NVidia Ada Speculation, Rumours and Discussion

Status
Not open for further replies.
RT
Kimi overestimated Hopper's die size, not transisor count, and NVIDIA hinted at that number already in their teasers.




RTX40-hero.jpg

Calling that number plays a recorded message asking “how fast would you like to go”.

I’ve seen some theories that 629 is the die size and 7538 is the transistor count (in millions). No idea what 208 would be though.
 
I’ve seen some theories that 629 is the die size and 7538 is the transistor count (in millions). No idea what 208 would be though.
That theory is being discussed above, with biggest issue being H100 being 80 Bn @ 814 mm^2 on same (or very similar) process. Possible explanations offered include more dark silicon (not counted?) in H100 and more dense caches in AD, but to get over 21 % better density from those seems outlandish.
The 208 has somewhere been suggested to mean 208 % performance of current gen
 
Calling that number plays a recorded message asking “how fast would you like to go”.
The latest teaser on Twitter has someone called User40 asking how much faster?

The overall focus of the massage is about speed this time, which means we should be seeing something substantially substantial.

RTX-40-ADA.jpg
 

208 is Idaho, capital is Boise, which is in Ada county.
 
Before the reveal, the looped screen on the broadcast has several interesting hints:

1-Logo with triangles .. new rasterization technique?
2-XXXXX with RTX .. some old game remastered with RTX?
3-G-Assist .. something related to G-Sync or related to Reflex?
4-Ti or Tie .. the reveal of 4090Ti?
5-R.O.N. .. previously an April's fool joke from NVIDIA about an AI assistant for gaming .. is this the same joke again? or a true project?
6-SETI .. stands for the Search for Extraterrestrial Intelligence, NVIDIA deployed their RTX hardware in the service of that goal in 2019. Why mention it now?
7-Version 208.629.7538 .. doubling down on the die size and transistor budget hints?
8-Project Beyond .. beyond what?


There's the big L2$ to consider. Nvidia is paying a lot of transistors to increase performance while barely budging on memory bandwidth. How would your analysis go if nvidia opted for HBM instead of a large cache to get the same performance with far fewer transistors?
AMD went from 10 billion transistor in RDNA1 (5700XT) to 27 billion in RDNA2 (6950XT), an increase of 2.7X times the budget, to achieve 2.33X the performance. Accounting for the caches doesn't change the outcome.
 
Last edited:
Over 21 % better density from just more cache?
Well, if I remember correctly, there was that NV blog post on using AI to design smaller circuits at a given performance level for Hopper. Maybe they applied the method much more broadly on Ada? But I agree that it seems more likely for the rumored transistor count number to be wrong.
 

Edit: Nvm. It was backside.

It's the same fan design principal as the existing RTX 3090/ti FE. Fans are slightly larger. Also going by the earlier leak the vapour chamber and cold plate design is different (larger). Hard to say at this point if fin density and therefore area is different or if the heatpipe configuration is different.

 
This is kinda nuts, crazy increases everywhere. Comparing their 3090 announcement - 69 "RT-Tflops" then vs 200 now, 285 Tensor-Tflops vs 1400 (surely Ada's not FP16?), 36 Tflops shader vs 90 so they're all massive gains if consistent. Morrowind RT/mod as well? Hello you have my attention

*They've quoted FP8 tensor performance for other products, I presume it's the same for the 4090 so 700 FP16
 
Last edited:
Status
Not open for further replies.
Back
Top