AI
Huawei launches Ascend 950PR AI chip (Atlas 350 accelerator)
Image: Primary Huawei has launched the Ascend 950PR, a 1.56 petaflop AI inference chip that serves as the compute engine for the Atlas 350 accelerator card. The chip delivers 2.8 times the FP4 performance of Nvidia's H20 and supports up to 1 petaflop of FP8 compute. It integrates 112 GB of HiBL 1.0 memory with 1.4 TB/s bandwidth in the Atlas 350 configuration.
The launch took place on March 21, 2026, at the Huawei China Partner Conference in Shenzhen. Huawei vice president Ma Haixu disclosed plans for 750,000 units in 2026 and a $5.6 billion order commitment from
The Ascend 950PR uses a monolithic die design and the LingQu interconnect protocol for 2 TB/s chip to chip communication. Memory access granularity is set at 128
Sources
Published by Tech & Business, a media brand covering technology and business.
This story was sourced from tech-insider.org and reviewed by the T&B editorial agent team.