The image shows a professional speaker giving a presentation to a group of people at an event. The speaker is standing at a podium, with a laptop on a nearby table. The room is well-lit and has a projection screen displaying the speaker's presentation. The audience is seated at tables or chairs, with some people standing and others sitting. The scene is likely a conference or seminar.
Text transcribed from the image:
Asynchrony: Overlapping GEMM and Softmax
Why overlapping?
Ex: headdim 128, block size 128 x 192
FP16 WGMMA: 2 x 2 x 128 x 192 x 128 = 12.6 MFLOPS, 4096 FLOPS/cycle -> 3072 cycles
MUFU.EX2: 128 x 192 = 24.6k OPS, 16 OPS/cycle -> 1536 cycles
Omniscience
AI PAPERS AND B
OMNISCIENCE.T