At the recent Huawei Connect 2025 conference held in Shanghai, Chinese tech powerhouse Huawei unveiled its ambitious roadmap for the upcoming iterations of its Ascend chip family.
Huawei’s Vision for Next-Gen Ascend Chips
During his keynote address, Huawei’s deputy board chair Eric Xu reflected on 2025 as a landmark year, highlighting the launch of the DeepSeek-R1 in January as a pivotal milestone. Xu candidly acknowledged that China’s semiconductor manufacturing capabilities will likely trail behind global leaders in process node advancements for an extended period.
In response to ongoing trade restrictions and tariffs, Huawei is doubling down on innovation in infrastructure and chip design. Notably, the company has embraced open-source strategies by releasing significant portions of its software ecosystem, including the openPangu foundational AI models and the Mind series software development kits (SDKs), fostering broader collaboration and adoption.
Introducing the Ascend 950, 960, and 970 Series
Huawei plans to launch three new Ascend chip lines: the 950, 960, and 970, each progressively more powerful and feature-rich.
The Ascend 950 series, including the 950PR and 950TO variants, share the same silicon die but enhance support for low-precision data types such as FP8. The 950 delivers a peak performance of 1 PFLOP (one quadrillion floating-point operations per second), while the MXFP8 variant doubles that to 2 PFLOPs. Improvements also include finer-grained memory access, reducing chunk sizes from 512 bytes to 128 bytes, and enhanced vector processing capabilities.
Interconnect bandwidth for the 950 chips reaches 2 TB/s, a 2.5-fold increase over the current Ascend 910C. The 950PR is slated for release in Q1 2026, followed by the 950DT in Q4 2026.
Building on this, the Ascend 960, expected in Q4 2027, will double the computing power, memory bandwidth, capacity, and interconnect ports compared to the 950. It will also introduce Huawei’s proprietary HiF4 data format, which the company claims offers superior precision relative to existing FP4 standards.
The flagship Ascend 970, targeted for Q4 2028, aims to push performance boundaries further. While specifications are still being finalized, Xu indicated ambitions for 4 TB/s interconnect bandwidth, 8 PFLOPs of FP4 compute power, and expanded memory capacity.
SuperPoDs: Scalable NPU Clusters for Hyperscale Computing
Huawei’s strategy centers on delivering massive compute clusters called SuperPoDs, designed to meet the demands of hyperscale data centers. The Atlas 950 SuperPoD, powered by Ascend 950DT chips, will debut in Q4 2026.
Comparatively, NVIDIA’s NVL144 SuperPod, launching mid-to-late 2026, will be outmatched by Huawei’s offering, which boasts 56.8 times more NPUs than GPUs in the NVL144 and nearly sevenfold greater processing throughput. Even with NVIDIA’s planned NVL576 release in 2027, Huawei’s Atlas 950 SuperPoD is projected to maintain superior performance.
Kunpeng 950: High-Core-Count General-Purpose Processors
For broader computing needs, Huawei will introduce two Kunpeng 950 processor models in Q1 2026. These chips will feature 96 cores with 192 threads and a higher-end variant with 192 cores and 384 threads. Complementing these processors, Huawei will launch the TaiShan 950 SuperPod, described by Xu as the world’s first general-purpose computing SuperPoD, also arriving in early 2026.
UnifiedBus 2.0: Open-Source Interconnect for SuperPods and SuperClusters
Both NPU and general computing SuperPoDs will utilize UnifiedBus 2.0, the successor to the UnifiedBus 1.0 protocol that underpins the Atlas 900 A3 SuperPoD, which has seen over 300 deployments since its March 2024 launch.
UnifiedBus 2.0 will be released as an open standard, with full technical documentation made immediately available to developers. This protocol will not only interconnect individual SuperPoDs but also enable the formation of larger SuperClusters.
The first such cluster, the Atlas 950 SuperCluster, will deliver 2.5 times more NPUs and 1.3 times greater compute power than xAI’s Colossus, currently the world’s most powerful computing cluster. Looking ahead to Q4 2027, Huawei plans to release the Atlas 960 SuperCluster, integrating over one million NPUs and achieving 4 ZFLOPS of FP4 performance (where a ZFLOP equals 10^21 floating-point operations per second).
Xu emphasized that these SuperPoDs and SuperClusters, powered by UnifiedBus, represent Huawei’s solution to the escalating global demand for high-performance computing resources, both now and in the future.
