Inside Rubin: NVIDIA's most powerful AI chip

======================================================== Learn to write the programs that keep these tensor cores busy, in Vizuara's Kernel Engineering Workshop: https://kernelworkshop.vizuara.ai =================================================== We take NVIDIA's Rubin apart, from the package down to a single tensor core, and find out why even the most powerful AI chip spends most of its time waiting for memory. Rubin holds 336 billion transistors, and nearly all of them exist to do one small job: multiply two numbers and add the result to a running total. Layer by layer, we look at the two compute dies and why there are two, the eight HBM4 memory stacks, the links on its edges, the 224 streaming multiprocessors, and the tensor cores where the multiplying happens. Then we put Rubin beside the H100 and look at the one ratio that decides how fast AI actually runs. Chapters: 0:00 Meet Rubin 0:20 One small job 0:43 Five layers 1:04 Why two dies 1:26 The memory 2:17 Links, clusters and SMs 2:55 The tensor core 3:34 Rubin and the H100 3:56 The catch ======================================================== Learn to write the programs that keep these tensor cores busy, in Vizuara's Kernel Engineering Workshop: https://kernelworkshop.vizuara.ai ===================================================