Sold in mainland China only. The Atlas 350 is distributed through Huawei's Chinese channel and partner OEMs and is not sold, hosted or rentable outside China, so no cloud provider on this site lists it. Huawei is on the US Entity List and Ascend silicon is treated as export-controlled wherever it is located. Huawei has also published no datasheet, product brief or catalogue entry for this card, and it does not appear in either of Huawei's public Ascend product catalogues: every figure here comes from a Huawei executive's launch remarks relayed by Chinese press.
HU

Huawei Atlas 350 NEW

Da Vinci v3 Custom 2026
VRAM
112 GB
HiBL 1.0
TDP
600 W
Bandwidth
1.4 TB/s
memory

Overview

The Huawei Atlas 350 is the accelerator card built on the Ascend 950PR processor, announced by Huawei VP Ma Haixu at Huawei China Partner Conference 2026 and formally on sale from 21 March 2026. It is the first Atlas card to break the long-running 300-series numbering, reflecting a new silicon generation rather than a refresh. The card is a de-rated bin of the 950PR chip rather than a straight repackaging: it carries 112 GB of HiBL 1.0 at 1.4 TB/s against the chip's 128 GB at 1.6 TB/s, exactly seven eighths of both, the signature of one memory stack fused off out of eight, and it delivers 1.56 PFLOPS of dense MXFP4 against the chip's 2 PFLOPS, a further shortfall that implies a clock or core bin on top. HiBL 1.0 is Huawei's own in-house high-bandwidth memory, developed as a lower-cost alternative to HBM3E and HBM4E and aimed at the prefill stage of inference and at recommendation workloads, which are compute-heavy and memory-light. Power is 600W, one and a half times the 400W of NVIDIA's H20, and capacity is 1.16 times the 96 GB H20, the comparison Huawei itself drew at launch. Seven Chinese server OEMs launched systems around the card, mounting up to eight of them in a 6U two-socket Kunpeng 920 chassis. Huawei has published no datasheet for the Atlas 350, no FP8 or FP16 figure, no process node and no transistor count, and its separate claim of 2.87 times the single-card compute of an H20 has no stated basis and cannot be reconciled with any published pair of numbers. The underlying silicon is catalogued separately on this site as the Ascend 950PR.

Performance Metrics

Peak theoretical throughput by precision type

PrecisionBitsPeak TFLOPSEfficiency
FP6 -- 804.0 1.340 TFLOPS/W
INT8 8 804.0 1.340 TFLOPS/W
FP8 8 804.0 1.340 TFLOPS/W
FP4 4 1561.0 2.602 TFLOPS/W
FP16 16 425.0 0.708 TFLOPS/W
BF16 16 425.0 0.708 TFLOPS/W
FP8 Efficiency
1.340 TFLOPS/W
804.0 TFLOPS / 600W
FP16 Efficiency
0.708 TFLOPS/W
425.0 TFLOPS / 600W

Power Specifications

TDP

600 W

Max Power

690 W

Power Connector

PCIe 16-pin

Cooling

Air

Memory Specifications

Capacity

112 GB

Type

HiBL 1.0

Bandwidth

1400 GB/s

Interface

--

Hardware & Design

Form Factor

Custom

Architecture

Da Vinci v3

Process Node

--

Launch Year

2026

Variant

Standard

Market Segment

Professional

Full Specifications

Memory
VRAM 112 GB
Memory Type HiBL 1.0
Bandwidth 1.4 TB/s
Interconnect & I/O
GPU-to-GPU Unified Bus 2.0
Power & Thermal
TDP 600 W
Enterprise Features
Compute APIs CANN, MindSpore, PyTorch (torch_npu), vLLM-Ascend
General
Form Factor Custom
Architecture Da Vinci v3
Launch Year 2026

Documentation & Resources

Common Use Cases

General Compute AI/ML Workloads Data Processing

The Huawei Atlas 350 is optimized for high-performance computing tasks with Da Vinci v3 architecture delivering high TFLOPS of compute power.

Where to Rent

Compare cloud providers offering on-demand GPU instances for AI training, inference, and HPC workloads.

Browse GPU Cloud Providers

Similar GPUs

Stay Updated on GPU Releases

Get notified when new GPUs are added or specifications are updated.

Loading verification...

No spam, unsubscribe anytime.

Back to GPUs