The rumors do not stop and in full shortage of GPUs the leaks continue to arrive about the new graphics cards from NVIDIA and AMD. The data just released is as optimistic as, in principle, difficult to believe, but if they are correct we should expect a massive performance jump at both AMD and NVIDIA, where Ada Lovelace it would not be the spearhead in the greens.
Again Kopite7Kimi to the attack and once again the data provided leaves both locals and strangers speechless. Although we obviously cannot confirm or deny what is leaked, the reputation of this user is already legendary and at the moment it has a practically flawless success rate, but also take the following with a little salt and we are going to have trouble.
NVIDIA and AMD would push performance with Ada Lovelace and RDNA 3
-5nm (no matter TSMC or SEC)
-AD102 in transition or GH202 in revolution?
-RDNA3, GFX11, how does AMD reach its perf goal? Double the SIMD in CU?
-How about Intel?
– kopite7kimi (@ kopite7kimi) May 22, 2021
The data is as concrete as it is amazing, and it seems that the new thing from both companies would arrive in a 5 nm node that is not clear if it would be TSMC or Samsung, both with EUV and with a very similar density, a single mask and a jump in density and efficiency more than interesting.
But the data does not stop here, since the latest from NVIDIA would arrive with Ada Lovelace, but it would not be the most powerful, but a transition architecture in front of the main course: Hopper. AMD instead would focus only on RDNA 3, its new architecture that, as we already know, promises 50% more performance / watt.
The leaked numbers are really hard to believe, at least at first: NVIDIA would increase performance 2.2 times with Ada Lovelace and 3 times the performance with Hopper compared to the current Ampere, while AMD would achieve 2.5 times the performance of RDNA 2 with RDNA 3. Sounds difficult? Well, a priori of course they seem very optimistic numbers, but in addition, another figure has been filtered that makes the previous ones pale: a performance in FP32 of 100 TFLOPS.
Does NVIDIA have an ace up its sleeve with Hopper?
Taking into account that the current performance of the RTX 3090 is nothing less than 35.58 TFLOPS (unreal and marked by the controversy of dual function shaders), we are talking about almost 3 times the current performance, no matter how unreal.
How would this be possible? Behind the scenes NVIDIA and AMD are reported to have been studying the performance of their new chips with state-of-the-art lithographic processes for some time and have concluded that it can scale at a good pace in power, but they find a very clear bottleneck. : the memory bandwidth.
NVIDIA has alleviated this with GDDR6X and AMD with their Infinity Cache, but we don’t know exactly how they are going to address this in Ada Lovelace and RDNA 3. There is talk of the return to the HBM2 or even new cache systems with GDDR6X to ensure that the performance jump is not diminished by memory and above all not to shoot the final cost of each GPU.
In any case, there is almost a year left for the presentation of both architectures, where Hopper remains the great unknown, since NVIDIA could advance Ada Lovelace by the end of the year and in 2022 launch Hopper with architecture MCM to get ahead of AMD, if RDNA 3 is not already MCM. We will have to wait.
The post The next NVIDIA and AMD GPUs: at 5nm, double the performance and MCM? appeared first on HardZone.