Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Short answer: Huawei’s Ascend 920 is a real next-generation data-center AI-accelerator project reported in April 2025, positioned to serve Chinese customers affected by restrictions on NVIDIA’s China-focused H20. Reported figures—more than 900 BF16 TFLOPS, about 4 TB/s of memory bandwidth, HBM3 and a 6-nanometer-class process—suggest a serious domestic alternative. They are not, however, independently validated product benchmarks. No public evidence establishes that Ascend 920 matches H20 performance, software compatibility, price, or shipment volume.
What happened, and when?
In April 2025, reporting said the United States had expanded restrictions affecting NVIDIA H20 shipments to China. The H20 had been designed for the Chinese market under earlier rules, so the policy change threatened a product that Chinese cloud and AI companies had been able to procure.
Shortly afterward, Huawei was reported to have introduced or previewed the Ascend 920 at a partner event. Tom’s Hardware described it as an answer to the gap created by the H20 restrictions and reported a target for mass production in the second half of 2025 (Tom’s Hardware). The timing shows a market response, not necessarily that Huawei designed the chip after the new rule; development could have started much earlier.
Available coverage does not independently confirm production volume, widespread shipment, commercial pricing, or a later change in H20 eligibility. Any current purchasing decision therefore needs a check against the applicable US rules, NVIDIA disclosures and Huawei availability in the buyer’s jurisdiction.
Free tools Windows power users keep installed
One-click scans. No signup required.
#1 Best Overall
- NVIDIA Volta GV100 Architecture — 4,608 CUDA Cores, 640 1st-Gen Tensor Cores delivering 14 TFLOPS FP32 and 112 TFLOPS deep learning performance for AI training, inference, HPC, and scientific computing workloads
- 32GB HBM2 ECC Memory — 900 GB/s Bandwidth — High-bandwidth memory on a 4096-bit bus with ECC error correction provides the memory capacity and throughput required for the largest AI models, simulations, and datasets
- PCIe 3.0 x16 Interface — 250W TDP — Standard PCIe Gen3 connectivity with passive cooling designed for enterprise rack server deployment in HPE ProLiant, Dell PowerEdge, and Supermicro platforms with adequate chassis airflow
- NVLink — Scale to 96GB Unified Memory — Connect two V100 GPUs via NVLink at 300 GB/s bi-directional bandwidth to scale GPU memory from 32GB to 96GB for larger AI training and HPC workloads
- Multi-Precision Computing — Supports FP64 (7 TFLOPS), FP32 (14 TFLOPS), FP16 (112 TFLOPS) and INT8 precision modes for flexible deployment across training, inference, and scientific simulation workloads
What Ascend 920 is—and what it is not
Ascend 920 belongs to Huawei’s Ascend family of data-center AI accelerators, built around Huawei’s Da Vinci architecture. It is intended for training and inference rather than consumer graphics. “Chip,” “accelerator card,” “server” and “rack-scale system” describe different layers: a figure quoted for one chip cannot predict the throughput of a complete cluster.
Secondary reports describe Ascend 920 as a successor or higher-end development beyond Ascend 910C. The exact relationship between Ascend 920 and references to “920C” is not established by a definitive public Huawei product taxonomy; some analyst material treats the C designation as a specialized or enhanced variant (Mirae Asset/SKS research).
Reported specifications, with confidence attached
| Attribute | Reported Ascend 920 detail | What is actually established |
|---|---|---|
| Process | SMIC 6-nanometer-class, sometimes called N+3 | Reported estimate; no verified yield or transistor-density data |
| AI compute | More than 900 TFLOPS, generally discussed as BF16 | Reported target or estimate; workload, clock and sparsity assumptions are unclear |
| Memory | HBM3 | Reported, but capacity and supply arrangements are not publicly verified |
| Bandwidth | Approximately 4 TB/s | Reported estimate, not an independent measurement |
| Connectivity | PCIe 5.0 plus a higher-throughput Huawei interconnect | Reported; exact implementation and cluster performance need primary documentation |
| Workloads | Transformer and mixture-of-experts models | Reported design emphasis, not proof of performance on every model |
| Efficiency | About 30–40% better than Ascend 910C in relevant workloads | Huawei-related claim or industry estimate; not an independent benchmark |
These figures were reported by Tom’s Hardware and Japanese technology coverage, which explicitly noted that the chip had not been independently tested (Mynavi Tech+; Tom’s Hardware). A separate report characterized the process and launch timing as rumors rather than confirmed specifications (Notebookcheck).
Rank #2
- High-Performance AI Processing: The MX3 is designed to handle the most demanding AI computer vision workloads, delivering exceptional performance and efficiency.
- Flexible Integration: The MX3 can be easily integrated into your existing systems via its M.2 M-key form factor and support for Linux operating systems.
- Energy Efficient: The MX3 is designed to provide high performance while minimizing power consumption.
- Comprehensive Software Development Kit (SDK): The MX3 is supported by a comprehensive SDK that simplifies development and deployment.
- Hardware compatability: The MX3 is compatible with the PCI-SIG M.2 M-key 2280 Specification. It can be used with the Raspberry Pi 5 with a M-key 2280 HAT.
Ascend 920 versus NVIDIA H20
“Replace” can mean four different things:
- Regulatory replacement: a Chinese customer can buy a domestic accelerator when an H20 shipment is unavailable or requires a license.
- Functional replacement: it can run comparable inference and training models.
- Performance replacement: it delivers similar throughput, latency and utilization on production workloads.
- Commercial replacement: it is available in sufficient volume, supported, priced competitively and scalable in real systems.
The evidence supports Huawei’s first two goals as strategic aims. It does not prove the third or fourth.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallOutdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchHardware comparison
A reported 900-plus BF16 figure cannot establish H20 parity. Buyers need results that specify model, precision, batch size, sequence length, sparsity, software version and whether the result is for a chip, card, server or cluster. Memory capacity is absent from the public reporting; bandwidth alone does not show whether a model will fit or how efficiently data will move.
HBM3 could help workloads that repeatedly stream large weights and activations, but it does not guarantee capacity, latency or power efficiency. The same qualification applies to PCIe 5.0 and Huawei’s interconnect: topology, collective-communication software and node count determine cluster behavior.
Rank #3
- ✅Powered by 26 Tera-Operations Per Second (TOPS) Hailo-8 AI Processor. 2.5W typical power consumption
- ✅Scalable, enabling simultaneous processing of multi-streams & multi-models
- ✅Enabling real-time, low latency and high-efficiency AI inferencing on the edge devices
- ✅Supports TensorFlow, TensorFlow Lite, ONNX, Keras, Pytorch frameworks
- ✅Supports Linux and Windows. Supports the temperature range of -40°C to 85°C
Software and deployment
NVIDIA systems benefit from CUDA, mature libraries, profiling tools, documentation and a large developer base. An Ascend deployment may require model conversion, compiler tuning, new kernels and separate monitoring and debugging workflows. Support for PyTorch, MindSpore, optimized Transformer and MoE kernels, distributed training and enterprise service-level agreements should be evaluated on the exact software release, not inferred from arithmetic specifications.
Supply and total cost
For a Chinese buyer, assured domestic supply may outweigh a modest efficiency gap. Yet no available evidence confirms Ascend 920 lead times, server-OEM integration, cloud access, warranty terms, power requirements or total cost of ownership. A chip that works in a demonstration but cannot be packaged, installed and supported at scale is not a drop-in H20 substitute.
Why the manufacturing node matters
The reported SMIC N+3 or 6-nanometer-class process is strategically significant because China’s leading-edge manufacturing ecosystem faces restrictions on advanced lithography equipment. “6 nm” is a process label, not a guarantee of the same density, yield or energy efficiency as an unrestricted leading-edge process using EUV.
Rank #4
- 48GB AI graphics accelerator
SMIC has been associated with complex DUV multi-patterning at advanced nodes. That can increase process difficulty and cost. Yield, advanced packaging, substrates, HBM supply and test capacity may limit output even if the design is technically sound. The available reports provide no verified yield or volume figures (Mynavi Tech+; Notebookcheck).
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Why HBM3 is a supply-chain question, not just a specification
High-bandwidth memory can remove a major bottleneck in large-model inference and training, but supporting HBM3 on a design is different from securing reliable, high-volume HBM3 integration. Memory stacks, interposers, advanced packaging and testing must all be available in the required quantities. No source in the available reporting verifies Ascend 920 memory capacity or Huawei’s HBM3 supply arrangements.
“Domestic” therefore describes the accelerator designer, not complete supply-chain independence. Manufacturing equipment, EDA software, memory, packaging and networking components may still involve constrained or international inputs.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsBest Value
- Professional AI & Creator Workstation: AMD Radeon AI PRO R9700 GPU with 32GB GDDR6 is engineered for AI development, professional content creation, and compute-intensive workloads.
- Massive 32GB Memory Capacity: 32GB of GDDR6 memory on a 256-bit bus provides ample bandwidth for large AI models, 8K video editing, and complex 3D rendering.
- Advanced RDNA 4 with AI Accelerators: 64 Compute Units with 3rd Gen Ray Tracing and dedicated 2nd Gen AI Accelerators for groundbreaking AI performance and visual computing.
- Professional Blower Cooling: Efficient single blower design exhausts heat directly out of the chassis, ideal for multi-GPU workstation and server configurations.
- Enterprise-Grade Thermal Solution: Vapor chamber heatsink with industrial Honeywell PTM7950 thermal interface material ensures reliable cooling under sustained professional loads.
Ascend 910C, 920 and 920C: avoid name confusion
Industry comparisons commonly associate Ascend 910C with HBM2E and roughly 3.2 TB/s bandwidth, while reports on 920 or 920C point to HBM3, about 4 TB/s and stronger Transformer/MoE optimization. Those are attributed comparisons, not a Huawei datasheet. Confirm whether a source means an individual die, an accelerator card, a board or a rack system before comparing numbers.
What success would look like
Limited success
Huawei supplies selected domestic inference or training workloads where software teams can absorb porting costs and capacity is constrained.
Strategic success
Chinese companies accept somewhat lower efficiency in exchange for predictable domestic supply, local support and reduced exposure to US policy. This outcome does not require matching NVIDIA on every benchmark.
Full replacement
Huawei demonstrates independently reproducible performance, compatible tooling, competitive power and cost, dependable volume, server integration and support across the workloads H20 customers actually run. No available evidence shows that this threshold has been reached.
What buyers and investors should verify
- Independent benchmarks with precision, model, batch, sequence length, sparsity and software versions stated.
- HBM capacity, sustained bandwidth, power draw and performance per watt.
- Interconnect topology and scaling beyond one node.
- PyTorch, MindSpore, model-conversion, profiling and distributed-training support.
- Production volume, lead times, server configurations, cloud availability and warranty terms.
- Applicable export-control or licensing status on the purchase date and in the relevant jurisdiction.
Huawei’s enterprise information is available at Huawei Enterprise, while cloud buyers can check regional offerings at Huawei Cloud. NVIDIA’s data-center portfolio and AI Enterprise software are documented at NVIDIA Data Center and NVIDIA AI Enterprise. Public Ascend 920 pricing or an official order page was not established in the available material.
Bottom line
Ascend 920 matters because it gives Chinese AI companies a plausible domestic route around a threatened H20 supply channel. The reported specifications make it a credible project, not a proven one-for-one replacement. Manufacturing scale, HBM and packaging supply, software portability, independent benchmarks and commercial support will decide whether Huawei merely fills selected workloads or displaces H20-class systems broadly.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
