When it comes to GPUs, Nvidia has lost very little in the face of a resurgent AMD, and this is made possible by consistently maintaining technological dominance generation after generation. Nvidia’s top-end gaming GPU, the RTX 3090 is still uncontested in its bracket, and if a recent rumor comes to pass, Nvidia will most likely continue to dominate the space.
n31:5nm+6nm 256bit gddr6
n32 5nm+6nm 192bit gddr6
n33 6nm 128bit gddr6 perf>6900xt
ad102 5nm 384bit gddr6x
— Greymon55 (@greymon55) September 19, 2021
According to Greymon55 via Wccftech (having a very reliable record), Nvidia’s next-generation AD102 GPU could potentially hit and cross the 2200 MHz clock speed mark. If the rumor gets these numbers right or is even close to the figures on release, we can then stand to witness a significant generational jump, like the one brought on by Ampere powering the RTX 3000 series cards. The current reigning gaming GPU, the RTX 3090 boosts to about a max of 1700 MHz, so the rumored figures are certainly a significant jump from the previous generation.
Some earlier rumors on Lovelace have also corroborated the significant performance jump, with 3DCenter speculating the likely specifications.
- 12 Graphics Processing units
- 72 Texture Processor Clusters
- 144 Streaming Multiprocessors/18432 Cuda Cores
- 384-bit bus/GDDR6X Memory
So, nVidia's AD102 chip maybe is like:
18'432 FP32 units
~66 TFlops FP32 power (on 1.8 GHz) https://t.co/A8OnUktE1s
— 3DCenter.org (@3DCenter_org) December 28, 2020
This can theoretically amount to the AD102 GPU delivering an insane 66.4 TFLOPs (FP32). Putting this in context, the RTX 3090 maxes out at around 36 TFLOPs. Moreover, the earlier rumor assumed clock speeds of 1.8 GHz, so after calculating the new figure from the recent information, the revised floating-point compute performance stands at an even more ludicrous figure of 81 TFLOPs. This can translate to the AD102 GPU being over two times faster than an RTX 3090 in Floating-Point calculations.
As Wccftech rightly point out, FP32 compute scores don’t scale linearly with real life performance and the actual gains are significantly less. Rumors also suggest the “AD102” GPU will retain the 384-bit bus width with GDDR6X memory, similar to the RTX 3090. This is expected because GDDR6X already provides a significant amount of bandwidth (Over 40% Uplift from GDDR6), and most likely won’t be a bottleneck point in an intended use case scenario.
This contrasts with TSMC who announced a 1.8x density improvement and a 15% performance improvement or 30% lower power.
TSMC has announced a much larger density improvement than Samsung. My belief is that TSMC will have a significant density advantage over Samsung at 5nm.
– Scotten Jones (SemiWiki)
Another interesting aspect of the next-gen Ada Lovelace architecture is its rumored 5nm node, and according to Wccftech, this will be taken up by TSMC this time instead of Samsung. This can be surprising for two reasons, firstly because TSMC’s 5nm has a lot of demand and is already running at full capacity, and secondly, Samsung already manufactures Ampere GPUs for Nvidia, and now with their new 5nm node, Samsung were expected to be the frontrunners for the next generation of Nvidia GPUs.
Although this attributes to TSMC’s superior 5nm process, which is said to be denser than Samsung’s own 5nm node, leading to better performance gains and lower power consumption. Again, with TSMC’s production constraints, it’s likely that Nvidia will use the Taiwanese foundry exclusively for higher-end cards, and the rest will go to Samsung for volume production where a bleeding-edge node isn’t necessary. Nvidia’s Ada Lovelace line-up is still far away, most likely late 2022, so take the information here as speculation, at least until we have more concrete information closer to launch.