"First off the rank, in the first quarter of 2026, will be the Ascend 950PR which, according to slideware shown at the conference, will boast one petaflop performance with the 8-bit floating-point (FP8) computation units used for many AI inferencing workloads. The chip will also include 2 TB/s interconnect bandwidth and 128GB of 1.6 TB/s memory. In 2026's final quarter Huawei plans to deliver the 950DT, which will be capable of two petaflops of FP4 performance thanks to the inclusion of 144GB of 4 TB/s memory."
"As The Register has previously reported, users of Huawei's existing accelerators have reported results that suggest the parts don't always perform brilliantly or match claims they are equivalent to kit from the likes of Nvidia and AMD. Such comparisons stopped mattering on Wednesday, when China forbade local companies from shopping for American accelerators. Now that Chinese buyers can't acquire chips from offshore, they'll have to learn to live with Huawei's wares."
Huawei announced a multi-year roadmap of Ascend AI accelerators delivering rising compute and memory bandwidth from 2026 through 2028. The Ascend 950PR in Q1 2026 targets one petaflop FP8 performance with 2 TB/s interconnect and 128GB of 1.6 TB/s memory. The 950DT in Q4 2026 aims for two petaflops FP4 with 144GB of 4 TB/s memory. The Ascend 960 in 2027 will include 288GB of 9.6 TB/s memory, and the Ascend 970 in 2028 will reach 14.4 TB/s memory bandwidth. Domestic high-bandwidth memory sourcing and Chinese restrictions on American accelerators make Huawei's chips central to local AI deployment, despite mixed performance reports against Nvidia and AMD.
Read at Theregister
Unable to calculate read time
Collection
[
|
...
]