[1oooqooq]

interesting timing to recall that story. now the same trick is used for h100 vs whatever the throttled-for-embargo-wink-wink Chinese version is called.

but those companies are really adverse to open sourcing because they can't be sure they own all the code. it's decades of copy pasting reference implementations after all

[rfoo]

> now the same trick is used for h100 vs whatever the throttled-for-embargo-wink-wink Chinese version

No. H20 is a different chip designed to be less compute-dense (by having different combinations of SM/L2$/HBM controller). It is not a throttled chip.

A800 and H800 are A100/H100 with some area of the chip physically blown up and reconfigured. They are also not simply throttled.

[1oooqooq]

that's what nvidia told everyone in mar 23... but there's a reason why h800 were included last minute on the embargo in oct 23.

[rfoo]

That's not what NVIDIA claimed, that's what I have personally verified.

> there's a reason why h800 were included last minute

No. Oct 22 restrictions are by itself significantly easier than Oct 23 one. NVIDIA just need to kill 4 NVLink lanes off A100 and you get A800. For H100 you kill some more NVLink until on paper NVLink bandwidth is roughly at A800 level again and then voila.

BIS is certainly pissed off by NVIDIA's attempt at being creative to sell the best possible product to China. So they actually lowered allowed compute number AGAIN in Oct 23. That's what killed H800.

[1oooqooq]

I see. thanks for the details.