It looks like NVIDIA is going to put two flagships on the table that go way above what we currently have with Ampere. A new leak reveals what we suspected: the NVIDIA RTX 40 will have a very high frequency, in particular it is about 3 GHz for the RTX 4090 Yes 2.8GHz for the alleged RTX4090Ti. Is it really possible? Let’s see.
Again such a reliable leaker as Kopite7kimi on stage and again throwing a dart moderately clarifying where the shots will go. What has been said now seems more probable and the data that we are going to offer to give or take away credibility is completely in line with what can be expected from these new RTX 40s.
NVIDIA would have a much better litho node than AMD
Getting to 2.8 GHz is easy, at least not very hard.
– kopite7kimi (@kopite7kimi) June 21, 2022
Yes, TSMC’s 4N destined for NVIDIA is simply an act of fear vis-à-vis AMD and on the other hand, a movement of guarantees from the manufacturer of the chips in the contracts. However, it is the node that is largely responsible for the improvements, because if we look at what it brings, we can begin to understand everything that has previously been filtered.
If we compare the 8nm LPP with these new 4Ns, the differences are really interesting, since TSMC would have achieved a 40% more frequency and 60% more density per mm2. This would mean that the same architecture the RTX 4090 with its new node would obtain a frequency of 2.375MHz in Boost, but logically as we saw a month ago, NVIDIA’s architectural changes make greater frequency scaling possible.
Taking into account that the theoretical performance of the RTX 4090 will be around 100 TFLOPS and assuming you have 16,834 shaderscounts to do the frequency calculations are quite simple if we assume that 3GHz serial:
16,384 x 2 x 3 GHz -> 98,304 TFLOPS
A figure very close to what has been said and which would reveal the charismatic figure of 3 GHz. Of course, if everything said above is true.
The RTX 4090 Ti or TITAN RTX 40 would lower its frequency
We are talking about 800 watts for the whole chip and this would in theory be that of the RTX 4090 Ti. The AD102 would have in its largest version no less than 18,432 shaderswho makes the right accounts and knowing that it will be around 2.8 GHz approximately:
18,432×2×2.8GHz -> 103,219 TFLOPS
It might not seem like much of a difference, but the older of the two siblings surely hits their VRAM more frequently, and therefore, greater final bandwidth to the same bus, where it’s also more than likely to double the size of memory because it is already a tradition at NVIDIA.
This would leave a bigger difference in performance, surely close to 8% or 10%, which for those in Huang would be justification for raising the price. If, as the leaker says, reaching these numbers of 2.8GHz for the complete chip it is not very difficult, we will have to see how far the manufacturers and assemblers can take the custom models, many of which will surely be by water.