DeepSeek and Huawei announced a set of open-source tools for Huawei Ascend accelerators on September 30, 2026, according to Tom’s Hardware’s October 1 report, which cites Reuters. The reported release brings together a compute library, a distributed communication library, and Ascend support for TileLang. It expands the software available to Ascend developers; it does not establish broad CUDA feature parity or make these tools a complete CUDA replacement.
What DeepSeek and Huawei released
The reported release has three parts: tools for computation, tools for communication across accelerators, and a higher-level way to write accelerator kernels. The projects address different layers of AI development, so they should not be treated as one interchangeable package.
As an Amazon Associate I earn from qualifying purchases.
| Tool | Role | What is established |
|---|---|---|
| DeepGEMM-Ascend | Compute | Tom’s Hardware reports that it handles matrix multiplication and other calculations used in DeepSeek models, supports BF16, FP8 and FP4, and preserves programming interfaces from DeepSeek’s existing DeepGEMM library. These details are reported by the outlet; a primary project page was not available in the sources reviewed. |
| DeepEP-Ascend | Distributed communication | The project repository describes communication functions for machine-learning training and inference on Ascend NPUs, with mixture-of-experts expert-parallel dispatch and combine at its core. |
| TileLang support for Ascend | Kernel programming | The main TileLang project announced an Ascend 950 backend, while a separate adapter repository documents Ascend examples and tested-device scope. |
The release is best understood as work on an Ascend software stack, not as evidence that every CUDA workload or API can be moved over unchanged.
What DeepEP-Ascend does
DeepEP-Ascend is a communication library for Ascend devices. Its documented central use is expert-parallel all-to-all dispatch and combine for mixture-of-experts (MoE) models: tokens are routed to experts, and the resulting data is combined across the participating devices.
#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
The repository also lists pipeline communication, bucket collectives for context- and data-parallel work, and Engram remote-memory access. These paths do not all have the same maturity: the project labels several as experimental or in progress. Check the repository’s feature status before depending on a non-core function in a production system.
What TileLang adds—and which devices are covered
TileLang is a Pythonic domain-specific language for writing accelerator kernels, built on TileLang and TVM compiler infrastructure. Rather than requiring developers to write every low-level operation directly, it provides a higher-level kernel-authoring layer. The separate TileLang-Ascend adapter describes examples for GEMM, vector operations and attention.
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.
The main TileLang repository announced an Ascend 950 backend on September 30, 2026, describing native code generation, scheduling, synchronization, and SIMD/SIMT vector programming. The adapter page separately says it has tested A2 and A3 devices. Those statements refer to different project scopes; the adapter’s A2/A3 test claim is not evidence that those tests validate the Ascend 950 backend.
What hardware and software DeepEP-Ascend requires
The repository’s documented setup is specific, not a general support promise for every Ascend system. It calls for Linux on an Ascend host, Ascend 950 with UBMEM connectivity for multi-rank communication, CANN and Ascend C, Bisheng, HCCL/HCOMM, and a matching PyTorch/torch_npu stack.
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
The validated software and hardware combination listed by the project is:
- Ascend 950DT
- CANN 9.2.0
- Python 3.12
- PyTorch 2.13.0+cpu
- torch_npu 2.13.0rc1
DeepEP’s README says its measurements do not establish support for other Ascend generations or CANN versions. A team considering a different configuration should verify compatibility with the project and its hardware provider rather than assume the listed stack generalizes.
Rank #4
- 48GB AI graphics accelerator
How to interpret the performance and availability claims
DeepEP’s published measurements were produced on a manually configured proof-of-concept HDK supplied to the project—not on a broadly available commercial system. The README said a public Atlas 850E Q3 commercial HDK release was planned for around October 15, 2026, subject to Huawei’s schedule, and explicitly noted that the reported results were not collected on that planned release. That date was a plan, not confirmation that the hardware had become publicly available.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →The sources provide no release-specific published numeric benchmark or independently verified comparison with Nvidia hardware. Huawei’s separate 2025 article says its attention/FFN disaggregation design improved decode throughput by “over 50%”; that figure describes Huawei’s 2025 design, not these 2026 tools, and is not a benchmark for DeepEP-Ascend, DeepGEMM-Ascend or TileLang.
Best 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.
Does this replace Nvidia CUDA?
No evidence in the cited material establishes that. The reported goal of reducing reliance on Nvidia’s ecosystem is context for the release, but the available documentation does not demonstrate CUDA feature parity, broad workload portability, or comparative performance. An open-source implementation for Ascend can expand developers’ options without being a drop-in CUDA substitute.
For a practical evaluation, compare the specific Ascend generation and software versions you can use, the operations and kernels your workload needs, communication feature maturity, compiler and programming model, and availability of the required hardware. The repository documentation is the appropriate starting point for checking each tool’s current scope.
Why CANN matters
Huawei’s CANN platform is part of the documented foundation for this work: DeepEP lists CANN components among its prerequisites. Huawei has also described a broader open-source strategy for Ascend software. That context helps explain why libraries and programming tools matter, but it does not prove that every element previously announced by Huawei shipped on schedule.
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 reinstallCrashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteQuick 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.
