Question Speculation: RDNA2 + CDNA Architectures thread

[uzzi38]

All die sizes are within 5mm^2. The poster here has been right on some things in the past afaik, and to his credit was the first to saying 505mm^2 for Navi21, which other people have backed up. Even still though, take the following with a pich of salt.

Navi21 - 505mm^2

Navi22 - 340mm^2

Navi23 - 240mm^2

Source is the following post: https://www.ptt.cc/bbs/PC_Shopping/M.1588075782.A.C1E.html

 

[TESKATLIPOKA]

Is there somewhere at least one laptop with 5700M?

 

[uzzi38]

Is there somewhere at least one laptop with 5700M?

Some Dell laptop with only Comet Lake as the CPU choice.

Which is why nobody realised the device exists.

 

[skline]

I'll try to check the stepping of my 5600XT tonight with GPU-Z.

 

[KompuKare]

BTW I have to wonder when AMD finally fixes the high power consumption using two monitors, because this was not fixed for many years and It doesn't make a good impression.

That's a tricky one. iGPU is by far the best at running multiple monitors sipping a few watts at most, and that's same whether Intel or AMD AFAIK.
With dGPUs, for dual monitors, AMD ramp up the memory clocks to gaming levels AFAIK. While Nvidia only do that for three monitors (which TPU don't test - so must have seen those results elsewhere).
Think some people have set the multi-monitor memory clocks lower themselves with AMD cards and that mostly works, so it really looks like the drivers are just playing safe most of the time. Which would imply an embarrassing driver issue which hasn't been fixed for years.

 

[uzzi38]

I'll try to check the stepping of my 5600XT tonight with GPU-Z.

It's show a different value altogether. Iirc either C0 or D0 or something like that. >.>

 

[Mopetar]

I can understand why some people are pessimistic about RDNA2 considering the past, but we know big Navi has 80CU so I find It unreasonable when someone says It will perform between RTX 3060-3070. We know RTX 3090 has "only" 5248 Cuda so RTX 3070 will probably have 3072 Cuda cores and It's not possible that a 5120 core RDNA2 GPU performs the same as Ampere with 3072 cores. Ampere would need to have much better IPC and clockspeed compared to Turing or RDNA1(2) to erase the difference in the number of execution units between Big Navi and RTX3070.

Compare the shader counts for Fury/Vega against the NVidia cards at the time and it looks suspiciously like the same argument. And yet the results spoke for themselves.

Obviously the argument is that RDNA allows AMD to get around GCN limitations that made the additional shaders they had over the competition rather ineffectual, but AMD has had a history of more shaders that didn't perform as well as NVidia's fewer shaders.

It's easy to see why some people may be being overly pessimistic if they weren't spending a lot of time trying to understand why the past results were the way that they were.

 

[TESKATLIPOKA]

But in this case Ampere most likely won't be significantly different than Turing, so I don't expect significant IPC gain or much higher clockspeed to offset the difference in execution units.
It's possible I have a biassed view in this considering I have followed gpu architectures for some time now in comparison to some who may have started not so long ago or just weren't interested in them to such an extend.

 

[Stuka87]

Compare the shader counts for Fury/Vega against the NVidia cards at the time and it looks suspiciously like the same argument. And yet the results spoke for themselves.

Obviously the argument is that RDNA allows AMD to get around GCN limitations that made the additional shaders they had over the competition rather ineffectual, but AMD has had a history of more shaders that didn't perform as well as NVidia's fewer shaders.

It's easy to see why some people may be being overly pessimistic if they weren't spending a lot of time trying to understand why the past results were the way that they were.

This is partly because the chips were compute heavy cards. And at launch this did nothing. The cards supported features that would not be used for years. Which did give the cards long lives. A 7970 for instance performed significantly better than a GTX 680 5 years down the road. But that was/is kind of a moot point.

RDNA seems to offer much better performance for a given shader count than GCN did, which is partly due to AMD making RDNA less compute heavy, and leaving that for CDNA where it belongs.

 

[A///]

Zotac's leaked Twin Edge 3070 with dual fan and no bling looks awesome. How are Zotac cards these days? It also looks really compact.

RDNA seems to offer much better performance for a given shader count than GCN did, which is partly due to AMD making RDNA less compute heavy, and leaving that for CDNA where it belongs.

I learned this a few months ago when I looked into it myself. I wasn't sure why AMD were still relying on GCN/CDNA cards almost a year into RDNA. The more you know!

 

[NTMBK]

I learned this a few months ago when I looked into it myself. I wasn't sure why AMD were still relying on GCN/CDNA cards almost a year into RDNA. The more you know!

Possibly due to 7nm supply. They can only get so many wafers from TSMC, better to use them on high margin Epyc than on an RX590 replacement.

 

[A///]

Possibly due to 7nm supply. They can only get so many wafers from TSMC, better to use them on high margin Epyc than on an RX590 replacement.

And this has to do with CDNA being better for compute how?

 

[uzzi38]

Possibly due to 7nm supply. They can only get so many wafers from TSMC, better to use them on high margin Epyc than on an RX590 replacement.

590 replacement?

CDNA?

You what now?

 

[NTMBK]

590 replacement?

CDNA?

You what now?

I was talking about "why AMD were still relying on GCN cards". Easy to keep pumping out low-margin RX590 cards on GloFo 12nm, where chip supply is plentiful, instead of moving budget cards on to 7nm.

 

[uzzi38]

I was talking about "why AMD were still relying on GCN cards". Easy to keep pumping out low-margin RX590 cards on GloFo 12nm, where chip supply is plentiful, instead of moving budget cards on to 7nm.

Er, just wondering, but has nobody mentioned Navi24 here yet?

https://twitter.com/x/status/1291796703767457792

 

[TESKATLIPOKA]

Navi24? Interesting. I wonder how much CU It will have.
Will they update the whole RDNA1 lineup?

 

[uzzi38]

Navi24? Interesting. I wonder how much CU It will have.
Will they update the whole RDNA1 lineup?

My guess is that all RDNA 2 GPUs this time around will have 5WGPs per SA.

So it'll be a simple 80/60/40/20 in the order of the codenames.

 

[A///]

All of this is nice and all but I want to see a product from AMD before the end of the year. Don't care about rumors. I want to see products.

 

[Qwertilot]

Navi24? Interesting. I wonder how much CU It will have.
Will they update the whole RDNA1 lineup?

If they've genuinely got anything like a 50% perf/watt improvement then you have to imagine that with the money from Zen coming in they'll fund a top to bottom refresh of their GPU's.

 

[uzzi38]

All of this is nice and all but I want to see a product from AMD before the end of the year. Don't care about rumors. I want to see products.

FFS AMD have said that's happening multiple times already.

Navi21 will launch this year.

 

[uzzi38]

If they've genuinely got anything like a 50% perf/watt improvement then you have to imagine that with the money from Zen coming in they'll fund a top to bottom refresh of their GPU's.

Been saying for a while now, Zen's physical optimisation team found their way over to RTG and have been working some magic.

 

[A///]

FFS AMD have said that's happening multiple times already.

Navi21 will launch this year.

Are you sure? I'd like Lisa Su to confirm a 10th time just for clarity. You missed the point of my post.

 

[DDH]

Been saying for a while now, Zen's physical optimisation team found their way over to RTG and have been working some magic.

What is a physical optimisation team?

 

[uzzi38]

What is a physical optimisation team?

Somebody else can probably give you a real explanation as I'm really bad at it, but so your modern chip is dozens if not hundreds of different IPs pieced together. Often you'll use a program to put these together for the final layout of your design. A physical optimisation team is a team that will go and make small tweaks here and there to try and tweak and tune different areas of the chip by hand effectively. The end result is usually improved V/f properties in some way or another, depends on what the aim of said physical optimisation team was.

 

[kurosaki]

ody else can probably give you a real explanation as I'm really bad at it, but so your modern chip is dozens if not hundreds of different IPs pieced together. Often you'll use a program to put these together for the final layout of your design. A physical optimisation team is a team that will go and make small tweaks here and there to try and tweak and tune different areas of the chip by hand effectively. The end result is usually improved V/f properties in some way or another, depends on what the aim of said physical optimisation team was.

Back in the days, very much was done like this at Intel, while AMD took the cheaper approach going through software. Good to hear they have people in place squeezing out the last percents as well now a days!

 

[TESKATLIPOKA]

My guess is that all RDNA 2 GPUs this time around will have 5WGPs per SA.

So it'll be a simple 80/60/40/20 in the order of the codenames.

It look simple at first, 20CU more for each tier, but there will be an uneven performance difference between each tier.
20 -> 40CU 100% increase
40 -> 60CU 50% increase
60 -> 80CU 33% increase