"As we've previously written, US export rules mean Nvidia must limit the card to approximately 581 teraFLOPS of FP4 and 1.4TB/s of memory bandwidth. While faster at 4-bit precision, the card's performance will be roughly equivalent with an H20 at high precision with less than half the bandwidth. The lack of high-speed chip-to-chip interconnectivity also limits the unit's usefulness for AI training."
"Despite the performance limits, the FT reports that Chinese companies were lining up to purchase tens of thousands of the Blackwell-based accelerators and had even begun working with server suppliers to validate them for deployment, presumably for use in inference or model fine-tuning. The ban comes just weeks after Chinese authorities issued letters to top Chinese tech firms discouraging the use of Nvidia's H20 accelerators, particularly for sensitive government workloads."
China's Cyberspace Administration directed top technology companies, including Alibaba and ByteDance, to halt testing and cancel orders of Nvidia's RTX Pro 6000D Blackwell GPUs. The Blackwell card is intentionally limited to about 581 teraFLOPS FP4 and 1.4TB/s memory bandwidth to comply with US export rules, offering stronger 4-bit performance but roughly H20-equivalent at high precision and lacking high-speed chip-to-chip interconnects needed for large-scale training. Chinese firms had been preparing to buy and validate tens of thousands of the accelerators for inference and fine-tuning. The move follows recent Chinese discouragement of H20 use and previous US export restrictions and negotiated resumptions.
Read at Theregister
Unable to calculate read time
Collection
[
|
...
]