Is a desktop APU similar to the Series S APU likely or even possible?

[g67575]

I've been thinking that the Series S's RDNA 2 based GPU seems quite impressive (120 fps in many games at 1080p).

Is it possible (given current production capacity of CPU/GPU dies) that we might see a similar spec, RDNA 2 based APU for desktops, with around 20 Compute Units? Something like this would probably make a lot of cheap graphics cards / old graphics cards obsolete. The whole APU has a tdp of just 100w.

It saddens a little me to see the last gen. RTX 2060 go for a similar price new (or more in some cases), as what I payed for my RTX 3070 FE... A hell of a good deal for Nvidia I bet.

 

[Poneros]

Nah, no chance. It's going to be 8/12 CUs at best, and also not having unified memory will hurt it further by comparison (plus not GDDR 6). The GPU situation is a mess, for sure, but there's no real solution for it that could also be implemented before the next generation of cards arrives. In the end people simply have to wait for Q4 '22/ Q1 '23 or pay extra for having one now. I imagine supply might still not meet demand but by then extra fab capacity will be available, as well as the rest of the component shortage should have lessened.

 

[g67575]

Nah, no chance. It's going to be 8/12 CUs at best, and also not having unified memory will hurt it further by comparison (plus not GDDR 6). The GPU situation is a mess, for sure, but there's no real solution for it that could also be implemented before the next generation of cards arrives. In the end people simply have to wait for Q4 '22/ Q1 '23 or pay extra for having one now. I imagine supply might still not meet demand but by then extra fab capacity will be available, as well as the rest of the component shortage should have lessened.

Simply because AMD has given priority to console APUs over desktop APUs, in 2020 and 2021? Couldn't this change by 2022? Why couldn't AMD produce powerful desktop RDNA 2 APUs with 8 or more GB of VRAM by next year?

I suppose what I'm wondering is, why isn't it possible just to mostly copy the design of the Series S APU, to create a desktop APU equivalent? Especially with AMD transferring production moving over to TSMC 5/6nm fab processes in 2022.

 

[CuriousTomCat]

The number of compute units is irrelevant because it will have DRR4. Surely not even a 6900XT could keep up with a series S if you slow it down with DDR4.

 

[g67575]

I'm wondering if AMD will save it's most powerful RDNA2 iGPUs for 5nm based Zen 4 chips in q3/q4 2022, based on this:
https://overclock3d.net/gfx/articles/2021/04/12103048819s.jpg

This rumour suggests Zen 4 CPUs will have iGPUs built in.

I'm not sure why AMD would stick with DDR4/5 for RDNA 2 APUs/iGPUs in 2022, based on the fact the latest consoles all use GDDR6 VRAM and released in 2020. I think the roadmaps are guessing this incorrectly.

The rumour mill incorrectly suggested Warhol CPUs and Rembrant APUs would be coming based on 6nm, but there's no sign of them. Instead, I think the 6nm process will be reserved for RDNA 3 and Intel high end GPUs.

There's also this slide from TSMC, showing their plans to increase 6nm (EUV) production to 48% (relative to 7nm) by the end of 2021. Which is a bit puzzling... A lot of this could be mobile chips tbf.

Link:
https://www.pcinvasion.com/wp-content/uploads/2021/06/tsmc-6nm-amd-7nm-ryzen-radeon-rdna-3-zen-3.jpg

 

[Journey]

I'm not sure why AMD would stick with DDR4/5 for RDNA 2 APUs/iGPUs in 2022, based on the fact the latest consoles all use GDDR6 VRAM and released in 2020. I think the roadmaps are guessing this incorrectly.

You do realise an APU is a CPU with a GPU on/integrated with it? It isn't designed to use GDDR RAM, it uses system RAM DDR4/5 etc. The only way they could put faster RAM on an APU would be by including HBM as part of the package, and that would make it very, very expensive.

A console chip is a custom design, with very specific design parameters and can't just be shoehorned into a desktop PC.

 

[g67575]

A console chip is a custom design, with very specific design parameters and can't just be shoehorned into a desktop PC.

Huh? The series S chip is still an APU though, with the GPU based on RDNA 2 architecture.

There's some speculation here that AMD is already selling these APUs in PCs, as then 4700s, only with the GPU die disabled. Link:
https://videocardz.com/newz/amd-4700s-might-be-xbox-series-xs-apu-for-pcs

You're right though, they would have to redesign it to include VRAM, as the VRAM chips are separate to the main APU chip in the Series S.

 

[Journey]

There's some speculation here that AMD is already selling these APUs in PCs, as then 4700s, only with the GPU die disabled. Link:
https://videocardz.com/newz/amd-4700s-might-be-xbox-series-xs-apu-for-pcs

So it is a CPU, not an APU if the GPU is disabled.

You're right though, they would have to redesign it to include VRAM, as the VRAM chips are separate to the main the main APU chip in the Series S.

I know, that is why I said it, but you can't put 16GB of DRAM on CPU package.

 

[PrincessFrosty]

Not to solve the GPU shortage problem, no. The GPU shorage is primarily due to competition to get hold of the wafers the fabs produce, and existing APUs share them with GPUs and many CPUs, as well as mobiles and all the other devices that are driving up demand. The bottleneck in supply of GPUs is way back in the supply chain right where the chips are made and that's because so many high demand products share the same resource.

I think the problem with low graphics power APUs on PC is that not only would that be very niche but also offer quite bad upgrade paths for users. One of the benefits of dGPUs is that you can have a system with a motherboard, RAM and CPU that is spec'd reasonably well and have it last about 6 CPU generation, in which time you can upgrade your GPU 2 or 3 times without swapping other parts. Much more awkward to do that with any kind of APU setup.

 

[g67575]

The costs of HBM2 VRAM was estimated to be $120 for 16GB in 2019 (without the cost of the package). Probably half the cost for 8GB, which is what the Series S has. So, not that expensive, assuming prices are roughly the same as in 2019.

Source here:
https://semiengineering.com/whats-next-for-high-bandwidth-memory/

I think it's more a question of, does AMD want to do it? and will there be sufficient production capacity?

If there's a full size PCI express slot on a mobo, you can install a graphics card, so don't see why anyone would be put off. AMD could always make versions without an iGPU, like Intel does.

This would basically mean that ~60 fps (console level) would be become the baseline level of performance for PCs with AMD APUs.

It's interesting that bespoke designs like the Series S are currently a much better deal than low-mid end gaming PCs right now. Is it in AMD or Nvidia's interests for this to continue?

 

[Journey]

The costs of HBM2 VRAM was estimated to be $120 for 16GB in 2019 (without the cost of the package). Probably half the cost for 8GB, which is what the Series S has. So, not that expensive, assuming prices are roughly the same as in 2019.

You are flip-flopping somewhat, first saying GDDR6, now saying HBM(2), which is it? GDDR6 would need to be on the main board, and adding a ~120mm2 HBM stack on to a CPU package means possibly a new socket, maybe? How big is the die for the Xbox, ~200mm2 or something close to that.

Don't get me wrong it would be amazing, but it isn't going to happen soon, maybe 2024 at the earliest when 3nm helps out.

 

[g67575]

@Journey if die space is a problem, perhaps AMD could use their new 3D stacking chiplet tech, to add HBM VRAM to the die above the main CPU chip, instead of extra L3 cache.

EDIT - apparently AMD may already be planning to use 3D stacking and HBM for Milan server chips, link here:
https://www.extremetech.com/computi...g-on-milan-x-with-3d-die-stacking-onboard-hbm

It might need some clever (possibly expensive) engineering to work around this problem, or, they might just simply opt for much larger die areas for Zen4 + AM5. Who knows at this point?

 

[KompuKare]

You are flip-flopping somewhat, first saying GDDR6, now saying HBM(2), which is it? GDDR6 would need to be on the main board, and adding a ~120mm2 HBM stack on to a CPU package means possibly a new socket, maybe? How big is the die for the Xbox, ~200mm2 or something close to that.

Don't get me wrong it would be amazing, but it isn't going to happen soon, maybe 2024 at the earliest when 3nm helps out.

Those AM5 renders do look deeper but that might just be the render as making AM5 thicker than AM4 would kill heatsink compatibility.

With so much soldered these days, LP-DDR5 might be possible for AM5 but not really sure what that gains consumers?
Higher bandwidth at lower power but at the cost of being soldered?
With probably a max of 16GB which will be too little for gaming sooner or later.
GDDRs are also soldered so again, for throw-away consumer stuff but an upgradeable desktop PC? No thanks! Soldered RAM on laptops is bad enough already.
HBM has been too expensive with the only mainstream parts really only AMD's Fuji and later Vega, and Intels Kaby Lake G with the Radeon Vega.
(Fuji's reason for existence was possibly to showcase HBM after AMD spend years developing it.)

The only future I see for mid/hight performing APUs is when AMD adds 3D stacked cache to some models. Since the cache benefits the CPU as well, that is going to be an interesting product.

 

[g67575]

Those AM5 renders do look deeper but that might just be the render as making AM5 thicker than AM4 would kill heatsink compatibility.

With so much soldered these days, LP-DDR5 might be possible for AM5 but not really sure what that gains consumers?
Higher bandwidth at lower power but at the cost of being soldered?
With probably a max of 16GB which will be too little for gaming sooner or later.
GDDRs are also soldered so again, for throw-away consumer stuff but an upgradeable desktop PC? No thanks! Soldered RAM on laptops is bad enough already.
HBM has been too expensive with the only mainstream parts really only AMD's Fuji and later Vega, and Intels Kaby Lake G with the Radeon Vega.
(Fuji's reason for existence was possibly to showcase HBM after AMD spend years developing it.)

The only future I see for mid/hight performing APUs is when AMD adds 3D stacked cache to some models. Since the cache benefits the CPU as well, that is going to be an interesting product.

Large amounts of L3 Cache is not really the point of buying an APU (it's expensive), current APUs are limited by graphics performance, particularly in TFlops and memory bandwidth.

I don't think anyone with an APU is worried about being limited to 16gb of VRAM (or even 8GB like the Series S APU). 8GB of HBM would be large upgrade vs DDR4/5.

The Kaby Lake G (2018, fabbed on 14nm) is a very interesting example of a PC APU integrating 4GB of HBM2, picture here:
https://en.wikichip.org/w/images/th...900px-kaby_lake_g_with_amd_radeon_package.png

But only a Quad core CPU. Still, it looks like there's still plenty of room for another 4GB of HBM VRAM, if Intel had wished to do so.

Full specs of the GPU here:
https://www.techpowerup.com/gpu-specs/radeon-rx-vega-m-gh.c3056

It appears to be one of the most powerful desktop APUs ever built (more powerful than a R9 380), current gen APUs are limited to ~2 TFlops of processing power and poor memory bandwidth. It was an unusual collaboration between AMD and Intel.

 

[Journey]

he Kaby Lake G (2018, fabbed on 14nm) is a very interesting example of a PC APU integrating 4GB of HBM2, picture here:
https://en.wikichip.org/w/images/th...900px-kaby_lake_g_with_amd_radeon_package.png

But only a Quad core CPU. Still, it looks like there's still plenty of room for another 4GB of HBM VRAM, if Intel had wished to do so.

HBM can be stacked already (and is) so there would have been no extra space need in the X, Y axis to accommodate the extra 4GB. What you are not seeing is the package size, it is ~60mm x ~30mm, an EPYC/TR package ~58mm x ~75mm, where as AM4 is only 40x40mm.

If anything happens it is more likely to be vertical (3D), and further down the line, as there is little point is designing a custom BGA solution, vs a socketable one, for a socket that will already exist.

 

[g67575]

I've read some articles suggesting 'Rembrandt' Zen 3(Zen3+ ?) APUs will be available in 2022 with DDR5 RAM support (therefore, these chips would be the first using the AM5 socket). If so, these would likely launch before Zen4 in Q4 2022. I think the trouble with this is, AMD is already launching new APUs in August this year, so maybe 1 year on from then?

Links here:
https://liliputing.com/wp-content/uploads/2021/05/amd-2022-1200x630.jpg
https://www.techpowerup.com/282637/amd-socket-am5-package-underside-pictured

Supposedly, AM5 will only support DDR5, which is a shame. So far, the rumours don't point towards 'Rembrandt' being that much more powerful than existing APUs.