The RTX 40 will be a simple evolution of the RTX 30, only at 5nm

According to what has just been leaked as a rumor, it seems so, since although the FP64 will return, in theory, to 1:1 levels, the changes could be minimal from the point of view of architecture, which would be quite interesting. for many reasons.

NVIDIA RTX 40: just an evolution with TSMC’s 5nm


Apparently, although it is true that Ada Lovelace will bring small changes in the form of improvements, it would seem that they would not be enough to be called a new architecture as such. This would be a less natural evolution than Huang’s are used to, as Ada Lovelace will be based on Ampere and will share much of her units as they are now.

What does this imply? Well, the SM would have barely evolved and everything is achieved thanks to the TSMC N5 node, so the AD102 chip would include those already mentioned in other articles 12 GPC, 72 TPC and 144 SMwhich and each one of them would continue to have 128 Shaders or CUDA cores with the drag of the controversy of the duplication of their real number (there would be 64 CUDAS per SM in reality).

The total sum knowing this are those already known 18,432 shaders, which are no less than 71% higher in number compared to the current 10,752 of the GA102. The performance should be, if not above, close to 90 TFLOPS, which is a +152% versus the RTX 3090 and a +125% compared to the RTX 3090 Ti, mainly because they talk about crazy frequencies: 2.5GHz.

The jump to MCM and Hopper


NVIDIA would thus cling to TSMC’s 5nm and small improvements to achieve that performance, all based on pure muscle and few optimizations, which suggests a change in trend.

It seems that AMD’s blow with Instinct MI200 has caught them off foot and Intel’s hook with Xe and MCM would be the trigger for a relevant fact: NVIDIA could be technologically behind after almost two decades at the top. It seems that the efforts of the greens would be focused on Hopper and MCM, as well as their CFR technologythe replacement for AFR and SFR that would arrive not for two or more GPUs, but for the interconnection of dies.

This would be the reason why consumption is going to shoot up so much, mainly because a worsening of efficiency is expected since, in theory, there are not enough architectural improvements to mitigate the high frequency to achieve greater performance. This reminds us of what happened with Intel and all the architectures after Skylake (at least in part), except that NVIDIA will only need two years to make the successive changes, not like in the RT Cores and Tensor Cores where they had worked for a decade.

Therefore, Ada Lovelace is expected as pure muscle and a lot of watts to consume, a high frequency and possibly no overclock that will leave the thermal margin in a more than tight range. AMD has with it a golden opportunity to deal an accurate blow to NVIDIA, we will see if she succeeds.

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *