AMD Execution Thread [2023]

[DavidGraham]

It looks different if you assume the future of PC gaming is APU.

That argument is recurrent since 10, 15 or even 20 years ago, it always comes up whenever NVIDIA seems to have a monopoly on dGPUs, as if APUs can abolish that monopoly.

But in short, that argument has always failed the test of time, especially now, prices of silicon is higher than ever, GPUs are bigger than ever, power has increased substantially too, thus cooling became more complex as well. Bandwidth needs are much higher, and caches alone can't help, you now need caches + bigger memory bus + faster clocks to achieve good bandwidth, and HBM remains too expensive and too scarce to help.

And it gets worse, CPUs have fared much worse than GPUs, with massive power and temps increases, cooling became much more complex than GPUs as a result. 10 years ago a typical CPU consumed 65w on air cooling, now its in the range of 170w with water cooling. Memory became more power hungry than ever too.

So in short, high end APUs are not happening, the factors that could have made them happen 10 years ago are no longer here now. The future is clear now, bigger, and hotter chips with exotic cooling are the way to go to satisfy our ever increasing computing itch: high refresh rate gaming, high resolution, ray tracing/path tracing/lumen, machine learning/artificial intelligence ...etc.

 

[Nisaaru]

Rumor: Kepler is claiming RDNA4 is not going to have any high end options, just like RDNA1.

https://twitter.com/x/status/1687388515670056960

Makes no sense to me with their expected move into multi chip modules for GPUs unless they canceled these plans.

 

[del42sa]

.... unless they canceled these plans.

seems they did https://forum.beyond3d.com/threads/amd-execution-thread-2023.63186/page-9#post-2308957

to resume RTG strategy :

*they don´t want to build huge die GPU aka Fiji
*they don´t want their GPU become too complex (more GCD´s chiplets, interposer and expensive interconnect)
*they don´t need direct competition to all Nvidia GPU
*they can build faster GPU, but they don´t want to
*they can fix RDNA3 but they won´t bother
*they only target max. $999 price as acceptable by gamers
*they don´t care about dGPU market share because they have console

conclusion, don´t expect too much from RDNA4

 

[DegustatoR]

conclusion, don´t expect too much from RDNA4

I mean wouldn't a 7900XTX perf level at ~$500 be exactly what people seem to want more than anything?

 

[Bondrewd]

to resume RTG strategy :

*they don´t want to build huge die GPU aka Fiji
*they don´t want their GPU become too complex (more GCD´s chiplets, interposer and expensive interconnect)
*they don´t need direct competition to all Nvidia GPU
*they can build faster GPU, but they don´t want to
*they can fix RDNA3 but they won´t bother
*they only target max. $999 price as acceptable by gamers
*they don´t care about dGPU market share because they have console

That's not what it is but I digress.

I mean wouldn't a 7900XTX perf level at ~$500 be exactly what people seem to want more than anything?

Don't think you're getting that either.

 

[del42sa]

wait for RDNA5 ... Lol

 

[Bondrewd]

wait for RDNA5 ... Lol

Correct!
That's where the horryfing tiled dGFX parts now live at.

 

[DavidGraham]

wait for RDNA5 ... Lol

It's probably the case that getting compute chiplets to work with graphics is a nightmare to go through.

 

[techuse]

I mean wouldn't a 7900XTX perf level at ~$500 be exactly what people seem to want more than anything?

If they release that next year it would be fantastic but that’s almost certainly not what is going to happen. Im thinking RDNA 3 all over again but they only release the 7600 equivalent.

 

[WhiningKhan]

10 years ago a typical CPU consumed 65w on air cooling, now its in the range of 170w with water cooling. Memory became more power hungry than ever too.

I object your idea of "typical". You are describing enthusiast CPUs there, which are pretty small part of market.
E.g. in this household there are 3 PCs. Their current CPUs are Ryzen 5600, 5600x and 5600G. All 65W parts.

Steam hw survey indicates huge majority of systems with this or lower class.

 

[JoeJ]

So in short, high end APUs are not happening, the factors that could have made them happen 10 years ago are no longer here now. The future is clear now, bigger, and hotter chips with exotic cooling are the way to go to satisfy our ever increasing computing itch: high refresh rate gaming, high resolution, ray tracing/path tracing/lumen, machine learning/artificial intelligence ...etc.

Only an enthusiast niche will continue to buy this argument.
They say we need to buy and run a data center HW at home just to play games.
But the increasing majority will say goodbye to chasing photorealism for premium costs. They just want to play some games for fun.

People do not want photorealism, they only want to see progress. If Switch2 looks better than Switch1, it makes them just as satisfied as comparing PS5 to 4.
So by accepting the lower standard they save money, and can still enjoy the progress beside the games. That's what most will continue to do, looking at current PC->console player movement.
If PC fails to find the sweet spot, some other platforms doing better will eat it's market share, until even the enthusiast niche becomes too small to survive, no matter how ambitious and willing to spend they are.

That's why you need the affordable APU mainstream just as much as i want it. \/
They will pick a solution for the bandwidth problem, and 100tf are not needed.

 

[Bondrewd]

high end APUs are not happening

Bro wtf is Strix-Halo then?

 

[Seanspeed]

Bro wtf is Strix-Halo then?

Lots of information bounding about for it, but I'm still unsure on how AMD expects to solve the bandwidth problem.

 

[Bondrewd]

but I'm still unsure on how AMD expects to solve the bandwidth problem

MALL + 256b L5x@8533.
Very simple, really.

 

[techuse]

That argument is recurrent since 10, 15 or even 20 years ago, it always comes up whenever NVIDIA seems to have a monopoly on dGPUs, as if APUs can abolish that monopoly.

But in short, that argument has always failed the test of time, especially now, prices of silicon is higher than ever, GPUs are bigger than ever, power has increased substantially too, thus cooling became more complex as well. Bandwidth needs are much higher, and caches alone can't help, you now need caches + bigger memory bus + faster clocks to achieve good bandwidth, and HBM remains too expensive and too scarce to help.

And it gets worse, CPUs have fared much worse than GPUs, with massive power and temps increases, cooling became much more complex than GPUs as a result. 10 years ago a typical CPU consumed 65w on air cooling, now its in the range of 170w with water cooling. Memory became more power hungry than ever too.

So in short, high end APUs are not happening, the factors that could have made them happen 10 years ago are no longer here now. The future is clear now, bigger, and hotter chips with exotic cooling are the way to go to satisfy our ever increasing computing itch: high refresh rate gaming, high resolution, ray tracing/path tracing/lumen, machine learning/artificial intelligence ...etc.

Do AMD CPUs typically consume that level of power? Isn’t it just Intel’s poorly engineered products?

 

[Bondrewd]

Do AMD CPUs typically consume that level of power?

Yeah vanilla Raphael has a 230W PPT or therein (it kinda does not much there Vs 160W PPT but I digress).

 

[Seanspeed]

MALL + 256b L5x@8533.
Very simple, really.

lol

Super reasonable RAM expectations....

 

[DegustatoR]

256b LPDDR5X is nowhere near "high end".
It may be a high end APU but it won't compete with high end GPUs.

 

[Bondrewd]

Super reasonable RAM expectations....

Yea you get decent amount of RAM (up to like 48GB).

256b LPDDR5X is nowhere near "high end".

It's very much high-end by APU standards.

but it won't compete with high end GPUs.

You can always build a 512b one with more MALL and on-carrier memory.
Is there a market for that? not really.

 

[trinibwoy]

But the increasing majority will say goodbye to chasing photorealism for premium costs. They just want to play some games for fun.

People do not want photorealism, they only want to see progress.

Sure but photorealism is still a long ways off. The immediate problem is price / performance stagnation in all market segments so even incremental IQ gains are increasingly expensive. Gaming at 1080p medium settings is getting more expensive not less. I don’t see how APUs solve that problem. They’re subject to the same silicon economics and physics as discrete CPUs and GPUs.