
Previously reported that NVIDIA's new China-reduced AI GPU model may be B30, with Blackwell architecture, ConnectX-8 SuperNIC connection, and PCIe 6.0 switch. It is not considered by the outside world to achieve multi-GPU expansion with NVLink. The price is between $6,500 and $8,000, far below the $10,000 to $12,000.
Wccftech reports that a certain Chinese Internet company estimates that the performance of the new China-reduced AI GPU-B30 is about 75% of H20. It has received hundreds of thousands of orders for films in late June, with a total of US$1 billion, and delivery will begin in August.
Market news, NVIDIA originally hoped to ship millions of B30 AI GPUs by the end of the year, but the production speed was a bit slow. Because GDDR7 replaces the HBM module, there are also technical downgrades, and even if the performance is degraded, there are still bottlenecks produced.
At present, the demand for AI GPUs in the Chinese market is still strong, because everyone is looking for cost-effective inference solutions. Most model training has been completed with previously purchased AI chips, such as H100, and then the requirements fall into the inference model. The B30 can also be used in small and medium-sized equipment, and for large Chinese users, new products can provide low-cost computing.
The performance of the engine-level solution constructed with B30 is standard and more cost-effective will be more attractive. The power consumption of B30 is 30% lower than that of H20 and the purchase cost is 40%. Although the other specifications of B30 are not clear, it is obvious that NVIDIA wants to comply with the US export restrictions, aiming at China's broad artificial intelligence market and constantly launching special versions to attract Chinese customers.