Lately it seems quite clear that the AMD Navi 31 GPU is going to have an MCM (Multi Chip Module) design, while the models Navi 33 and lower will continue with the usual monolithic architecture that we have had to date. New information about the GPU has also arrived AD102 Ada Lovelace architecture benchmark from NVIDIA regarding its memory configuration and (potential) operating speed. Let’s see it all.
NVIDIA breaks all records: the RTX 4090 would have 81 TFLOPS
Lately there have been rumors about the Ada Lovelace architecture from NVIDIA, and more specifically about the AD102 chip that will be its top of the range. Based on previous rumors, it appears that NVIDIA will use the TSMC node N5 for these GPUs and that their design will be monolithic and not MCM like AMD. The latest rumors assure that this GPU will work at a 2.2 GHz frequency or even more, which gives us a hint of the performance we can expect from this GPU.
The NVIDIA AD102 GPU would initially have a CUDA core count of 18,432, almost double the number of cores present in the Ampere architecture (which was already a big increase compared to Turing). Therefore, an operating speed of 2.2 GHz would give us a raw performance of nothing more and nothing less than 81 TFLOPS (FP32), more than double the performance of the current RTX 3090 (which has 36 TFLOPS).
This represents a 125% increase in performance, a figure that seems too high and, however, we must remember the increase in performance in FP32 numbers that we already had with Ampere (the RTX 3090 offers 36 TFLOPS in front of 13 TFLOPS of the RTX 2080 Ti, and that was already a 150% performance increase). However, we must also remember that these figures are FP32 gross performance, and then in games things change a lot (from Turing to Ampere the increase was 50-60%).
In addition to this, the leaker has also claimed that the RTX 4090 would retain a 384-bit bus interface, just like the RTX 3090; It is not specified if the GDDR6X memory type will be maintained, so this could mean that NVIDIA will not make the leap in terms of memory type in this new generation, but it is likely that it will use the latest iterations of this interface with 20 Gbps (or even more). This could be an interesting performance leap on this RTX 4090.
AMD to battle with the RX 7900 XT
This information, which comes from the same leaker, ensures that AMD will use a MCM design on the RX 7900 XT and will be based on a 5 nm + 6 nm lithography for this. The leaker assures that AMD will keep the GDDR6 memory with a 256-bit interface, on the Navi 31 chip, while on the Navi 32 chip (which will enhance the supposed RX 7800 XT) will also have an MCM design but with a memory interface of only 192 bits.
For its part, the Navi 33 chip (for the supposed RX 7700 XT) will be a monolithic piece with a 6nm lithography and a 128-bit GDDR6 memory interface, so we would be facing an already almost entry-level GPU in this respect and that certainly will not compete with the high-end of NVIDIA. Of course, it is said that AMD could take advantage of Infinity Cache to compensate for the performance losses when using these buses as “scarce” as currently happens with the RX 6000 series.
In principle, the Navi 33 chip should offer superior performance to the current RX 6900 XT and, although it will obviously not compete with NVIDIA’s RTX 4090 in performance, it could give a twist to the market because there is talk that AMD will launch it At a price of 450 dollars.