Is Nvidia's Tensor and Ray Tracing cores gimping conventional performance?

[atavax]

Has been any talk from knowledgeable people about like, If Nvidia is losing any significant performance pushing these two features?

[da na]

No

[Fasauceome]

The cores themselves are taking up die space and power budget which technically could be pushed towards more cuda cores, but the binning selection for that many cuda cores would be ludicrous.

 

Not to mention, RTX cores and tensor cores are actually extremely efficient at what they do.

[porina]

49 minutes ago, Fasauceome said:

The cores themselves are taking up die space and power budget which technically could be pushed towards more cuda cores, but the binning selection for that many cuda cores would be ludicrous.

Without specifics, I'd hope the tensor/RT parts would be idle when not in use, and essentially take no power. Silicon area is a manufacturing cost. For the same core count with and without these extra features, I'd not expect any notable performance difference.

 

There might be small factors that could swing it either way. It will spread out the heat production and may make cooling a bit easier, in a similar way some older Intel CPUs often run much cooler than AMD CPUs on the same cooling, even though AMD CPUs used less power. The smaller process condensed the heat output making it more difficult to cool. However, you could argue a bigger size might increase internal latencies. I don't feel this is a problem in practical implementations, but a more spread out die could have higher latency overall compared to a more compact one, which would be taken into consideration at design step.

 

If we go the other way, and ask what if we put that used die area into more "non-RTX" stuff, what could we gain? Link below claims 22% increase in area requirement for a TPC from adding those features. Note TPCs aren't the only things in a GPU, there's other parts like L2 cache and external IO that will dilute the RTX cost. Also we have to consider if you increase the core compute, you need to scale VRAM bandwidth if you want perf to scale better. Not doing so would reduce the benefit from the cores. Likewise consider power budget. More cores without increasing power budget will result in much less improvement than might be expected. Take all the above into consideration, dropping RTX stuff to allow more basic stuff probably would make a pretty insignificant difference in performance unless you start scaling everything else to go with it. 

 

https://www.techpowerup.com/254452/nvidia-rtx-logic-increases-tpc-area-by-22-compared-to-non-rtx-turing