The second generation of Nvidia accelerators with Blackwell architecture represented by the B300, as reported by Tom’s Hardware with reference to SemiAnalysis, will offer a 50% increase in performance compared to the GB200, but the TDP level will only increase from 1200 to 1400 W. Chips of the B300 family, according to analysts, will appear approximately six months after the B200, deliveries of which were already supposed to begin this quarter.
Image source: NVIDIA
TSMC will produce B300 chips using the previous 4NP technology, but this will not prevent them from achieving a 50% increase in computing performance. Another important change will be the use of twelve-tier memory stacks of the HBM3E type by the B300 series chips. It will provide a memory capacity of 288 GB per accelerator and a bandwidth of 8 TB/s. Such changes together will reduce the cost of training neural networks by up to three times compared to their predecessors.
The advent of the ConnectX-8 800G class network controller will double the network interface bandwidth relative to the current ConnectX-7, and increasing the number of PCI Express lanes from 32 to 48 will expand the possibilities for integrating these accelerators in server systems. An important change in the production of B300 accelerators will be Nvidia’s refusal to supply motherboards or server systems strictly in the reference design. By expanding partners’ access to the production of such accelerators and systems, Nvidia will increase product supply volumes. In general, the design of systems based on the B300 and GB300 will be based on the principles of greater freedom, and customers will only benefit from this.