News
Amid rumors that Huawei is set to launch next-gen AI chips like the Ascend 910D and 920, the tech giant has reportedly teamed up with DeepSeek to integrate its Ascend series with DeepSeek’s budget-friendly AI models. According to TechNews, citing Chinese media outlet 53AI, an “all-in-one machine” powered by Ascend 910B and 910C could slash costs by 60-70% compared to NVIDIA’s H100. Here’s a quick look at what makes this machine stand out.
GPU Capabilities/ Compute Power
The integrated machine is said to be equipped with Huawei’s Ascend 910B (built on 7nm) and 910C (built on SMIC’s N+2 process). The reports suggest that Ascend 910C could deliver up to 320 TFLOPS of FP16 performance, achieving 60-70% of the performance of NVIDIA’s H100.
As per Reuters, Ascend 910C matches NVIDIA’s H100 performance by packaging two 910B chips together using advanced integration methods.
Meanwhile, 53AI reveals that the machine features a distributed architecture, with either the Ascend 910B or 910C paired with the Kunpeng 920 CPU. It has NVMe SSD memory, with up to 16TB of storage per unit, the report indicates.
Cost/ Pricing Advantage
According to the reports, the all-in-one machines come in two main product lines: the inference-focused Atlas, preloaded with DeepSeek’s R1 models (32B, 70B, and 671B), and the FusionCube A300 DS Edition, designed for both training and inference, supporting DeepSeek V3 (671B) and R1 models.
As for pricing, the inference-only Atlas units start at around RMB 300,000 to 500,000 for the 32B model, with the high-end 671B version ranging from RMB 3 to 5 million.
For machines handling both training and inference, prices start around RMB 2 million and can top RMB 10 million. While not cheap, they still offer a 60–70% cost saving compared to NVIDIA’s H100 solution, which goes for about RMB 20 million, the reports note.
In terms of API pricing, the reports indicate that DeepSeek’s V3 model input is priced at just RMB 1 per million tokens and R1 model output at RMB 16 per million tokens—far lower than OpenAI’s RMB 60 per minion tokens. To boost adoption, especially among small and medium-sized enterprises (SMEs), DeepSeek is also offering free versions as part of its initial promotion, the reports add.
Notably, the all-in-one machine reflects China’s progress in domestic localization, with key players like SMIC, Hua Hong, and YMTC playing important roles in the supply chain, as per 53AI.
Read more
(Photo credit: Huawei)
