Atlas 950 SuperPoD
Rack-scale AI supercomputing cabinet built from 64 Huawei 950DT NPUs per cabinet and scalable to 1,024 NPUs over a UB Link fabric for trillion-parameter model training and inference.
Find alternatives to Atlas 950 SuperPoD
What it does
- AI compute hardware
- Server systems
- Model training
- Interconnect
Features 3
- SuperPoD architectureSupports up to 1,024 NPUs with 1.72 PB/s UB Link bandwidth, in 64-NPU or 1,024-NPU configurations.
- High performanceUp to 0.9 EFLOPS at mxFP8/FP8/HiF8 and support for formats from FP32 to MXFP4.
- High reliabilityMulti-level reliability design across devices, network and systems, with UB link self-recovery.
How it compares
-
NVIDIA Vera Rubin
NVIDIA plans to launch their NVL144 system in the latter half of 2026. Our Atlas 950 SuperPoD will have 56.8 times more NPUs than they've got GPUs, and will deliver 6.7 times more computing power.
Press release · 13 Oct 2026
Sources
- Pricing
- Sold within
-
The Atlas 950 SuperPoD page gives a full specification - "The largest SuperPoD, supporting up to 1,024 NPUs with 1.72 PB/s UB Link bandwidth" - and the same single control, "Project Consulting". It states no price and no order route.
- Intended for
-
"Designed for large-scale AI data centers... trillion-parameter model training & inference" for finance/telecom/internet enterprises, plus "Education & Research" institutions listed as a served sector
Also from Huawei
-
Atlas Hardware
Huawei's line of AI training and inference processors (NPUs) and the systems built on them -- the platform brand for the silicon itself (still called Ascend in some regional markets and in the underlying chip generation names), sold standalone and in Atlas-branded servers and SuperPoD clusters, now recorded in their own separate hardware and software-stack products.
-
Atlas 650E AI Server Hardware
14U AI server powered by eight Huawei 950DT NPUs, rated at up to 12.4 PFLOPS at mxFP4, for on-premises AI training and inference in finance, government and healthcare deployments.
-
CANN Developer tool
Huawei's heterogeneous compute architecture for its Ascend/Atlas NPUs, supplying the operator libraries, compiler and programming interfaces that bridge AI frameworks to the hardware.
-
MindIE Developer tool
Inference engine and serving framework for Atlas/Ascend hardware that deploys LLM and diffusion models behind unified APIs compatible with vLLM, OpenAI and Triton interfaces.
-
MindSpore Developer tool
Open-source AI framework originated by Huawei for building, training and deploying models with native distributed training, best optimised for Huawei's Ascend/Atlas processors.
-
MindStudio Developer tool
End-to-end development toolchain for Atlas/Ascend AI applications, covering custom operator development, model conversion and compression, accuracy debugging and performance profiling via MindStudio Insight.
-
Celia Application
A voice assistant that answers questions, performs device tasks and translates content.
-
HUAWEI ADS Application
Huawei's intelligent-driving system for passenger vehicles that fuses multi-sensor input and uses a world action model to plan driving, today shipped as an assisted-driving feature with over-the-air updates.
-
Huawei AICC Application
All-in-one contact-centre solution for carriers and enterprises with voice, video and messaging channels, AI-driven service flows and a digital human, offered in Voice Basic, Omni-channel Advanced and Premier Featured editions.
Something wrong here? Send a correction — quote this product id: huawei-atlas-950-superpod.