对比

数据来源：华为云生态大会 2025、NVIDIA GTC 2024 官方资料；示意图基于本文分析，对比数据为双方官方口径。

4.2. Deployment Milestones

1. 2025-04: CM384 released, brought online at scale in the Wuhu data center;
2. 2025-09: the Ascend Industry Summit disclosed 300 cumulative sets deployed, serving 20+ customers; the HUAWEI CONNECT the same month published large-EP practices across six industries and the 950/960/970 roadmap;
3. 2026-06: the Shenzhen Hetang platform completed post-training of the 1.6 trillion-parameter DeepSeek-V4-Pro (see section 6.3);
4. 2026-07: per the Baidu Baike figures, cumulative deployments exceeded 750 sets (side-by-side figures, see section 6.4).

5. Harness Design

5.1. Six-Layer Capability Overview

5.2. L1 Context Engineering Layer

The 48 TB global memory pool + unified memory addressing enable cross-NPU semantic access to the KV Cache (officially, a 3× improvement in data-interaction efficiency; B/C-grade Huawei figures) — directly relevant to Agentic AI long-context workloads: context generated by multi-round tool calls can flexibly reside in the supernode memory pool rather than being hard-partitioned by single-card memory boundaries.

5.3. L2 Tools and Execution Layer

1. MindIE Motor: builds in large-EP deployment, PD-disaggregation configuration, DP load balancing, and minute-level fault recovery (A-grade, Ascend Community);
2. vLLM-Ascend plugin: supports Chunked Prefill, Prefix Caching, Speculative Decoding, and multimodality, and contributes back upstream with features such as W8A8 hybrid parallelism and Function Call Multi-LoRA — Function Call support makes it directly serve agentic workloads (A-grade);
3. Open-source interfacing: MindIE interfaces with the vLLM/SGLang open-source ecosystem (see 09-inference-serving.md in detail).

5.4. L3 Orchestration and Control Layer

Large-EP "1 card, 1 expert" scheduling + LingQu supernode 200 ns-class communication support large-scale expert-parallel orchestration (A-grade). This is the Ascend stack's unique structure: expert routing in MoE models is hardware-ized as a cross-card communication problem, and the supernode's low-latency interconnect becomes the physical substrate of L3 orchestration.

5.5. L4 Memory and State Layer

Relying on the global memory pool to realize cross-card state sharing; on the training side, 40-day stable operation and 10-second-class checkpoint recovery, and on the inference side, minute-level fault recovery together constitute state-continuity guarantees (A/B-grade).

5.6. L5 Evaluation and Observability Layer

1. The CCAE cluster self-awareness engine (Ascend official website, cluster-enabling portfolio) provides cluster-level automated O&M;
2. The availability metrics for training/inference (40-day stability, 10-second-class recovery) are all officially disclosed figures; the gap in independent third-party validation coexists with the absence of MLPerf.

5.7. L6 Governance and Security Layer

Full-stack domestic localization (chip — CANN — MindSpore — complete system) is deeply tied to government/finance compliance scenarios; third-party figures put the 910C localization rate at ~55% (B/C-grade). In the context of Xinchuang (IT application innovation) and domestic substitution in critical industries, L6 is Ascend's absolute strength relative to all overseas vendors.

6. Real-World Cases

6.1. Postal Savings Bank of China: Large-EP Inference Solution

Background: the Postal Savings Bank of China runs fine-tuning optimization and reinforcement-learning iteration of its "YouZhi" LLM on the Ascend 384 supernode (A-grade, Ascend AI Industry Summit 2025).

Solution: adopts a large-scale expert-parallel (large EP) inference solution; official figures show a 3× throughput improvement.

Outcomes: applied across five scenarios — risk-control smart review, smart marketing, smart customer service, a loan-approval assistant, and post-training data synthesis — a benchmark case for domestic-LLM inference in finance, and the finance customer case with the highest A-grade credibility in this group.

6.2. Large-EP Practices Across the Six Major Industries

Released at the 2025-09-18 HUAWEI CONNECT (A-grade, Ascend Community official):

Official figures for the large-EP solution: "1 card, 1 expert"; system inference throughput improved by more than 4×, latency within 15 ms; deployed in the core scenarios of more than 50 customers.

6.3. Other Industry and Regional Deployments

6.4. Side-by-Side Listing of Deployment and Shipment Figure Conflicts

This article contains two figures conflicts; per discipline, all are listed side by side, with no single choice:

Conflict 1: Supernode deployment volume:

The time baselines of the two figures differ by ~10 months, which can be explained as temporal evolution; but there is no official quarterly disclosure series between 2025-09 and 2026-07 (gap, see the information-gap statement); citations must indicate the time baseline.

Conflict 2: 2025 shipments: see section 1.3 (expert minutes 630—640K units, D-grade vs. IDC 812,000 cards, B-grade).

7. Summary

Strengths:

1. Full-stack domestic localization is the only complete solution for compliance-mandated scenarios (L6 strength);
2. Structural differentiation of the supernode route: optical-interconnect long-distance deployment, 48 TB global memory pool, large-EP hardware-ized expert scheduling;
3. The highest deployment-case density in this group: finance (Postal Savings Bank), operators, manufacturing (Midea), government affairs, and education all have A/B-grade cases;
4. A pragmatic open-source convergence strategy for the software stack: vLLM-Ascend contributes back upstream, lowering migration barriers.

Weaknesses:

1. Gaps in the validation system: no MLPerf-class third-party benchmark results; key metrics (energy efficiency, compute benchmarking) are all official figures;
2. No official spec sheet for 910C single-card specifications; uncertainty exists in third-party teardown figures;
3. Multiple figures sets coexist for deployment and shipments, introducing noise into external observation.

Applicability boundaries: the first choice for strongly constrained domestic-compliance scenarios such as government affairs, finance, operators, and energy; mixed training-and-inference workloads can rely on the supernode and large-EP solutions; scenarios with hard requirements for third-party benchmark validation need careful evaluation.

Selection recommendations: prioritize case metrics from official A-grade sources (Postal Savings Bank's 3× throughput, large EP's 4× throughput / 15 ms latency) as feasibility anchors; for inference scenarios, prioritize evaluating the MindIE + vLLM-Ascend combination over a purely self-developed stack; watch the 2026—2028 delivery pace of the 950/960/970 roadmap.

Information Gap Statement

1. Huawei official has no formal spec sheet for 910C compute/memory (no single-card parameter page on the official site; existing figures are mostly third-party teardowns);
2. The official whitepaper on CloudMatrix 384 power consumption and energy efficiency is missing ("energy-efficiency advantage over NVIDIA" is a presentation figure, lacking third-party validation);
3. There is no official quarterly disclosure series between "750 sets (2026-07)" and "300 sets (2025-09)" for the Ascend 384 supernode.

8. References

1. Excellent Practices of Large-Scale Expert Parallelism Across Six Ascend Industries — Ascend Community, 2025-09-18. <https://www.hiascend.com/activities/dynamic-news/560?pageSize=10&pageNo=1&domainCode=&id=560&lang=zh>
2. Rising with the Times: the Ascend AI Industry Summit Successfully Held (300 supernode sets, Postal Savings Bank case) — Ascend Community, 2025-09. https://www.hiascend.com/activities/dynamic-news/555
3. Ascend 384 Supernode — Baidu Baike (cross-checked against Huawei's public materials), 2025—2026. https://baike.baidu.com/item/%E6%98%87%E8%85%BE384%E8%B6%85%E8%8A%82%E7%82%B9/68303349
4. Huawei Ascend 910C — Baidu Baike (cross-checked), 2025—2026. https://baike.baidu.com/item/%E5%8D%8E%E4%B8%BA%E6%98%87%E8%85%BE910C/67777523
5. Ascend Community official website (CANN/MindIE/MindSpeed/MindCluster full-stack landscape) — Huawei, 2025. https://www.hiascend.com/
6. MindIE Supported Model List (1.0.RC2 official documentation) — Ascend Community, 2025-03. https://www.hiascend.com/document/detail/zh/mindie/10RC2/whatismindie/mindie_what_0002.html
7. MindSpore official website (2.6—2.10 releases, vLLM-MindSpore plugin, DeepSeek support) — MindSpore Community, 2025. https://mindspore.cn/en
8. The Rise of Domestic AI Chips and Scenario-Based Selection (IDC 2025 China AI accelerator card market data) — Tencent Cloud Developer Community, 2026. https://cloud.tencent.cn/developer/article/2686867
9. In-Depth Insight: LLM Inference Capabilities and the MindIE Engine (vLLM-Ascend ecosystem) — republished via the Volcano Engine ADG community, 2025. https://adg.csdn.net/695251d05b9f5f31781b844f.html
10. Smart-Computing Landscape Brief (Huawei Cloud CM384 specifications and stability data) — informatization think-tank compilation, 2025 (C/D-grade, leads only). https://ima.qq.com/wiki/?shareId=da93ec88352050a760e7af693803074aae16789a9a0e0e81ae70e79558f04186
11. Ascend Research Minutes (2025 shipments, third-party figures; D-grade, for figures comparison only) — Xueqiu, 2025-09. https://xueqiu.com/6800509262/355219046