A100 is nvidia's ML card in this lineup - and it has a much saltier premium.

The 3090 will probably stay at a premium due to a few factors - I don't think ML performance plays into this at all though:

DLSS - There's a reason AMD cites "raster performance" since with DLSS enabled the 3090 has a major advantage.

AMD-specific optimizations - near the end of the presentation AMD disclosed that with all the optimizations on (including their proprietary cpu->gpu optimizations, only available on the latest gen cpus)- they could pass 3090's raster performance in some games.

I think for these two reasons, and the fact nvidia can't seem to get cards to vendors (and customers), there won't be a price drop on this SKU. They may however release a watered down version as a 3080ti and compete there.

Aren't the optimizations also restricted to the 500 series motherboard chipsets?

Yes, which is one of the two supported chipsets for the zen3 arch.

Not according to the tensorflow benchmarks so far, it smokes the Titan RTX. Whether it could be faster is irrelevant, it's the fastest currently by a lot for the money.

FP16 multiply with FP32 accumulate performance is on par if not slower than Titan RTX due to half rate which is what the parent was likely referring to.

[1] https://www.nvidia.com/content/dam/en-zz/Solutions/geforce/a... (Appendix A)

BTW - Whether it "could" be faster is indeed relevant because some of us are holding out for a Titan GPU next year with this unlocked. If you have unlimited budget or are under time constraints then by all means get the 3090, it is a beast. But if one has a 2080 TI then it's an important consideration.

I'm obviously aware of the theoretical limitations they referred to. My point that it smokes the Titan RTX in real world benchmarks stands, there's more to machine learning performance than just that one stat. Performance should be measured against real-world use cases, not whether there are lines in the whitepaper you object to.

If you can wait until next year you should always wait until next year, because there will (almost) always be something better than what is currently out. That's unrelated to whether or not the 3090 is good for doing ML research; it objectively is.