Huawei Ascend 950DT vs Huawei Ascend 950PR

The Ascend 950DT delivers 3% more FP8 throughput than the Ascend 950PR (1,034 vs 1,000 TFLOPS dense). The Ascend 950DT also carries 16 GB more memory (144 GB vs 128 GB).

Ascend 950DT: Da Vinci v3, 2026 Ascend 950PR: Da Vinci v3, 2026

FP8 dense lead

+3%

Ascend 950DT: 1,034 vs 1,000 TFLOPS

Memory

144 vs 128 GB

16 GB more for the Ascend 950DT

Bandwidth

4.0 TB/s vs 1.6 TB/s

Ascend 950DT moves data faster

Download comparison sheet Printable version →

Huawei Ascend 950DT

Leads on FP8 Full specs →

What is the Huawei Ascend 950DT?

The Ascend 950DT is an NPU from Huawei on the Da Vinci v3 architecture, an accelerator built for AI inference. It is an OAM module, the open accelerator format used in multi-GPU servers. Ascend 950DT full specifications →

Can I rent the Ascend 950DT in the cloud?

Flopper is not tracking any live rental listings for the Ascend 950DT yet. The provider directory lists the clouds that rent comparable accelerators on demand. GPU cloud providers →

How does the Ascend 950DT compare with other accelerators?

This page sets the Ascend 950DT against the Ascend 950PR. Both parts launched in the same year. Both are Huawei parts, and both ship as an OAM module, so they fit the same class of system. The compare tool builds the same comparison for any two parts in the catalogue. All GPU comparisons →

Where do the Ascend 950DT figures come from?

Every Ascend 950DT figure on this page traces to Huawei's own datasheet or specification table. How Flopper verifies specifications →

Huawei Ascend 950PR

Full specs →

What is the Huawei Ascend 950PR?

The Ascend 950PR is an NPU from Huawei on the Da Vinci v3 architecture, an accelerator built for AI inference. It is an OAM module, the open accelerator format used in multi-GPU servers. Ascend 950PR full specifications →

Can I rent the Ascend 950PR in the cloud?

Flopper is not tracking any live rental listings for the Ascend 950PR yet. The provider directory lists the clouds that rent comparable accelerators on demand. GPU cloud providers →

How does the Ascend 950PR compare with other accelerators?

This page sets the Ascend 950PR against the Ascend 950DT. Both parts launched in the same year. Both are Huawei parts, and both ship as an OAM module, so they fit the same class of system. The compare tool builds the same comparison for any two parts in the catalogue. All GPU comparisons →

Where do the Ascend 950PR figures come from?

Huawei published the Ascend 950PR figures in marketing material rather than a datasheet, and Flopper has no independent confirmation of them. How Flopper verifies specifications →

Specifications

Ascend 950DT

Ascend 950PR

Architecture

Da Vinci v3Da Vinci v3

Launch Year

20262026

Form Factor

OAMOAM

Memory

144 GB128 GB

Memory Bandwidth

4.0 TB/s1.6 TB/s

TDP

—900 W

Process Node

——

Spec

Ascend 950DT

Ascend 950PR

Architecture Da Vinci v3Da Vinci v3

Launch Year 20262026

Form Factor OAMOAM

Memory 144 GB128 GB

Memory Bandwidth 4.0 TB/s1.6 TB/s

TDP —900 W

Process Node ——

Performance (TFLOPS)

Ascend 950DT

Ascend 950PR

30 TFLOPS

273 TFLOPS

547 TFLOPS 500 TFLOPS

1,034 TFLOPS 1,000 TFLOPS

FP6

2,007 TFLOPS 2,000 TFLOPS

1,034 TOPS

Precision

Ascend 950DT

Ascend 950PR

30 TFLOPS

273 TFLOPS

547 TFLOPS 500 TFLOPS

1,034 TFLOPS 1,000 TFLOPS

FP6

2,007 TFLOPS 2,000 TFLOPS

1,034 TOPS

Figures are dense. Where a vendor publishes a figure it sits under the dense one and never decides a winner. Hover any precision for its definition, or read the GPU glossary.

FLOPS by Precision

What actually differs

The suffix names a workload phase, not a speed grade. Huawei states near-identical compute for both chips, about 1,000 TFLOPS FP8 dense, and splits them on memory. The 950PR serves prefill and recommendation, which are compute-bound, so it pairs 128 GB of Huawei’s own HiBL 1.0 with 1.6 TB/s. The 950DT serves decode and training, where streaming weights sets the pace, so it carries 4.0 TB/s.

Read the 950DT as the general-purpose part and the 950PR as a prefill tier for disaggregated serving fleets. Certainty differs too: the 950PR figures come from keynote disclosure, the 950DT figures from Huawei’s chip specification page.

The trap is modelling the catalog row. The 950PR ships as the Atlas 350, a de-rated bin at 112 GB and 1.4 TB/s, and Huawei’s 950DT systems specify 96 GB. For a reference point outside China, see the 950DT against the B200.

New to the terms? The glossary explains FP32, TF32, BF16, FP16, FP8, FP4, INT8, Sparsity and TFLOPS.

Frequently asked questions

What do PR and DT mean on the Huawei Ascend 950?

They name the workload phase each chip is built for. PR is prefill and recommendation, compute-bound work that tolerates slower memory. DT is decode and training, where moving weights and KV cache sets the pace. Huawei states near-identical compute for both and separates them on memory and interconnect.

Is the Ascend 950DT faster than the 950PR?

On compute, no: Huawei states 1,034 TFLOPS FP8 for the 950DT and 1,000 TFLOPS FP8 for the 950PR, both dense. On memory, yes: 4.0 TB/s against 1.6 TB/s. Expect the 950DT to lead wherever bandwidth is the limit, which includes LLM decode and most training.

How much memory do the Ascend 950PR and 950DT have?

Huawei lists the 950PR at 128 GB of HiBL 1.0 and the 950DT at 144 GB of HiZQ 2.0, with a 96 GB configuration too. Shipping products sit lower: the Atlas 350 card carries 112 GB, and Huawei's 950DT systems specify 96 GB.

Can you buy or rent the Ascend 950PR or 950DT?

Both are sold in mainland China only, the 950PR as the Atlas 350 card and the 950DT inside Atlas 650E servers and the Atlas 950 SuperPoD. Ascend silicon is export-controlled, so no provider tracked by Flopper rents either chip.

Get Comparison Updates

New GPUs added weekly. Be the first to see how they compare.

Open in Full Compare Tool → Ascend 950DT full specs Ascend 950PR full specs Estimate cluster cost in the calculator →

Huawei Ascend 950DT vs Huawei Ascend 950PR

Huawei Ascend 950DT

What is the Huawei Ascend 950DT?

Can I rent the Ascend 950DT in the cloud?

How does the Ascend 950DT compare with other accelerators?

Where do the Ascend 950DT figures come from?

Huawei Ascend 950PR

What is the Huawei Ascend 950PR?

Can I rent the Ascend 950PR in the cloud?

How does the Ascend 950PR compare with other accelerators?

Where do the Ascend 950PR figures come from?

Specifications

Performance (TFLOPS)

FLOPS by Precision

What actually differs

Frequently asked questions

What do PR and DT mean on the Huawei Ascend 950?

Is the Ascend 950DT faster than the 950PR?

How much memory do the Ascend 950PR and 950DT have?

Can you buy or rent the Ascend 950PR or 950DT?

Get Comparison Updates

Stay Updated on GPU Specs