amazing to think there was a time (literally not too long ago) where you would throw 80 GB of HBM onto a chip and turn 90% of it off

▲

cromka a day ago | parent | next [-]

Those were binned cards, though.

▲

octoberfranklin a day ago | parent [-]

Binned for compute failures, not memory failures.

That's why this works. The yield on those HBM stacks is excellent. The binned cards had failures in the logic die, and hyperscalers want uniformity in their clusters. That's why the mining market took the downbinned cards.

▲

londons_explore 20 hours ago | parent [-]

Do you not test the logic die before mounting the memory on top?

	▲	octoberfranklin 12 hours ago \| parent \| next [-]
		Some testing can be done with a wafer probe, but other tests require stressing the die -- high current, heat output -- so it has to be in-package to do those tests. Mounting the HBM on the interposer has to happen before putting the interposer in the package. So it's the "must be in-package" yield tests that drive this.
	▲	monster_truck 16 hours ago \| parent \| prev [-]
		back then, no

▲

qurren a day ago | parent | prev | next [-]

I'm wondering if this is actually the case or whether they were binned for defects in the HBM

▲

kmike84 a day ago | parent | next [-]

Anecdotically, 64GB works extremely reliably on 8GB cards after unlock.

It is much more of "let's sell cards to crypto people, but nerf them, to not harm our enterprise margins", not "we've got a ton of subpar hardware, let's to something with it".

	▲	londons_explore 20 hours ago \| parent [-]
		crypto was a real boom... I could imagine these cards were already in production when the crypto guys came along willing to give lots of $$$ for them, but wanting really short lead times...

▲

legitronics 15 hours ago | parent | prev [-]

I have a grey market a100 that had hbm errors but the card used some of its spare rows to remap. I believe they introduced this with ampere, and I was a little annoyed that my “new” card had these but I got it so cheap I didn’t really care. It was 7k usd (vs 9k and 10k) and I remember being unsure if it was worth it… None of them needed to remap any more in they years I’ve had them but I think I got that one so cheap because of it. The most expensive one came from the best rated seller and appears to be an engineering sample but has passed every test I’ve thrown at it, the only thing “wrong” with it is that I can set the power limit to whatever I want? This was back when b100 was about to be formally announced and before the pricing so eBay sellers were trying to unload them.

▲

tencentshill a day ago | parent | prev [-]

Sam Altman is probably buying them all up now