Isn’t this how these things are always done? Same with number of cores in a CPU, the higher core and lower core come from the same fabrication process, it’s just that the lower core CPU had more faulty cores that are disabled.

Yes, that's why I question the approach here.

Testing hardware is a complicated matter (you can do a PhD on the topic), and nvidia probably does a better job at it than a hacker in their basement who thinks they can upgrade their GPU for free.

On the other hand ... perhaps a few bit errors is not a problem for AI workloads ... but they could also crash your system ...

The good news is that Nvidia did test the hardware, and left the "defective" masks right there in the fuses for us to read. If you have a card that has been fuse-segmented well below its defective part count, you have a product segmentation product on your hands.

>a hacker in their basement who thinks they can upgrade their GPU for free

Except he literally did.

Depends a lot on when its from in the product cycle, the yield at the end of production runs are usually so good that it turns into artificial product segmentation

There's an aspect of that, but the primary reason for differences between various CPUs and GPUs which have the same underlying chip is product segmentation.

But nowadays they can make much more money with enterprise targeted products.

Which is why I think these devices are more likely to be faulty.

Nowdays - but this was a mining card, meaning bought for the crypto boom and before AI.

The author mentions about 90% yield on getting the full 80GB of memory (in a limited 10 card sample).

Historically, there are products with high enough yield rates that it makes it irrelevant. See ATI Radeon 9500 Pro. A two byte patch in the firmware gives you an ATI Radeon 9800 Pro.

Sometimes people order more lower-tier chips than there are defects, and then you get fully flawless silicon with some disablement.

Used to be now we have chiplets. The defective chips ship as some weird OEM part usually now with parts disabled. Example: Ryzen 5 3500

I don't know what Intel is doing as I quit keeping up with them. I do miss the core unlocked boards that you'd put a 2 core cpu and maybe win the lottery on and get all 4 cores.

This is why, for a while, AMD was selling 3-core CPUs. They were quads with a bad core, which could sometimes be unlocked.

They definitely did, but maybe you can get some good results running your own tests and determining if they really ARE bad or not, or they just overly tagged them out.

it appears the opposite is true and that's why it's faster than an A100, which is optimized for yield and allows more defects

[delayed]

The sheer recklessness on display here. Everyone knows random bits is how something becomes sentient! You'll kill us all!

It would be hilarious if it became sentient but locked in an inferior GPU.

we call that dithering in the mechanical and analog worlds

They say ECC works (?) so you would know.

I recently had a faulty A100, for some reason the provider turned off ECC on that card, so it silently failed and gave junk results... Took me a minute to figure out.

How did you ever figure this out?

The weights loaded fine, but as I was training the loss made no sense. So I wrote some basic matmul test.

HBM is so expensive that I doubt you would find significant gains on average. They are already yieldmaxxing like crazy.

The article mentions that some die areas are factory-marked as defective and have to be "re-disabled" after the unlock. Ctrl-F _DEFECTIVE