Despite concerns that mass production of Blackwell’s ‘B200’ GPU could also be delayed than expected resulting from design issues, predictions have emerged that the discharge schedule of the 2nd generation Blackwell ‘B300’ GPU will proceed as planned.
Semianalysis reported on the twenty sixth (local time) that the B300 GPU is scheduled to be released roughly six months after the discharge of the B200. For the reason that B200 and GB200 models have been released now, the B300 series is anticipated to be released within the second or third quarter.
This schedule indicates that NVIDIA is releasing its next-generation products ahead of schedule, and appears to be intended to dispel concerns about the potential of a delay in mass production of the B200 and GB200, which have had heating problems resulting from design issues.
One in all the most important differentiating aspects of the B300 GPU is its memory capability. While the previous model used 8-layer ‘HBM3E’ memory, the B300 adopts 12-layer HBM3E memory, providing a memory capability of 288GB per GPU and a bandwidth of 8TB per second.
This is anticipated to significantly improve the speed of AI model training and inference, and specifically, reduce inference costs by as much as 3 times. That is as Nvidia’s pursuers proceed to release AI chips that deal with inference.
B300 may handle larger batch sizes, support prolonged sequence lengths, and help address latency issues in user interaction.
Not only will there be a rise in memory capability, but there can even be significant changes when it comes to power consumption. The B300 is manufactured using the identical TSMC 4-nano (nm) process because the B200, but its computational performance is improved by 50%, and its thermal design power (TDP) is 1400W, which is 200W greater than the B200. This extra 200W is anticipated to extend performance by 50%.
There’s also a possibility of adopting Nvidia’s ‘800G ConnectX-8’ network interface card (NIC). This provides twice the bandwidth of the prevailing ‘400G ConnectX-7’ and increases the variety of PCIe lanes to 48, which is anticipated to be a very important advantage in large-scale cluster operation.
Lastly, NVIDIA plans to significantly change the availability chain method for the B300 and GB300 in comparison with the B200 and GB200 models. Now, fairly than selling entire servers or motherboards, NVIDIA plans to sell the B300 mounted on an SXM Perc module and supply it with a Grace CPU and Axiado’s Host Management Controller (HMC). This may allow more corporations to take part in the Blackwell system, and the usage of Blackwell-based systems is anticipated to spread.
With the B300 and GB300, NVIDIA plans to provide hyperscalers and OEM partners greater freedom in Blackwell system design. This will likely lead to differences in price and performance.
Reporter Park Chan cpark@aitimes.com
