Is NVidia really selling cards with 90% of their provisioned RAM unused? Explains a lot if so.
I do wonder how well the card's power bus and cooling supports using 150% of the listed compute units. Cooling at least isn't that hard to upgrade, but power?
The thing about HBM memory is that it is part of the same chip as the GPU itself – the memory die and logic die get bonded to an interposer die during manufacturing.
Because of this, it is quite plausible that the GPU die cannot be fully yield-tested early enough in the manufacturing process to install reduced amounts of RAM in the mining cards.
The bit I don't understand, and the reason I posted was to get other's opinions on, is: if the die was binned in this way surely the compute units that are "turned back on" should be faulty in some way?
This isn't new, there are two aspects: there may be more functional units than a given device class, but fewer than the next size up, and yields increase over time leading to fewer failed units but you still want to maintain the cheaper end (I've always assumed so that the higher end ones can have a much higher premium or something, but who knows?)
icefox | 22 hours ago
Is NVidia really selling cards with 90% of their provisioned RAM unused? Explains a lot if so.
I do wonder how well the card's power bus and cooling supports using 150% of the listed compute units. Cooling at least isn't that hard to upgrade, but power?
meithecatte | 20 hours ago
The thing about HBM memory is that it is part of the same chip as the GPU itself – the memory die and logic die get bonded to an interposer die during manufacturing.
Because of this, it is quite plausible that the GPU die cannot be fully yield-tested early enough in the manufacturing process to install reduced amounts of RAM in the mining cards.
icefox | 18 hours ago
Heckin wild, I didn't know that. Thanks!
[OP] wgx | 20 hours ago
The bit I don't understand, and the reason I posted was to get other's opinions on, is: if the die was binned in this way surely the compute units that are "turned back on" should be faulty in some way?
ceph | 20 hours ago
Not necessarily. If yields exceed supply they will just bin non-defective chips to maintain price tiering.
[OP] wgx | 20 hours ago
I had no idea. Interesting!
ner0n | 13 hours ago
That's exactly what was bugging me while reading this, surely this was binned this way because it failed to be a full A100. This is a different era.
olliej | 9 hours ago
This isn't new, there are two aspects: there may be more functional units than a given device class, but fewer than the next size up, and yields increase over time leading to fewer failed units but you still want to maintain the cheaper end (I've always assumed so that the higher end ones can have a much higher premium or something, but who knows?)