Speculation: Ryzen 4000 series/Zen 3

[Glo.]

GCN vs RDNA2.

Which will be even more memory bandwidth reliant, than previous generations of AMD GPU architectures.

 

[uzzi38]

There are two versions that float around about VGH. That it is 16 CU APU and that it is 24 CU.

I don't believe in second option mainly for obvious reason: RX 5500 XT replacement will have 24 RDNA2 CUs, and there is absolutely no sense for AMD to release APU with such CU count.

If anything VGH has to be below that level in CU count.

Why? Because if by any chance this will not be premium-only part, and might land in mainstream market, it has to be below next generation small dGPU in performance.

The other problem with VGH is that in order to give decent performance level on 16 CUs it has to have HBM2.

And, adding to Uzzi's point I think it gives away completely for whom might this iGPU be. There are two, very specific customers who would enjoy low-power SoC, that have powerful CPU and iGPU side in their lightweight, premium laptops.

And those customers are direct competitors to each other.

Oh come on, I was planning on not giving away CU count precisely because I was hesitant on them. The first is what I heard though. 16CUs.

To fit in an APU at all, the CUs would have to be clocked rather low, probably around 1400mhz (depending on V/f curve for RDNA2). That's... not like AMD at all, and is exactly why I'm hesitant.

Also, I don't think it's HBM2. Pretty sure it's LPDDR of some kind, but relying on rumours again. I don't know. But with the RDNA2 CUs clocked so low and RDNA in general being so memory bandwidth efficient it might be OK?

Idk, taking quite a few guesses here. But you can probably imagine, if that's what the specs are, the performance will be pretty nice.

 

[uzzi38]

Which will be even more memory bandwidth reliant, than previous generations of AMD GPU architectures.

Lol no, RDNA in general is grwat with memory bandwidth. See the 5600XT, even with 3/4s the bandwidth or less in it's original format the GPU does great and performs only a few percentage points behind a 5700. Increasing core clocks even at this stage can improve performance notably which also shows the GPU isn't bottlenecked significantly.

RDNA1 had bottlenecks elsewhere in the design, either cache wasn't enough to keep the GPU fed properly or somewhere in the fixed function hardware AMD skimped out. But in terms of memory bandwidth usage RDNA is fine.

That is, unless AMD allows RTRT on Van Gogh. Uh... then I'm wrong, very wrong. That stuff is a huge memory bandwidth strain.

 

[Tup3x]

Could LPDDR5 not provide sufficient bandwidth for a powerful GPU?

But I agree that 24 CUs just seems bonkers. The PlayStation 5 only has 32 CUs, I don't see a laptop APU getting two thirds of the way there. 16 CUs, lower clocks, and LPDDR5 is my guess. That would make Apple very happy.

It would be awesome if they would make console class APUs though. OEM/motherboards manufactures could then sell motherboards with different memory configurations and APUs (all soldered). It would be enough for many and could open up some rather nice compact form factors. I hope that'll eventually happen... but I doubt that will happen anytime soon. It would make upgrading simple since you'd only have to buy one thing.

 

[Glo.]

That is, unless AMD allows RTRT on Van Gogh. Uh... then I'm wrong, very wrong. That stuff is a huge memory bandwidth strain.

THAT is the reason why you are wrong .

RDNA2 will have RTRT top-to bottom, at least based on what I am hearing. And here comes the interesting part of what I heard. Mero and Van Gogh could be the same APU, but one with RTRT, and the other with it disabled.

If that is the case, one has HBM2 the other - does not.

The memory bandwidth reliance is the reason why next gen small dGPU, RX 5500 XT replacement with 24 CUs, might have 192 bit GDDR6 memory bus with over 300 GB/s.

 

[uzzi38]

THAT is the reason why you are wrong .

RDNA2 will have RTRT top-to bottom, at least based on what I am hearing. And here comes the interesting part of what I heard. Mero and Van Gogh could be the same APU, but one with RTRT, and the other disabled.

If that is the case, one has HBM2 the other - does not.

I think Van Gogh is based off Mero from what I'd seen in the aforementionned unrelated-to-the-huge-mess Github repo, so uh, yeah. I would assume VGH is the cut down variant.

Somewhere down the line (around April/May 2019) references Mero is no longer mentionned alone and is instead mentionned as "Mero/VGH".

 

[Glo.]

I think Van Gogh is based off Mero from what I'd seen in the aforementionned unrelated-to-the-huge-mess Github repo, so uh, yeah. I would assume VGH is the cut down variant.

Somewhere down the line (around April/May 2019) references Mero is no longer mentionned alone and is instead mentionned as "Mero/VGH".

I don't know which one is based on which. Maybe you are correct.

 

[NTMBK]

Hardly, for example RX 560 has 16CU and 112 GB/s of memory bandwidth - LPDDR5-5200 translate to "just" 83.2 GB/s and it must be shared with CPU.

Samsung is talking about LPDDR5-6400, which would give >100GB/s on a 128-bit bus. Or a semicustom part could go for a wider bus, like 192-bit- if you're replacing a dGPU+CPU combo, you're still ending up with a lower number of memory chips and traces overall.

 

[rainy]

Samsung is talking about LPDDR5-6400, which would give >100GB/s on a 128-bit bus. Or a semicustom part could go for a wider bus, like 192-bit- if you're replacing a dGPU+CPU combo, you're still ending up with a lower number of memory chips and traces overall.

At this moment, LPDDR5-5500 is the fastest, 6400 variation would be released later.
Wider bus, it's a different story - 192-bit plus LPDDR5-5500 will deliver 132 GB/s and this should be enough for 16CU, especially with RDNA 2 improvements over GCN parts like Polaris.

 

[Shivansps]

Well they could go with just a smaller version of console APUs that has GDDR6 only. For a premium, niche product what would work and the R&D is already done. I dont think Apple would care too much about using GDDR6, in fact, im petty sure they would love the idea of no upgradeable ram.

But i would think that high speed LPDDR5 on 128 bit bus is the way to go for everything else. Just like they did on Renoir, maybe is not avalible now, but it will eventually.

 

[maddie]

Which will be even more memory bandwidth reliant, than previous generations of AMD GPU architectures.

Meaning? Are you saying that bandwidth/CU need rises with RDNA2?

 

[NostaSeronx]

We don't know yet if RDNA2 is the Super-SIMD architecture, either.

GCN => 1 CU -> 1 Scheduler -> 1 SIMD64 unit
RDNA1 => 1 CU -> 2 Schedulers -> 2 SIMD32 units
Super-SIMD => 1 CU -> 2 Two-issue Schedulers -> 4 SIMD32 units <== Same ALU rate as the previous dual-CU in a single-CU.

One of the big instruction additions for Super-SIMD is psuedo-CISC mode; ALU op + Read or Store op.

 

[Glo.]

Meaning? Are you saying that bandwidth/CU need rises with RDNA2?

Ray Tracing technology in RDNA2 will require quite a lot of memory bandwidth.

As I have said in one of previous posts. There is a chance that next gen low-end might START with 192 bit memory buses and 6 GB's of VRAM, because of this very reason, based on what I am hearing. Next gen Navi 14 replacement may look something like this: 24 CUs, 192 bit memory bus. There is a good reason why David Wang said some time ago that Ray Tracing will not get wide adoption till you will have RT capable hardware in top-to-bottom stack.

Memory bandwidth requirements will go through the roof with next gen GPU tech.

 

[lobz]

Which will be even more memory bandwidth reliant, than previous generations of AMD GPU architectures.

while also having much better memory compression.

 

[jpiniero]

Ray Tracing will not get wide adoption till you will have RT capable hardware in top-to-bottom stack.

Not really. It's all about how quickly developers adopt it with the new consoles in mind, and how quickly they ditch the current gen.

 

[Glo.]

while also having much better memory compression.

While also REQUIRING much better memory compression.

Fixed, for you .

 

[lobz]

While also REQUIRING much better memory compression.

Fixed, for you .

So RDNA2 won't have much better memory compression in your opinion. They just brainlessly put in features that need a lot more BW and call it a day. In that case, thank you for the fix..... Geez

 

[maddie]

Ray Tracing technology in RDNA2 will require quite a lot of memory bandwidth.

As I have said in one of previous posts. There is a chance that next gen low-end might START with 192 bit memory buses and 6 GB's of VRAM, because of this very reason, based on what I am hearing. Next gen Navi 14 replacement may look something like this: 24 CUs, 192 bit memory bus. There is a good reason why David Wang said some time ago that Ray Tracing will not get wide adoption till you will have RT capable hardware in top-to-bottom stack.

Memory bandwidth requirements will go through the roof with next gen GPU tech.

Thanks for that. Am looking forward to be raytracing with my APU.

 

[moinmoin]

To me Van Gogh and Mero seem pretty clear semi custom solutions so far. Similarly to the specs of the two upcoming new consoles there are parts that I'm pretty sure AMD on its own wouldn't want to pursue in a mainstream part it offers as its own product.

 

[uzzi38]

To me Van Gogh and Mero seem pretty clear semi custom solutions so far. Similarly to the specs of the two upcoming new consoles there are parts that I'm pretty sure AMD on its own wouldn't want to pursue in a mainstream part it offers as its own product.

If we define mainstream as what RR/PCO/RNR are, then VGH is absolutely not mainstream. Not even remotely so. Budget is RAVEN2 and so on. As far as I know we don't know a codename for a future successor to VGH/Mero either, so they could easily be one-offs. Ultimately, Mero is definitely semi-custom, and from what I can see Van Gogh is based off it. Whether it's more semi-custom for a single client or it's a slightly more 'mass-market' (kinda not really) version of Mero that's cheaper and easier to implement but still on the pricy side and focuses on an aspect most laptop manufacturers don't care about (iGPU) is still unknown. But in either case I wouldn't call it mainstream.

 

[DisEnchantment]

Thanks for that. Am looking forward to be raytracing with my APU.

Why... You dont like Ray Traced shadows in SolidWorks or CATIA in a notebook for those sleek presentations to prospective clients?

 

[RetroZombie]

Which one is the premium $$$ cpu, the 4800u or the 4800h model?

Hardware unboxed discussed the threadripper mobile possibility in their last video, due to intel high tdp mobile cpus where amd could do the same:

 

[Glo.]

So RDNA2 won't have much better memory compression in your opinion. They just brainlessly put in features that need a lot more BW and call it a day. In that case, thank you for the fix..... Geez

What I meant by that is Ray Tracing tech in next gen GPUs requires BOTH: much better memory compression technology, and much higher Memory bandwidth.

 

[Glo.]

Thanks for that. Am looking forward to be raytracing with my APU.

Ray Tracing is not only for games, and we should all know that by now.

Which is the whole point of having RT tech in hardware from top to bottom.

 

[uzzi38]

4800U score. Not much to say that isn't said in the post itself. cTDP is likely 25W, which is enough to allow it to score better than the 10750H in CPU and the iGPU edges out the 4800H as well. Though it's worth remembering it does have an extra CU and higher iGPU clocks to help get there.

But to be honest, the feeling I'm getting here is the APU isn't horribly memory bottlenecked, even with 8CUs clocked at 1850mhz. Not in Fire Strike at least, though Time Spy pushes your memory config harder, so will be better to really put the memory subsystem to the test.

https://twitter.com/x/status/1246763743620579329

EDIT: Comparison vs 4800H and 4900HS as well. Latter has the same iGPU configuration, just higher power budget and better binned CPU cores.

https://twitter.com/x/status/1246771540118642689