Huawei Ascend 950PR
Huawei's inference and recommendation chip on the new Ascend 950 die: 1 PFLOPS FP8 and 2 PFLOPS MXFP4, with Huawei's own low-cost HiBL 1.0 HBM. Shipping in limited volumes in 2026, and Huawei says it sells every unit it can make. Sources and verification dates.
FP81,000 TFLOPS
FP42,000 TFLOPS
Every published figure
Source per value- Owner The company that specified the chip and runs it in its own cloud, data centers or products.
- Huawei 1
- Role What the owner says the chip is for: training, inference, both, or a general-purpose server CPU.
- Inference 1
- Design partner The ASIC house that co-designed or physically implemented the chip (Broadcom, Marvell, MediaTek, Alchip, GUC and others), as stated by the owner, the partner, a filing or named trade-press reporting. 'In-house' when the owner states it designed the chip itself, including through an acquired team such as Annapurna Labs.
- In-house (HiSilicon) 1
- Memory On-package memory type (HBM2e, HBM3, HBM3E, LPDDR5) or, for CPUs, the memory the socket drives (DDR5).
- HiBL 1.0 (Huawei in-house HBM) 1
- FP8 TFLOPS Published peak FP8 throughput per chip in TFLOPS. The precision note records sparsity where the owner states it. Blank when the owner gives no FP8 figure.
- 1,000 1
- FP4 TFLOPS Published FP4 throughput per chip or accelerator module in TFLOPS. MXFP4 and MX4 have separate columns; see the precision note for scope.
- 2,000 1
- Precision note What the throughput figures mean: dense or sparse, which precision, per chip or per package, and any figure the owner states only in another unit.
- Huawei states 1 PFLOPS in FP8, MXFP8 and HiF8, and 2 PFLOPS in MXFP4, per chip. 1
- Status Announced: public roadmap without verified deployment. Sampling: first silicon or limited internal use. Preview: customers can request early access, without a GA claim. Deployed: stated production use or general availability. Discontinued: program stopped.
- Deployed 2
- Announced Year the owner first announced the chip publicly.
- 2,025 1
- Deployment scale The largest deployment the owner or a named report states: chip counts, clusters, gigawatts or regions.
- Available in limited volumes; Huawei says it is selling everything it can produce (October 2026). 2
- How to get it How an outside buyer can use the chip: a named cloud instance or service, a product, or 'Internal only'.
- Huawei computing cards and SuperPoD servers 1
Not verified in the sources reviewed: partner basis, foundry, node, memory capacity, memory bandwidth, bf16, fp16, mxfp8, mxfp4, mx8, mx4, int8, power, largest pod.
Sources
- Huawei: Eric Xu keynote at Huawei Connect 2025 accessed 2026-11-11
- TrendForce: Huawei says Ascend market share tops NVIDIA in China accessed 2026-11-11