In the last few months, there has been a lot of new information coming from leakers about the RDNA3 and ADA Lovelace graphics cards, so it’s time for a summary!
The next generation of wattage-hungry graphics cards?
First of all, most of the rumors come from 2 leakers, @kopite7kimi and @greymon55 who revealed what could be the characteristics and performances of the next graphics cards from AMD and NVIDIA.
On the AMD side and the RDNA3 architecture, there is talk of using the MCM technology, consisting of combining several die of different engravings. Three chips would be produced, Navi 31, 32 and 33.
Navi 31, the high-end version of RDNA3 would come with 2 dies of 60WGP engraved in 5nm, each containing 256 shaders, giving a total number of 15360 shaders (almost 3 times more than Navi 21). On the memory, we would be on 32 GB of GDDR6 in 256 bits, accompanied by 256 or 512 MB ofInfinity Cache engraved in 6nm and using 3D technology.
The frequency should also progress, around 2.4-2.5 GHz stock, allowing performance up to 75 TFLOPS in FP32. But all this would come with a power consumption exploding and exceeding 500 watts.
Navi 32, the mid-range, would have 2 die of 40 WGP this time if giving 10240 shaders, accompanied by 16 GB of GDDR6 on 192 bits and 256 MB of Infinity Cache always using the MCM.
Navi 33, the entry of range, would use only the engraving 6nm with only one die of 40 WGP giving 5120 shaders (as Navi 21) and would have 8 Go of GDDR6 on 128 bits and 128 Mo of Infinity Cache.
At NVIDIA, the new generation will be called ADA Lovelace and should like AMD, double or even triple the current performance. We have for the moment, only information on the highest-end chip, namely ADA102.
Like RDNA3, NVIDIA will return and use the 5nm from TSMC for its gaming graphics cards, and should also achieve a power consumption exceeding 500W.
On the specifications, we would be on 144 SM giving 18432 cuda core (against 84 and 10752 for the GA102 equipping the 3090/3080ti/3080). So an increase of 70% of cuda core!
The whole would be accompanied once again by the GDDR6X on a 384-bit bus. The frequencies should also make a jump, with around 2.3-2.5 GHz stock, allowing to reach up to 92 TFLOPS of performance in FP32.
Of course, these are all rumors! These generations of graphics cards are scheduled for the end of next year. So we’ll have to wait a while before we see any “reliable” information.
