Ascend 950 / Atlas 950 SuperPoD by Huawei: status and timeline

Ascend 950 / Atlas 950 SuperPoD

AI accelerator competing with NVIDIA H100/B200. 950PR targets prefill; 950DT handles decode and training. Both use proprietary HBM. SuperPoD integrates 8,192 chips with optical interconnect, 1 EFLOPS FP8.

ProtoStage 2 of 5

950PR in Q1 2026; 950DT with HBM in Q4 2026. Atlas 950 SuperPoD (8,192 NPUs) revealed at MWC 2026.

Updated Mar 2026Checked 11 Oct0 updates this week

Milestones

Next · Ascend 950PR mass production begins

Ascend 950 roadmap announced at HC 2025Sep 2025Complete.
Ascend 950PR mass production beginsTarget Q1 2026Current milestone.
Ascend 950DT with HiZQ 2.0 HBM launchesTarget Q4 2026Not yet reached.
Atlas 950 SuperPoD deployment in South KoreaTarget Q4 2026Not yet reached.

Most important updates

Upcoming

Q1 2026Ascend 950PR mass production begins (next)
Q4 2026Ascend 950DT with HiZQ 2.0 HBM launches
Q4 2026Atlas 950 SuperPoD deployment in South Korea
Q4 2026Ascend 950DT enters production; Atlas 950 SuperPoD ships to early customers

Current obstacles

US export controls on critical technologyRestrictions on design tools, chipmaking equipment and memory supplies. Must achieve full domestic supply-chain independence.
Software ecosystem adoption vs CUDARequires porting models, retraining, rewriting kernels. Early customers (DeepSeek, Baidu) proving viability.

Physics limits

Memory bandwidth vs computeEven 4 TB/s on the 950DT is finite. Very large models or small batch sizes can still be bandwidth-bound. Optical interconnect helps but cannot overcome physics.
Design-tool ecosystemBuilding chips without access to US EDA tools (Cadence, Synopsys) requires reverse-engineering or domestic replacements. Both are slower than the incumbent tool chain that NVIDIA uses.

How it works

3 parts

Prefill vs Decode

Separate chips for LLM inference stages

950PR handles prefill (prompt tokens), 950DT handles decode (output streaming). Each optimizes for its workload's compute and memory patterns.

HBM Integration

Proprietary memory stack

Huawei packages HBM with each die (HiBL on 950PR, HiZQ on 950DT), cutting cost and supply-chain risk. 950DT reaches 4 TB/s.

Optical Cluster Network

Microsecond interconnect at scale

4,096 optical transceivers (800G) connect 8,192 NPUs with 3-microsecond inter-NPU latency, enabling tight tensor-parallel training at cluster scale.

Spec sheet

Spec	Ascend 950 / Atlas 950 SuperPoD
Peak compute (FP8)	1 PFLOPS^R (reported)
Peak compute (FP4)	2 PFLOPS^R (reported)
Ascend 950PR memory (HiBL 1.0)	128 GB @ 1.6 TB/s^R (reported)
Ascend 950DT memory (HiZQ 2.0)	144 GB @ 4 TB/s^R (reported)
Atlas 950 SuperPoD NPU count	8,192 Ascend 950DT^R (reported)
Atlas 950 SuperPoD unified memory	256 TB^R (reported)
Inter-NPU RTT latency	3 microseconds^R (reported)

^R reported by the company

Update log

2 updates

Mar 2026

Major: Huawei unveils Atlas 950 SuperPoD at Mobile World Congress with 8,192 Ascend 950DT NPUs and optical interconnectPressHardware

Sep 2025

Major: Huawei announces three-year roadmap for Ascend 950, 960 and 970 chips with annual release cyclePressHardware

About Huawei

The team behind Ascend 950 / Atlas 950 SuperPoD

Huawei

Ascend NPU, Atlas SuperPoD, telecommunications

Huawei designs semiconductors, cloud systems and telecoms gear in 170 countries. Founded 1987, it grew to lead telecoms infrastructure. Now competing directly with NVIDIA in AI accelerators with the Ascend chip series.

Operates in 170+ countries across cloud, AI, telecoms, consumer devices and transport systems.
Ascend roadmap: 950/960/970 series at one-year cycles, each doubling compute. 950PR Q1 2026, 950DT Q4 with proprietary HBM.
Atlas 950 SuperPoD (8,192 NPUs) delivers triple H20 inference performance at quarter the cost, shown at MWC 2026.

huawei.com (opens huawei.com)Company profile →

Founded: 198739 yrs
Headquarters: China
Status: Private
Valuation: private
Staff: ~213,000Jan 2026
Works in: ComputingAI accelerators & CPUsVehicles
Coverage: 4 programs · 10 updateslatest 10d agochecked 10 Oct
Customers: DeepSeek (opens deepseek.com)BaiduAlibaba
Suppliers: SMIC
People: Ren Zhengfei (opens en.wikipedia.org)Founder and CEO

Spec

Ascend 950 / Atlas 950 SuperPoD

Peak compute (FP8)

1 PFLOPS^R (reported)

Peak compute (FP4)

2 PFLOPS^R (reported)

Ascend 950PR memory (HiBL 1.0)

128 GB @ 1.6 TB/s^R (reported)

Ascend 950DT memory (HiZQ 2.0)

144 GB @ 4 TB/s^R (reported)

Atlas 950 SuperPoD NPU count

8,192 Ascend 950DT^R (reported)

Atlas 950 SuperPoD unified memory

256 TB^R (reported)

Inter-NPU RTT latency

3 microseconds^R (reported)