The Blackwell GPUs feature two reticle-limited compute dies, which communicate via a 10TB/sec NVLink-HBI interconnect, allowing them to function as a single accelerator. They are also flanked by eight HBM3e memory stacks, offering up to 192GB of capacity and 8TB/sec of bandwidth. Nvidia's most powerful GPUs can be found in its GB200, which combines a 72-core Grace CPU with Blackwell GPUs, offering 40 petaFLOPS of FP4 performance and 384GB of HBM3e memory. However, these new chips are not expected to ship until the second half of the year, with the B200 and GB200 potentially not ramping up until early 2025.
Key takeaways:
- Nvidia has debuted its new Blackwell GPU architecture, aiming to extend its dominance in AI infrastructure with improved performance and power consumption.
- The Blackwell chips, successors to Nvidia's Hopper generation, claim to be roughly 5x faster in terms of raw FLOPS, and can do 20 petaFLOPS when using its new 4-bit floating point data type and opting for liquid-cooled servers.
- The Blackwell GPU architecture includes the B100, B200, and Grace-Blackwell Superchip (GB200), all of which share the same silicon and feature a pair of reticle limited compute dies communicating via a 10TB/sec NVLink-HBI interconnect.
- Nvidia's most powerful GPUs can be found in its GB200, which combines a 72-core Grace CPU with Blackwell GPUs, using the NVLink-C2C interconnect, and is capable of 40 petaFLOPS of FP4 performance and 384GB of HBM3e memory.