They may not be disabling them maliciously -- they may be "binning" them -- running tests on the parts and then fusing off/disabling broken pieces of the silicon in order to avoid throwing away a chip that mostly works.

▲

swiftcoder 3 days ago | parent [-]

Yep, that's likely the case - but they still charge double for the reduced-performance binned chip, just because it's a "professional" GPU (which, last I heard, really just means it can use the pro variant of the GPU drivers)

	▲	derefr 3 days ago \| parent [-]
		Funny enough, maybe the fusing itself (if they go a bit above-and-beyond on it) is exactly why it is a pro model. I.e. maybe Nvidia say "if we're going to fuse some random number of cores such that this is no longer a 3050, then let's not only fuse the damaged cores, but also do a long burn-in pass to observe TDP, and then fuse the top 10% of cores by measured TDP." If they did that, it would mean that the resulting processor would be much more stable under a high duty cycle load, and so likely to last much longer in an inference-cluster deploy environment. And the extra effort (= bottlenecking their supply of this model at the QC step) would at least partially justify the added cost. Since there'd really be no other way to produce a card with as many FLOPS/watt-dollar, without doing this expensive "make the chip so tiny it's beyond the state-of-the-art to make it stably, then analyze it long enough to precision-disable everything required to fully stabilize it for long-term operation" approach.