What is Nvidia up to with its next-generation GPU? According to recent leaks, the specs of the RTX 5090 are expected to leave everything that has come before it behind by a significant margin
With the release of the Geforce RTX 4090, graphics card manufacturer Nvidia is running lonely circles in the top performance class of our GPU ranking. This gap to its own upper-class models and yes, also to AMD, is supposed to be widened even further by the upcoming flagship of the Blackwell
christened generation.
We are of course talking about the Geforce RTX 5090 as well as a potential Ti version, which is still missing from the RTX 4090. Although it will still take some time until a possible release – the talk is about a release in the year after next – the rumor mill is already running hot, because the well-known leaker kopite7kimi claims to have found out concrete details about the RTX 5090.
According to the leak, the Geforce RTX 5090 seems to step up a gear in almost every specification.
Specifically, kipite7komi talks about a 12×8 structure of the GB202 chip that is used in the flagship. This means twelve GPCs (Graphics Processing Clusters) with eight TPCs (Texture Processing Clusters) each. With twelve streaming multiprocessors (SM) within a GPC, nothing should change.
As I mentioned before, GA100 is 8*8, and GH100 is 8*9. GB100 will have a basic structure like 8*10. GB202 looks like 12*8.
– kopite7kimi (@kopite7kimi) September 28, 2023
Now it gets technical for a moment: If this leak is true, we can also deduce how many CUDA cores are installed in the RTX 5090. A total of 192 SMs of 64 FP32 and FP32/INT32 units each result in the following equation:
Number of CUDA cores = Number of SMs * Number of FP32 + Number of SMs * Number of FP32/INT32.
So, inserting the numbers from the rumor, we get that the RTX 5090 can fall back on 24,576 CUDA units – growing by pretty much exactly 50 percent from the RTX 4090.
According to the leak, Nvidia also wants to make a big splash, not a small one, when it comes to video memory.
How much L2 cache and memory bit bus?
– Warren Tarbiat (@Dangerman1337) September 28, 2023
Recent rumors already say that the RTX 5090 has a GDDR7 memory of 32 GBytes, which is supposed to be connected via a 512-bit interface according to kopite7komi. The L2 cache is supposed to increase to 128 MB.
In order to clarify the assumed performance leap between the RTX 4090 and RTX 5090 once again, you can find all relevant and supposed specifications summarized in the following table:
Now, of course, it’s conceivable that Nvidia will use a slightly trimmed version for the RTX 5090’s GB202 processor, similar to the RTX 4090’s AD102 chip.
Finally, even the RTX 4090 does not exhaust the theoretical maximum capacities of the AD102 chip, which is why a potential release of an RTX 4090 Ti was considered likely for a long time – until recent leaks probably shelved it entirely.
In addition, it remains questionable why Nvidia should even fully utilize the GB202 chip, because as can be seen in the comments below the leak as well as in our GPU ranking mentioned at the beginning, the current flagship is already running out of competition.
It only remains to be seen whether AMD can somehow close this gap with Radeon RX 8000 – it would certainly be good for the competition.
Now it’s your turn: What do you think of the RTX 5090 leaks? Would these specs finally cement Nvidia as the frontrunner in the high-end segment, or do you trust AMD to catch up? Let us know in the comments!
