Cat and mouse: Nvidia's B40 to circumvent China ban
A new GPU accelerator from Nvidia is set to find its way to China. It could be similar to the GeForce RTX 5090.
Render image of a GB202 GPU from Nvidia. Unlike the GB200 for the large Blackwell accelerators, it has no HBM on the carrier.
(Image: Nvidia)
Nvidia is reportedly planning to mass-produce a new GPU accelerator for Chinese customers as early as June. The newcomer is to be called B40 or RTX Pro 6000D and will circumvent the latest US export restrictions. The latter most recently affected the H20 from Nvidia's last generation Hopper and AMD's Instinct MI308.
The news agency Reuters reports on a new GPU accelerator for China. It cites internal sources and the Chinese securities company GF Securities. According to these sources, the newcomer is similar to the similarly named RTX Pro 6000, which in turn is a professional sister model of the GeForce RTX 5090 for workstations and servers.
GDDR7 instead of HBM
Most recently, the US government made GPU exports dependent on the transfer rate of the memory and the interconnect speed between several GPUs. Large and fast memory is necessary for training large AI models. A fast interconnect helps to distribute the training across numerous GPUs.
Due to the restrictions, Nvidia reportedly dispenses with fast memory of the high-bandwidth memory (HBM) type and instead relies on GDDR7 components. The memory should be able to transfer 1.7 to 1.8 terabytes per second. For comparison: The high-end GPUs B200 and B300 achieve 8 TByte/s with HBM3(e).
The RTX Pro 6000 and GeForce RTX 5090 with GB202 chip and GDDR7 memory on a 512-bit interface can also achieve the 1.7-1.8 TByte/s mentioned above. Up to 96 GB of RAM is possible on one card. Considering the rumored near-term series production, Nvidia may not have had time to launch a completely new chip for China. The use of the GB202 graphics chip seems logical.
The fast NVLink interconnect would be completely absent here; several GPUs would have to communicate with each other via comparatively slow PCI Express. If the USA wants to enforce speed limits from Biden's time in office, Nvidia would also have to throttle the GPU. Recently, however, computing power has hardly been an issue when it comes to export restrictions. Under President Donald Trump, the USA no longer seems to want to enforce the quantity limits on AI chips to many countries.
Videos by heise
Own China chip could follow
Nvidia is reportedly targeting a price of between 6500 and 8000 US dollars for the B40 and RTX Pro 6000D. The H20 is said to have cost between 10,000 and 12,000 US dollars. High-end models such as the B200 and B300 are many times pricier.
According to Reuters, Nvidia is also currently designing a new chip for the Chinese market. Such a chip could save money in production in the long term if it reduces unnecessary parts such as 3D functions, thereby reducing the space required. Even with AI accelerators such as the B300, only a small proportion of the shader cores can render 3D scenes.
However, a chip designed specifically for China seems risky if the USA changes its export restrictions again and bans the sale of corresponding models.
Empfohlener redaktioneller Inhalt
Mit Ihrer Zustimmung wird hier ein externer Preisvergleich (heise Preisvergleich) geladen.
Ich bin damit einverstanden, dass mir externe Inhalte angezeigt werden. Damit können personenbezogene Daten an Drittplattformen (heise Preisvergleich) übermittelt werden. Mehr dazu in unserer Datenschutzerklärung.
(mma)
