Nvidia's next gaming architecture rumoured to be called Lovelace

by Mark Tyson on 29 December 2020, 10:11

Tags: NVIDIA (NASDAQ:NVDA)

Quick Link: HEXUS.net/qaepyf

Add to My Vault: x

Despite the rollout of GeForce Ampere graphics cards being very much incomplete, Twitter tech leaker wires are buzzing with talk of next gen Lovelace GPUs. Various well known leakers have purported GPU specs and roadmap info but it is admittedly rather early for this kind of information as Lovelace isn't expected to be official until 2022.

Kopite4kimi seems to be the source that first turned the spotlight onto Ada Lovelace as the next gen family of GPUs for GeForce gamers. It is thought that Lovelace is stepping to the fore as a monolithic series of GPUs for gamers as the multi-chip design of Hopper might be prosumer/professional/HPC only and might be somewhat delayed. Lovelace GPUs will come with a codename beginning with AD, with the top dog being the AD102. Kopite4kimi has more recently proposed that while GA102 has a 7x6 structure, the AD102 GPU will have a 12x6 Texture Processing Clusters structure.

Sites like 3DCenter and VideoCardz have taken the AD102 ball and run with it to come to the following educated guesses for Lovelace AD102 GPU specs:

  • 12 Graphics Processing Clusters
  • 72 Texture Processing Clusters
  • 144 Streaming Multiprocessors
  • 18,432 CUDA cores

So, AD102 could have a significant increase in CUDA cores compared to GA102 (71 per cent more) and it is further estimated that the above top Lovelace GPU could achieve 66TFlops @1.8 GHz.

Nvidia GTC 2018 Heroes T-Shirt detail

Even if the above is true there is quite a lot of detail up in the air right now. For example, no one knows at what foundry/process Lovelace will be fabricated, though a hat is tipped at Samsung 5nm. As for timescales, as per our intro, Nvidia is still in the midst of its GeForce Ampere rollout. In spring one might hope for it to have completed addressing the various market segments (especially lower tier) before announcing RTX 30 Super GPUs at Computex in June perhaps. It might be a year later before RTX 40 series graphics cards become a thing.



HEXUS Forums :: 15 Comments

Login with Forum Account

Don't have an account? Register today!
how is it that AMD don't need so many cores and still be up there?
So the Samsung 5nm has me a little bit at odds, they failed their 7nm EUV and not sure they've even gotten it fully on track to being a fulfilled competitive node.

So to jump to 5nm Samsung? Hmm, i'm a little unsure.
Tabbykatze
So the Samsung 5nm has me a little bit at odds, they failed their 7nm EUV and not sure they've even gotten it fully on track to being a fulfilled competitive node.

So to jump to 5nm Samsung? Hmm, i'm a little unsure.

Don't fall into the trap of comparing nodes. Samsungs 8nm is roughly comparable to TSMC's 7nm and whilst the yields started off bad I'm hearing that they are improving quite well. Intel have failed to get their 10nm node working well so far so whilst I'm sure that TSMC are ahead Samsung appear to be on track if a little later than expected
QuorTek
how is it that AMD don't need so many cores and still be up there?

Didn't Nvidia update their ‘cores’ to do ‘up to’ two instructions per clock so they're now counting them as double? I.e. the 10496 CUDA ‘cores’ in the 3090 is actually 5248 physical cores which is comparable to the 4352 in the 2080Ti. Nothing wrong with that but I think in terms of physical cores AMD and Nvidia are still comparable.
Hopefully its more Ada Lovelace and not more Linda Lovelace! :P