AMD Reportedly Readying RX 6900 XTX, Bringing the Battle to NVIDIA RTX 3090

[Vayra86]

So tsmc's 7nm is 91mm but got 51mm on rx 6900xt. So tsmc's scalling of sram and logic is lesser than advertised ?

The density of different blocks on a chip varies a lot.

As you well understood, my post was deliberately ignoring Nvidia there

There is a reddit post where detailed Turing die shots were analyzed. What he came up seems to be correct enough, Tensor cores and FP16 capabilities may be more nuanced but RT Cores are distinguishable and straightforward. RT Cores make up about 6% of TPC and about 3% of total die size. The increase for Tensor cores and/or FP16 capability concurrent to FP32 has more/most uses outside RT, same for cache. Implementation for AMD and Intel should not be too much different in terms of transistors and area cost, possibly less.

I wish there were good/readable enough die shots for RDNA2 and Ampere but apparently not so far. Would also need comparisons without RT and in case of RDNA where RT capability is part of some other block (TMU?) it is probably impossible to read.

3090 is on Samsung 8N, 6900XT is on TSMC N7:
- 3090 die is 28.3B transistors on 628 mm² - 45 MTr/mm²
- 6900XT die is 26.8B transistors on 520 mm² - 51 MTr/mm²
This highlights the differences in manufacturing processes more than anything.
In terms of transistors/area cost of latest improvements RDNA2 has huge amount of transistors (at least 6.4B plus some control logic which is 24% of total transistors) in Infinity cache, Ampere no doubt has a lot of transistors in the doubled ALUs in shaders.

More cache has been the go-to improvement for a few generations before RDNA and Turing. More likely than not adding more and more cache (at different levels) would happen with or without RT.

Assuming similar transistor density as 6900XT, 3090 on N7 would be 5.5% larger, about 30 mm².
That assumption is obviously suspect though. Without Infinity Cache 6900XT die would be noticeably less dense. On the other hand, there is A100 on TSMC's N7 with 54.2B transistors and 826mm² making the density out to 65,6 MTr/mm².

Great insights, thank you.

 

[londiste]

So tsmc's 7nm is 91mm but got 51mm on rx 6900xt. So tsmc's scalling of sram and logic is lesser than advertised ?

The maximum transistor density numbers are always for low power/high density variation of the process. High Performance variations have 60-70% of that maximum density. Also, the maximum numbers tend to be from something dense like memory.

TSMC themselves state 91.2 MTr/mm² for N7 (high density variation). A good rule of thumb is that high performance variation is 2/3 of that which would be ~60 MTr/mm². With logic being less dense than whatever the official numbers come from the 51 MTr/mm² for Navi21 is a respectable and expected result.

 

[wolf]

79% do not CARE about RT!

Were we even reading the same thing? Mind-boggling. There was literally an option for not caring about RT and it was selected by 28% of respondents.

Ahh well, small sample of us clearly flogging a dead horse at this point, can certainly see from the post reactions how a few of you feel and that I should seek the answers I want from a much wider and more balanced audience. We are in the thread about a 6900XTX after all.

 

[N3M3515]

No f**ks where given....

 

[wolf]

Seems that way. Willfully misrepresent the numbers, people who want to believe that reality upvote and feel better about themselves, damned with the truth. I think this way, a few other people agree, I must be right, right?

 

[Vayra86]

Seems that way. Willfully misrepresent the numbers, people who want to believe that reality upvote and feel better about themselves, damned with the truth. I think this way, a few other people agree, I must be right, right?

That is actually what you are misinterpreting here. Not trying to fight you - hear me out.

The poll sums up like this: 5 degrees of 'want' for emphasis on RT effects relative to the emphasis on rasterized performance.

And the 79% is the total of the poll outcomes that lean towards prioritizing raster performance - or at the very least that raster perf should never be LESS important than RT perf. Another way to say: RT is a nice to have, but not a must have.

And like I tried to explain earlier, a big part of that result is simply price and availability. Sure, if all things are equal you would pick the GPU with the bigger feature set. But they never really are even just on the basis of local differences. But also hardware wise; RDNA2 has unique selling points there, and so does Ampere. Its not a clear cut 'Nvidia is better' like it has been for quite some time.

Another element is the proprietary approach and how much trust you place in continued support. Nvidia has developed a track record there and it provides no guarantees.

 

[wolf]

That is actually what you are misinterpreting here.

Just calling out someone else's blatant misrepresentation, which several of you like reacted. I'll happily bow out.

 

[seth1911]

Great for a GPU i can buy a 2. hand Car and can drive with it the next few Years.

Ill take the Car instead of a Stupid GPU