AMD: Navi Speculation, Rumours and Discussion [2019-2020]

Status
Not open for further replies.
About the bandwidth dependence of NAVI 10, I think a good idea would be checking the differences between the 5600XT, 5700 and 5700XT. I am looking at Techpowerup averages (because I'm lazy) and the fact that the 5600XT manages to reach around 79% of the performance of 5700XT and it is a few percent less performing than the 5700 when having 64% less available bandwidth, tells me the bandwidth itself is not the major culprit there. It seems more a question of shading or, probably primitives/scheduling as the performances seems quite in line with the ratio of the game clocks for all three cards.

EDIT: just saw Bondrewd's post, but I have similar concerns.
 
Last edited:
Bondrewd said:
5600XT is the living example.
6% less perf for 25% memory b/w chopped off.
Literally made 5700 obsolete (unless you need extra 2GB VRAM).
Click to expand...

I'd almost forgotten about that card. You're right, its barely 6-8%.

So if we take the 36 CU 5600XT with its 192 bit memory bus @ 14 Gbps as a baseline, an 80 CU N21 with 384 bit @16/17 Gbps with some cache and delta compression algorithm improvements should do fine.

Edit: @Leoneazzurro5, yes I was just checking that same comparison on TPU! I think bandwidth isn't really going to be a major concern in the end.
 
5600XT is already on the verge of losing perf due to bandwith. With the numbers above, Big Navi had +142% bandwith, but # of CUs alone is already +122% - without µarch improvements and (way?) higher clocks.
 
@CarstenS: how do we know there are no microarchitectural improvements? High level architecture partitioning says nothing about low-level ol algorithm optimization, and those were not discussed in the Microsoft workshop. I mean, you can be right, but we cannot know that now.
 
Erinyes said:
Has anyone ever tried downclocking the memory of either 2070/RX5700 and seeing how much the performance drops? I'd wager its not linear. Conversely, memory overclocking never gets you a linear performance increase. While it may be a bit short on bandwidth with a 384 bit GDDR6 memory bus, I dont expect it to have a significant impact on the performance of Big Navi.
Click to expand...

It's not going to be linear. You can just look at TPU boost clocks and get "boost TFLOPs" for GPUs using the same architecture and memory bandwidth and compare the performance summary chart. The 5700 has a better perf/boost TFLOPs ratio than the 5700xt (by the way the 5600xt actually has a slightly higher boost TFLOPS vs 5700). On the Nvidia side the 2060 Super > 2070 > 2070 Super > 2080.

A problem is with the use of "bottleneck" as the term typically used to describe these situations, it tends to suggest that the scenarios are either negligible limitations (essentially all gains) or huge limitations (to a trickle) when it's more of a diminishing returns type situation the more imbalanced any facet of performance is to others.

In terms of a Big Navi hypothetical if memory bandwidth doesn't increase 2x (whether theoretical or real through other improvements) if the rest of it (such as the CUs) achieves >2x speed up thnn Big Navi can still be 2x faster versus Navi 10 (5700xt). Memory bandwidth itself is not going to be an absolute limiter especially if it's across a suite of workloads (as some will be more or less b/w dependent).

Alternative viewpoints is you can ballpark around 2.3 x 5600xt if the concern is memory bandwidth scaling (384x16/192x14).
 
Leoneazzurro5 said:
@CarstenS: how do we know there are no microarchitectural improvements? High level architecture partitioning says nothing about low-level ol algorithm optimization, and those were not discussed in the Microsoft workshop. I mean, you can be right, but we cannot know that now.
Click to expand...
Sorry, maybe I did not express myself clearly: I meant, even without improvements that in turn COULD require more bandwidth like more TMUs - while other improvements could offset this again. I did not mean to assert, that there are no improvements in either way, which I wouldn't know or wouldn't be allowed to talk about just yet. I was just taking those out of the equation, if you will.
 
Bondrewd said:
Gladly but we have a month (nearly exactly) left to talk before all hells go loose.

The contribution is higher up in the chain.
Click to expand...

I mean, you could easily debate with a "no I did not see the desktop marketing material but the very same persons showing that to me and other various sources I have told me that the same is valid in the desktop, too" instead of mocking the interlocutor, and your contribution could be much more appreciated by everyone.
 
Leoneazzurro5 said:
I mean, you could easily debate with
Click to expand...
I've already told him that it scales both wide and fast, see consoles, yet he tried to akshually me.
Not even quoting any "sources", just fairly publically available data you can see or deduce for yourself.

Fucks sake Maxwell and Pascal scaled the same way yet somehow N2x will magically break at 4SE variant just to make nVidia win in the high-end.
Just no.
 
Status
Not open for further replies.
Back
Top