The B100 is expected to be built on a two-die configuration utilizing TSMC's CoWoS-L (Chip-on-Wafer-On-Substrate-L) advanced 2.5D packaging technology, facilitating the creation of larger processors. While specific details of the GPU's architecture remain speculative, there is anticipation that it may either continue with a monolithic design or transition to a multi-chip module (MCM) approach in future iterations. The B100 is also rumored to incorporate HBM3e memory technology, with potential upgrades in memory capacity, specifications, and features, such as 16-Hi stacks, with a possibility of adopting HBM4 in subsequent models.
Cooling solutions for the B100 are being developed to support up to 1000W, reflecting the significant computational power these GPUs are expected to deliver. Despite the lack of official pricing information, it is projected that the B100 will be priced above its predecessor, the H100, which is currently valued at over $20,000. The B100's memory configuration might include 8 stacks of 8-Hi HBM3e memory, providing a total of 192 GB, a capacity already offered by AMD's Instinct MI300 GPU.
Looking forward, the next-generation B200 GPU is rumored to feature 12-Hi memory stacks, potentially increasing the capacity to 288 GB. Details regarding the specific type of memory (HBM3e or HBM4) for the B200 remain undisclosed.
Leaks and predictions about the B100 and its potential performance and specifications have emerged, notably from tech tipster AGF/XpeaGPU and Reuters, despite NVIDIA's efforts to maintain confidentiality. These leaks suggest a significant leap in performance and capacity, with the B200 expected to further enhance these aspects a year following the B100's release. The enterprise variants of these GPUs are anticipated to be available first, followed by consumer-grade models for gaming and other applications.
