Huawei has stated the brand new chips will strengthen the inspiration of AI computing energy, each in China and around the globe
In sum – what to know:
Huawei lays out three-year chip roadmap – Ascend 950 (2026), Ascend 960 (2027), and Ascend 970 (2028), with every launch doubling compute capability.
950 sequence delivers main upgrades – New low-precision codecs, stronger vector processing, 2 TB/s interconnect, and specialised HBMs for prefill, suggestion, coaching, and decode workloads.
Scaling via proprietary innovation – Huawei provides customized HiF knowledge codecs and HBMs, aiming to maintain efficiency regardless of sanctions and intensifying world AI demand.
SHANGHAI — Chinese language vendor Huawei unveiled its roadmap for the following technology of Ascend AI chips throughout the current Huawei Join occasion in Shanghai, China, outlining plans for 3 new chip sequence over the following three years.
“As we glance forward, a few of you is likely to be questioning about Huawei’s chip roadmap,” stated Eric Xu, Huawei’s rotating chairman, throughout his keynote speech. “We’ll hold evolving Ascend chips to strengthen the inspiration of AI computing energy, each in China and around the globe.”
The roadmap begins with the Ascend 950 household, which incorporates two chips: the Ascend 950PR, optimized for prefill inference and suggestion, and the Ascend 950DT, designed for decoding and coaching. Each chips share the Ascend 950 Die and ship enhancements in a number of dimensions, the manager stated.
The 950 sequence helps new low-precision codecs comparable to FP8, MXFP8, MXFP4, and Huawei’s proprietary HiF8. Xu famous that these enhancements will present “considerably larger coaching effectivity and inference throughput,” with the chips able to 1 PFLOPS in FP8, MXFP8, and HiF8, and a couple of PFLOPS in MXFP4.
Different enhancements embody stronger vector processing, refined reminiscence entry granularity, shrinking from 512 bytes to 128 bytes, and a 2 TB/s interconnect bandwidth—2.5 instances that of the Ascend 910C.
The Ascend 950PR will hit the market within the first quarter of 2026, accessible in card and SuperPoD server codecs. Xu defined its design responds to the rising compute wants of agent-based functions and suggestion techniques: “Our Ascend 950PR chip is designed to help these two eventualities with HiBL 1.0, our proprietary, low-cost HBM. HiBL 1.0 is less expensive than HBM3E and HBM4E.”
The Ascend 950DT, set to be launched within the final quarter of 2026, targets coaching and decoding workloads with larger calls for for reminiscence and bandwidth. It integrates HiZQ 2.0 HBM, providing 144 GB reminiscence, 4 TB/s reminiscence entry bandwidth, and a 2 TB/s interconnect.
Xu highlighted that Huawei can even introduce the Ascend 960 chip within the final quarter of 2027. In contrast with the Ascend 950 chip, it’ll double compute energy, reminiscence, and interconnect capability. The chip additionally introduces HiF4, a proprietary 4-bit precision format designed to ship larger accuracy than customary FP4 options, promising a significant increase to inference throughput.
The roadmap culminates within the Ascend 970, scheduled for the fourth 2028. Whereas nonetheless beneath improvement, Xu stated the goal is to double FP4 and FP8 compute energy in contrast with the 960, double interconnect bandwidth, and improve reminiscence bandwidth by no less than 1.5 instances. “I’m certain its efficiency can be well worth the wait,” he added.
“Usually, we are going to comply with a 1-year launch cycle and double compute with every launch. All through this course of, we are going to hold evolving our Ascend chips, making them simpler to make use of, supporting extra knowledge codecs, and rising their bandwidth. The aim is to remain on prime of ever-growing demand for AI compute,” Xu added.
“Gaps nonetheless exist between Huawei and Nvidia by way of particular person chips and ecosystems, and we’re working to shut these gaps. However we’ve got the aptitude to construct SuperPoDs and SuperClusters – that are the supply of our confidence,” Xu instructed Chinese language media.
“Huawei is totally different from firms that work on AI fashions and functions: We’re a supplier of ICT infrastructure and sensible gadgets. We’re dedicated to benefiting from our strengths to supply stable infrastructure. We search to monetize infrastructure. Our core focus is SuperPoDs and SuperClusters,” he added.
Xu additionally unveiled the corporate’s roadmap for AI computing platforms, with plans to launch what it claims to be the world’s strongest single system within the final quarter of 2026 and double the efficiency with a brand new launch a 12 months later.
Throughout Huawei Join occasion, Xu additionally unveiled the corporate’s latest SuperPoD merchandise: the Atlas 950 SuperPoD — with 8,192 Ascend chips, and the Atlas 960 SuperPoD with 15,488 Ascend chips. He defined that SuperPoDs are designed to function as single logical machines, combining a number of bodily items for large-scale AI workloads.
The manager stated that the Atlas 950 SuperPoD can be accessible within the final quarter of 2026, whereas the Atlas 960 SuperPoD will hit the market within the final quarter of 2027.
