AMD Raven Ridge 'Zen APU' Thread

[Mopetar]

And how much will a 704 SP GPU really benefit from HBM? Not enough IMO.

This might happen eventually, but I seriously doubt it will be for Raven Ridge.

I think the eventual plan is that the HBM replaces all of the RAM and AMD delivers what is essentially an X86 system on a chip for notebooks and desktops.

I suspect that on-die HBM would have better latency than regular memory and could give AMD a significant leg up in some workloads.

Right now the big problem is cost and capacity. You'd want at least 8GB if you weren't going to have other RAM, but at the moment that seems difficult for the manufacturers to do, or at least for the price that AMD would need it to be and using more than a single stack just balloons die size and cost.

 

[ao_ika_red]

Why not both though? It would obviously be the same die, so I don't think them supporting a premium option with HBM is that far fetched. Even if they're not doing it with Raven Ridge it's definitely going to happen eventually. I feel like a premium ultra-portable with HBM2 could be a really appealing product.
That's why I'm wondering about latency and power consumption, since that would be deal breakers in an ultra-portable. It's only a matter of time until an HBM APU is happening, what I'm curious about is if it's feasible to replace DDR4 with HBM entirely, and especially in a ultra-portable laptop.

As far as I know, the only AMD client that try to push HBM on APU is Apple. It makes sense because Apple will able to make very good ultrabook (MacBook in this case). But, it will be an hybrid option(HBM +DDR4) because HBM is more expensive than DDR4 and also CPU doesn't work well with graphics memory.

What does an HBM interface add to the die for connections/complexity?

I notice that AMD has never used the same GPU die for both GDDR and HBM, so there is likely a significant hit to have dual memory interfaces.

And how much will a 704 SP GPU really benefit from HBM? Not enough IMO.

This might happen eventually, but I seriously doubt it will be for Raven Ridge.

Manufacturing GPU with HBM is a collaboration work.

I quote it from 3DInCities, “This project required collaboration with 20 different companies and government organizations before delivering the final product,” said Black. “We worked with many partners along the way. Some disappeared, some delivered. Some we’re still working with.” The final product integrates graphics chips from TSMC, OSAT partners ASE and Amkor; HBM was procured from SK Hynix, and UMC provided the interposer.

Because there's more company involved in this project, there's also more cost and failure rate, inevitably. That's why Fiji and Vega came very late to market.
HBM memory controller also behave differently than DDR because it works in much wider bus and lower latency. That's why I found it interesting because if somehow they work it out (hybrid HBM-DDR controller, or some sort of) the cost of non-HBM and HBM APU will be lowered because they will get diverse products from one die.
Lots of people are crying for HBM APU because in mobile space, there's bandwidth starvation in high end (8 CU) APU and unlike desktop APU, mobile users have limited option to address the problem. But sometimes people tend to forget that Zen-Vega APU works on much faster DDR4 memory. Vega also implement advanced memory compression tech that will help in limited bandwidth environment. That's why I think majority of RR APU won't get HBM and only limited to certain SKU (like Intel Broadwell-Iris Pro SKU).

 

[CatMerc]

As far as I know, the only AMD client that try to push HBM on APU is Apple. It makes sense because Apple will able to make very good ultrabook (MacBook in this case). But, it will be an hybrid option(HBM +DDR4) because HBM is more expensive than DDR4 and also CPU doesn't work well with graphics memory.



Manufacturing GPU with HBM is a collaboration work.

I quote it from 3DInCities, “This project required collaboration with 20 different companies and government organizations before delivering the final product,” said Black. “We worked with many partners along the way. Some disappeared, some delivered. Some we’re still working with.” The final product integrates graphics chips from TSMC, OSAT partners ASE and Amkor; HBM was procured from SK Hynix, and UMC provided the interposer.

Because there's more company involved in this project, there's also more cost and failure rate, inevitably. That's why Fiji and Vega came very late to market.
HBM memory controller also behave differently than DDR because it works in much wider bus and lower latency. That's why I found it interesting because if somehow they work it out (hybrid HBM-DDR controller, or some sort of) the cost of non-HBM and HBM APU will be lowered because they will get diverse products from one die.
Lots of people are crying for HBM APU because in mobile space, there's bandwidth starvation in high end (8 CU) APU and unlike desktop APU, mobile users have limited option to address the problem. But sometimes people tend to forget that Zen-Vega APU works on much faster DDR4 memory. Vega also implement advanced memory compression tech that will help in limited bandwidth environment. That's why I think majority of RR APU won't get HBM and only limited to certain SKU (like Intel Broadwell-Iris Pro SKU).

That was simply due to the development of an entirely new technology and standard. Of those only packaging and manufacturing companies are needed now. Like amkor and samsung, both of which AMD would work either way.

 

[The Stilt]

Infinity Fabric will allow CPU use HBCC in Vega GPU, and if the SoC will have HBM2 treat it as "system RAM" with appropriate bandwidth.

So basically you can get 200 GB/s instead of 50 GB/s with this "feature".

Only beneficial is AMD was to create a MCM chip with a server CPU core (such as Zeppelin) in it.
Infinity Fabric refers either to DFI, GMI or xGMI and much like the 15h APUs / SoCs which lacked HTT, Raven doesn't support GMI or xGMI (Global Memory Interface & Xternal Global Memory Interface).

 

[ao_ika_red]

That was simply due to the development of an entirely new technology and standard. Of those only packaging and manufacturing companies are needed now. Like amkor and samsung, both of which AMD would work either way.

Excuse me for being off topic.
I'm just wondering. We all know that SK Hynix has exclusive contract to deliver first HBM to AMD. But as the time goes by, Hynix lagged behind Samsung in HBM development war and eventually nVidia got first HBM2 delivered by Samsung on their GP100 chips. How deep is Hynix-AMD contract so AMD stubbornly waited for Hynix's HBM2 so their Vega product took a long time to hit consumer market?
And, which HBM version will be used in the future APU? Because, I believe HBM2 is in limited availability and HBM1 is not as energy efficient as HBM2 and has limited configuration if we compared to HBM2 (this is my pure speculation, there's no data to back it up)

 

[CatMerc]

Excuse me for being off topic.
I'm just wondering. We all know that SK Hynix has exclusive contract to deliver first HBM to AMD. But as the time goes by, Hynix lagged behind Samsung in HBM development war and eventually nVidia got first HBM2 delivered by Samsung on their GP100 chips. How deep is Hynix-AMD contract so AMD stubbornly waited for Hynix's HBM2 so their Vega product took a long time to hit consumer market?
And, which HBM version will be used in the future APU? Because, I believe HBM2 is in limited availability and HBM1 is not as energy efficient as HBM2 and has limited configuration if we compared to HBM2 (this is my pure speculation, there's no data to back it up)

AMD isn't limited to SK Hynix, they just get priority from them due to co-developing the standard. Current Vega FE is most likely using Samsung stacks if we're correct.

HBM1 will likely not see any use aside from Fiji. Due to various factors including economics of scale, HBM2 is/will be cheaper than HBM1.

 

[ao_ika_red]

AMD isn't limited to SK Hynix, they just get priority from them due to co-developing the standard. Current Vega FE is most likely using Samsung stacks if we're correct.

Thanks, I don't believe how messed up Hynix is, even AMD turn into Samsung for HBM2 supply.

HBM1 will likely not see any use aside from Fiji. Due to various factors including economics of scale, HBM2 is/will be cheaper than HBM1.

So, Fiji and HBM1 was purely technical exercise and HBM1 itself has very short lifetime (like GDDR4).

 

[PeterScott]

Manufacturing GPU with HBM is a collaboration work.

I quote it from 3DInCities, “This project required collaboration with 20 different companies and government organizations before delivering the final product,” said Black. “We worked with many partners along the way. Some disappeared, some delivered. Some we’re still working with.” The final product integrates graphics chips from TSMC, OSAT partners ASE and Amkor; HBM was procured from SK Hynix, and UMC provided the interposer.

Because there's more company involved in this project, there's also more cost and failure rate, inevitably. That's why Fiji and Vega came very late to market.
HBM memory controller also behave differently than DDR because it works in much wider bus and lower latency. That's why I found it interesting because if somehow they work it out (hybrid HBM-DDR controller, or some sort of) the cost of non-HBM and HBM APU will be lowered because they will get diverse products from one die.
Lots of people are crying for HBM APU because in mobile space, there's bandwidth starvation in high end (8 CU) APU and unlike desktop APU, mobile users have limited option to address the problem. But sometimes people tend to forget that Zen-Vega APU works on much faster DDR4 memory. Vega also implement advanced memory compression tech that will help in limited bandwidth environment. That's why I think majority of RR APU won't get HBM and only limited to certain SKU (like Intel Broadwell-Iris Pro SKU).

Those difficulties only strengthen the argument that AMD won't do HBM for such a mainstream part, with low end APU, as Raven Ridge.

And a niche offshoot part with only 11 CUs doesn't make sense for HBM either.

IMO if they are going to do an APU with HBM it will have a lot more GPU power than 11 CUs, as it is just wasting expensive HBM bandwidth.

Maybe Sony will contract something like this for PS5. Zen Cores, with 48 CUs and HBM memory.

But making RR more expensive and later to market, so they can sell some small portion into a tiny niche makes no sense. That niche is already better served by using dGPUs. Even NVidias lowest end mobile dGPU (1050) would destroy RR APU with HBM. A 1050 card beats RX 460 which is a 16 CU part.

Bottom line. Pairing HBM with an 11 CU APU is just pointless and expensive overkill. It won't happen.

 

[CatMerc]

So, Fiji and HBM1 was purely technical exercise and HBM1 itself has very short lifetime (like GDDR4).

Pretty much. HBM2 is the first one anyone expected to really take off.

 

[ao_ika_red]

Those difficulties only strengthen the argument that AMD won't do HBM for such a mainstream part, with low end APU, as Raven Ridge.

And a niche offshoot part with only 11 CUs doesn't make sense for HBM either.

Bottom line. Pairing HBM with an 11 CU APU is just pointless and expensive overkill. It won't happen.

I absolutely agree with that. This HBM-powered APU rumour I believe came from Macbook users. They're frustrated about Intel's Iris Pro solutions and hope AMD will come with better solution (HBM-powered APU) and bring the Macbook price down (yeah, good luck with that).
That's why my #852 post want to explain there's just faint hope of HBM-powered APU will happen in near future.

 

[mohit9206]

Faster than Desktop 550 is asking a lot. Firestrike standard is almost 2x faster on the 550 versus the A12-9800 (65W Bristol Ridge)

Then what good is it for? If it can't even be faster than 550 then it can't play any games other than those low spec popular games.
And as for htpc why not stick with pentium. Zen apu doesn't seem to fit anywhere like the previous apus from amd.

 

[moinmoin]

An RX 550 has a TDP of 50 watts, even for a desktop APU that's a big much.

 

[PeterScott]

Then what good is it for? If it can't even be faster than 550 then it can't play any games other than those low spec popular games.
And as for htpc why not stick with pentium. Zen apu doesn't seem to fit anywhere like the previous apus from amd.

Huge numbers of people play games on their Intel iGPs, and RR will be better for that. Not everyone is a card carrying member of the PC MR, running the latest AAA games at ultra settings on 4k monitors.

 

[ao_ika_red]

Then what good is it for? If it can't even be faster than 550 then it can't play any games other than those low spec popular games.
And as for htpc why not stick with pentium. Zen apu doesn't seem to fit anywhere like the previous apus from amd.

Lots of people still use HD monitor. By using Zen APU, they will able to play AAA games with better settings and higher fps. Recent 7890k apu only manage to ~30fps in gta v with medium settings. Future apu should deliver at least 40% better than that.
It will also useful for people who need gpu leverage for low to medium workload in graphic design / CAD.

 

[mohit9206]

An RX 550 has a TDP of 50 watts, even for a desktop APU that's a big much.

That's amds problem. Gt 1030 is faster at 30W tdp.

Huge numbers of people play games on their Intel iGPs, and RR will be better for that. Not everyone is a card carrying member of the PC MR, running the latest AAA games at ultra settings on 4k monitors.

That's what i said. It will be useless other than for low spec popular games.

Lots of people still use HD monitor. By using Zen APU, they will able to play AAA games with better settings and higher fps. Recent 7890k apu only manage to ~30fps in gta v with medium settings. Future apu should deliver at least 40% better than that.
It will also useful for people who need gpu leverage for low to medium workload in graphic design / CAD.

No, anyone who wants to play aaa games, will always buy a dedicated card unless they have no clue about pc market. $160 zen apu much worse than Pentium and 1050.

 

[PeterScott]

That's what i said. It will be useless other than for low spec popular games.

What is wrong with popular? This is AMDs CPU/APU for the masses. It may be their most important part, as the majority of the PC market only uses Intel iGPs, NOT discrete GPUs.

The majority of the PC market is a target worth chasing.

 

[Mopetar]

That's amds problem. Gt 1030 is faster at 30W tdp.

That's what i said. It will be useless other than for low spec popular games.

No, anyone who wants to play aaa games, will always buy a dedicated card unless they have no clue about pc market. $160 zen apu much worse than Pentium and 1050.

If you just want a cheap gaming PC, sure, but if you want a laptop that is used for gaming among other general tasks, a Zen APU is going to be the best bet. The top model will probably offer performance comperable to a 460 which is going to be enough for people playing e-sport titles on their laptops.

 

[iwulff]

An RX 550 has a TDP of 50 watts, even for a desktop APU that's a big much.

Actually that should fit the bill on an APU. An A10-7870K is between a R7 240/250. Which is between 30-65 TDP.

I think some people here are underestimating low GPU's/APU's. Sure you can't play most of the latest games on 1080p/60fps, but 30fps with moderate settings is sufficient for most games and the people playing them. Which games are supposed to be unplayable on a A12-980/A10-7870?

 

[iwulff]

That's amds problem. Gt 1030 is faster at 30W tdp.

That's what i said. It will be useless other than for low spec popular games.

No, anyone who wants to play aaa games, will always buy a dedicated card unless they have no clue about pc market. $160 zen apu much worse than Pentium and 1050.

Well I am one of them. I have a small case (In Win Chopin) where no dGPU fits in. I know from a functional aspect a dGPU would have been better, but I like this setup and the idea of an APU. I'm also a casual player and am satisfied with lowering settings for games to be playable.

 

[insertcarehere]

If you just want a cheap gaming PC, sure, but if you want a laptop that is used for gaming among other general tasks, a Zen APU is going to be the best bet. The top model will probably offer performance comperable to a 460 which is going to be enough for people playing e-sport titles on their laptops.

Uhh, so what you're arguing here is that the top laptop zen APU will have performance comparable to a RX 460? That would put it almost on par with the 1050 notebooks, but would also be way, WAY faster than any of the mobile Brisol Ridge stuff ever acheived, like, closer to a 400% difference than a 40% difference.

 

[Mopetar]

Uhh, so what you're arguing here is that the top laptop zen APU will have performance comparable to a RX 460? That would put it almost on par with the 1050 notebooks, but would also be way, WAY faster than any of the mobile Brisol Ridge stuff ever acheived, like, closer to a 400% difference than a 40% difference.

Supposedly the top APUs are supposed to have 11 CUs which would mean 702 (11 * 64) shaders or 78.5% of the shaders in a 460. Those typically run at around 1100 MHz (1200 boost) and we know that Vega clocks higher so between that and any IPC gains, the should be approaching 460-levels. Give it a max clock of 1300 MHz and a 7% IPC gain and an 11 CU APU is about 90% of a 460.

Obviously if performance is memory bound then the APU can suffer, but I suspect that the top APU is probably meant to replace small Polaris. If this weren't the case then AMD would only be building an APU with 8 CUs max.

 

[moinmoin]

Raven Ridge will likely have the most powerful iGPU yet, but still not beat current bottom level dedicated GPUs. 7LP (Zen 2) APUs may get there (though dedicated GPUs will make that switch as well).

 

[IntelUser2000]

Supposedly the top APUs are supposed to have 11 CUs which would mean 702 (11 * 64) shaders or 78.5% of the shaders in a 460. Those typically run at around 1100 MHz (1200 boost) and we know that Vega clocks higher so between that and any IPC gains, the should be approaching 460-levels. Give it a max clock of 1300 MHz and a 7% IPC gain and an 11 CU APU is about 90% of a 460.

Obviously if performance is memory bound then the APU can suffer, but I suspect that the top APU is probably meant to replace small Polaris. If this weren't the case then AMD would only be building an APU with 8 CUs max.

The biggest reason that it won't reach RX 460 performance is because AMD themselves said 40% improvement in GPU performance. Since its a marketing presentation it'll likely be an "upto" figure.

Earlier we were discussing about whether it'll even reach RX 550 performance, which is quite a bit slower than the RX 460. RX 550, nevermind the RX 460 is vastly faster than current AMD iGPU. Nevertheless the limiter for iGPUs are cost and TDP.

To be an effective replacement to the RX 550, it'll have to make that up in revenue. An RX 550 level of performance in an APU would mean significantly added costs to make up for the fact the discrete GPU on the level of RX 550 is obsolete. The number of ROPs in the current Carrizo APU is also half of RX 550(or RX 460 for that matter), meaning the differences are greater than number of SPs suggest.

Nevermind the only way to make up the vast difference in memory bandwidth is by having HBM of some sort. Since we're not seeing it even on $400 GPUs, what are the chances of a bargain basement APU design having one? While we wish for such solutions it's quite clear that if its not better than discrete GPUs in price and power use, it'll be snubbed again meaning it'll be pointless.

7LP (Zen 2) APUs may get there (though dedicated GPUs will make that switch as well).

20-30% gain per gen is an advancement, but to bridge the 2-3x gap it has from the bottom-rung discrete GPUs they would have to do more than that - meaning higher TDPs and costs.

I suspect that on-die HBM would have better latency than regular memory and could give AMD a significant leg up in some workloads.

On-package possible, but on-die no.

 

[PeterScott]

Supposedly the top APUs are supposed to have 11 CUs which would mean 702 (11 * 64) shaders or 78.5% of the shaders in a 460. Those typically run at around 1100 MHz (1200 boost) and we know that Vega clocks higher so between that and any IPC gains, the should be approaching 460-levels. Give it a max clock of 1300 MHz and a 7% IPC gain and an 11 CU APU is about 90% of a 460.

Vega clocks higher when chowing down on 300 watts. This is a laptop part, and so far the only leaks show the GPU clock at 800 MHz (72.7% of RX 460). 78.5% * 72.7% = ~57% of RX 460.

Obviously if performance is memory bound then the APU can suffer, but I suspect that the top APU is probably meant to replace small Polaris. If this weren't the case then AMD would only be building an APU with 8 CUs max.

You think if they can't beat low end discrete GPUs they would just give up? That makes no sense.

APUs are evolutionary, they get better as process, memory bandwidth allows. AMD added enough CUs for the available power/memory bandwidth envelope. APUs will keep getting better, abut so will discrete GPUs.

It is really better to think about this the other way around. Why would AMD bother building a low end GPU like the RX 550 if their soon to be released APU would beat it. That sets the floor for performance, not the 460/560.

I expect even the RX 550 will beat the RR APU for gaming.

 

[IntelUser2000]

I expect even the RX 550 will beat the APU for gaming.

This will likely be true for the next few years.

This is a laptop part, and so far the only leaks show the GPU clock at 800 MHz (72.7% of RX 460). 78.5% * 72.7% = ~57% of RX 460.

Nevermind the fact that 7700HQ paired up with RX 550 beats R7 graphics in Carrizo by 4-5x. If we normalize for GPU we're still looking at 2-3x. Compare it to to RX 460 and we're back up to 3-4x.

1/2 the ROP
45% of memory bandwidth if it uses DDR4-3200, and it has to share it with the platform
TDP shared with CPU, meaning CPU won't be as performant