Get the latest tech news
Blackwell: Nvidia's GPU
Nvidia has a long tradition of building giant GPUs.
Blackwell does away with subchannel switches, letting it more efficiently fill its shader array if applications frequently mix different work types on the same queue. Nvidia added floating point instructions to Blackwell’s uniform datapath, which dates back to Turing and serves a similar role to AMD’s scalar unit. Mixing local and global memory traffic can further increase achievable bandwidth, suggesting the LDS and L0 vector caches have separate data buses.
Or read this on Hacker News
