EmbeddingGemma 2: an open, lightweight multimodal embedding model
deepmind.google·6d ago
AI Engineering
Lobsters·Aug 29·5 min read
What GLM-5.3 Flash running on Chinese hardware actually means
TL;DR
Z.AI's latest model, GLM-5.3, runs entirely on Chinese hardware, specifically HiSilicon chips, showcasing China's advancements in AI technology. However, Western companies still maintain a significant competitive edge.
✦ Why It Matters
Evaluate your hardware choices and consider the implications of using Chinese AI technology in your projects.
Key Takeaways
- Assess the performance of HiSilicon chips against Nvidia's offerings.
- Monitor advancements in EUV technology for future hardware implications.
- Consider the cost-effectiveness of using Chinese hardware for AI inference.
- Stay informed about the competitive landscape in AI hardware development.
Full Summary
Z.AI has confirmed that its GLM-5.3 Flash model operates entirely on Chinese-manufactured hardware, likely using HiSilicon's 910c series chips. These chips achieve approximately 1.6 PFLOP/s of INT8 compute but lag behind Nvidia's H100, which is four years old.
The upcoming 950-series offers limited improvements and faces constraints due to the lack of advanced EUV (extreme ultraviolet) fabrication technology, which is crucial for producing high-performance chips. While China aims to compensate for hardware limitations by increasing production volume, this approach results in slower inference speeds and higher operational costs.
Without breakthroughs in fabrication technology, the performance gap between Chinese and Western AI hardware is expected to widen.