Huawei's Bold Leap: New Ascend Chips Set to Revolutionize AI Clusters and Beyond
Huawei is making waves in the tech world with its fresh designs for the next iteration of its Ascend chip line. The announcement came during the Huawei Connect 2025 event in Shanghai, stirring excitement among AI enthusiasts and tech aficionados alike.
Eric Xu, the deputy chairman of Huawei's board, took to the stage to share his insights on this pivotal year for the company. Highlighting the introduction of DeepSeek-R1 earlier this year, he hinted at a transformative future, despite acknowledging that China's semiconductor manufacturing may still lag behind global standards for the foreseeable future.
In response to ongoing international trade tensions, Huawei plans to enhance its technology infrastructure and has made a bold decision to open-source significant portions of its software. This includes the openPangu AI models and the Mind series SDKs, demonstrating a commitment to innovation rooted in collaboration.
Meet the New Ascend Chips!
Huawei is gearing up to roll out three new series of the Ascend chips—the 950, 960, and the 970 models. The Ascend 950PR and 950TO, crafted from the same die, will cater to low-precision data formats like FP8. Remarkably, the 950 is projected to achieve one petaflop (PFLOP) of processing power, while its companion, the MXFP8, is expected to reach up to two petaflops. Just to clarify, one PFLOP equals a staggering one thousand trillion floating-point calculations each second!
The enhancements don't stop there. These chips promise improved vector processing capabilities and finer memory access, breaking down memory operations into 128-byte chunks as opposed to the previous 512 bytes. And if that’s not enough, the Ascend 950 will boast a whopping 2 terabytes per second interconnect bandwidth, which is 2.5 times more than the current Ascend 910C. Mark your calendars—the 950PR is set for launch in Q1 2026, closely followed by the Ascend 950DT in Q4 2026.
Looking ahead, the Ascend 960 is on schedule for Q4 2027 and aims to deliver double the computing prowess, memory access, and capacity compared to the 950 series. This series will support Huawei's own HiF4 data format, which promises enhanced precision beyond existing technologies.
The crème de la crème, however, is the Ascend 970, anticipated for Q4 2028. While Xu remains coy about its exact specifications, he emphasizes an ambition to elevate its performance metrics significantly. This chip is expected to achieve interconnect speeds of 4 TB/s and around 8 PFLOPS of FP4 processing, along with a substantial memory upgrade.
The Advent of SuperPods
Huawei's vision extends further into the realm of hyperscaler clusters with the introduction of SuperPods. Starting late 2026, the Atlas 950 SuperPod will harness the power of the new Ascend 950DT chips. Competing directly with NVIDIA’s NVL144 system, Huawei claims its SuperPod will feature a staggering 56.8 times more NPUs compared to the NVL144's GPUs, offering nearly sevenfold the processing capabilities. That’s some serious power!
General Computing Pursuits
But that’s not all! For general computing needs, Huawei is prepared to unveil two Kunpeng 950 processors by Q1 2026. These will range from 96 cores with 192 threads to an impressive 192 cores and 384 threads in the high-performance model. There’s also the TaiShan 950 SuperPod, touted as the world’s first general-purpose computing SuperPod, hitting the market in early 2026.
Open-Source Connectivity for the Future
Both the NPU and general computing SuperPods will utilize the upcoming UnifiedBus 2.0—an upgraded version of the existing technology currently in use. The exciting part is that UnifiedBus 2.0 will be an open protocol, allowing developers immediate access to its tech specifications, fostering innovation and growth at a community level.
First off the block with this new connectivity will be the Atlas 950 SuperCluster, boasting 2.5 times more NPUs and 1.3 times more computing power than xAI’s Colossus, which is currently recognized as the globe's most formidable computing cluster. By late 2027, Huawei aims to launch the Atlas 960 SuperCluster, which will integrate over a million NPUs, delivering 4 ZFLOPS in FP4. For some context, that’s a mind-boggling \(10^{21}\) floating-point operations per second!
So, folks, whether you're a tech enthusiast or a seasoned professional, one thing is crystal clear: Huawei's ambitious plans are set to redefine not just AI computing, but the entire landscape for future technologies!
