Exploring Huawei’s DeepSeek All-in-One Machine: Achieving 60-70% Performance of the NVIDIA H100 at an Attractive Price
Huawei’s Next-Gen AI Chips and Collaboration with DeepSeek
Introduction to Huawei’s AI Innovations
Huawei is gearing up to revolutionize the AI sector with the anticipated launch of its next-generation AI chips, the Ascend 910D and 920. Recent reports suggest that the tech company has partnered with DeepSeek to enhance the capabilities of its Ascend series by integrating it with cost-effective AI models developed by DeepSeek. Sources from TechNews and 53AI indicate that this collaboration may lead to an “all-in-one machine” powered by Ascend 910B and 910C chips, significantly reducing costs by 60-70% compared to NVIDIA’s H100 chips.
The Technical Edge of Ascend Chips
GPU Capabilities and Compute Power
The upcoming integrated machine will utilize Huawei’s Ascend 910B and 910C chips. The Ascend 910B is built on a 7nm architecture, while the 910C is manufactured using SMIC’s advanced N+2 process technology. Notably, the Ascend 910C claims to deliver up to 320 TFLOPS of FP16 performance, providing around 60-70% of the processing power of NVIDIA’s H100. This performance is achieved by employing advanced integration methods that allow two 910B chips to work together effectively.
The machines also feature a distributed architecture. They can be outfitted with the Ascend 910B or 910C paired with the Kunpeng 920 CPU, and come with NVMe SSD memory—providing up to 16TB of storage per unit.
Cost Advantage of Huawei’s AI Solutions
Various Product Lines Offered
Huawei’s all-in-one machines are expected to be offered in two main categories:
- Atlas Units: These are optimized for inference, preloaded with DeepSeek’s R1 models (32B, 70B, and 671B).
- FusionCube A300 DS Edition: This version supports both training and inference, compatible with DeepSeek V3 (671B) and R1 models.
Pricing Structure
The pricing for the inference-only Atlas units ranges from approximately RMB 300,000 to 500,000 for the 32B model, whereas the high-end 671B version can cost between RMB 3 to 5 million. For units capable of both training and inference, prices start around RMB 2 million and can surge above RMB 10 million. Despite their price tags, these units still present a 60-70% cost reduction compared to NVIDIA’s H100, which retails for about RMB 20 million.
Competitive Pricing of DeepSeek Models
DeepSeek’s model pricing is notably competitive. For example, the input cost for the V3 model is just RMB 1 per million tokens, and the R1 output costs RMB 16 per million tokens—this is significantly lower than OpenAI’s pricing, which is RMB 60 per million tokens. Additionally, to support small and medium-sized enterprises, DeepSeek has introduced free versions as part of a promotional campaign.
Supply Chain and Localization Efforts in China
This initiative also highlights China’s strides in enhancing local capabilities in the tech sector. Key suppliers like SMIC, Hua Hong, and YMTC are crucial players in developing the components necessary for these all-in-one machines. This push reflects a broader movement towards self-sufficiency in technology within China.
Key Takeaways on Huawei and DeepSeek Partnership
- The collaboration aims to offer powerful AI processing capabilities at a reduced cost.
- The Ascend chips promise significant performance while maintaining a competitive edge against existing solutions.
- DeepSeek’s aggressive pricing strategy is likely to attract a broad range of users, especially smaller businesses looking for affordable AI solutions without sacrificing performance.
By leveraging innovative technology and strategic partnerships, Huawei is positioning itself to become a significant player in the AI landscape while fostering domestic supply chain growth.
