AMD Raven Ridge 'Zen APU' Thread

[msroadkill612]

"RavenRidge looks to be exactly what an AMD APU was supposed to be in the first place, cheap yet powerful CPU+iGPU and now at low wattage as well.

With 40% (iGPU) higher perf than BristolRidge at 50% less power, it could easily become the best APU in 2017 for Laptops."

Yep.

Random points:

Its not just a matter of amd amd having the best apu - they have the ONLY worthy apu.

Intel dont have an inhouse gpu to partner their cpuS, even if they had the expertise.

Any decent level of graphics using a discrete nvidia gpu has too many innate power and other disadvantages.

There are very compelling arguments for APUs on the desktop and servers also, but very compelling indeed in mobile.

AMD are old hands at apuS, and they worked well for general use, using mere system memory & a dull cpu, for ~all bar gaming.

Given amd now have; advances of 14nm zen & vega, faster ddr4 ram, fabric, hbcc & its exciting memory management, fabric linked resources,...

RR will be a hell of an advance in apuS, no matter what.

Missing and/or noteworthy imo, are wild cards that could add large additional impact beyond the 50%~better than prev apu amd claims.

One feature is that vega can have its own local nvme drives on its Fabric:

I am more optimistic about bandwidth, but the jist is:

"
two 1TB Samsung 961 M.2 SSD could potentially deliver as much as 6.4 GB/s read and 3.8 GB/s write for 2TB of capacity. Physical limit of the current configuration is probably 8GB/s read and 8GB/s write."

https://vrworld.com/2016/08/05/amd-radeon-pro-ssg-support-4tb-memory/

It can use it like a giant l3 cache for HBM2, and system memory as l2 cache. It could also provide big power and heat savings over dram, & it seems likely these resources would be shared by the cpu via fabric.

I think that may be a big theme with amd - use our HBCC tech, and you need far less costly, scarce & power thirsty dram

The sum is greater than the parts. Its hard not to underestimate fabric, but its full of potential.

all zen products are multiples (double ups) of the same Zeppelin, 2x zen ccx die. ~Straightforward.

1 x zen ccx and 1 x vega gpu on a single die, is very different die, surely?

Having covered that irresistible market, imo, they will soon move to a dual gpu die, similar to the zeppelin cpu die, and continue with the same formula as they have w/ ryzen/TR/epyc/2 socket epyc.

They will simply mix and match cpu & gpu dies on the MCMs, to suit clients bias toward cpu vs gpu processing power.

 

[msroadkill612]

"if this is correct than we can expect pretty nice boost in performance. 40% bigger GPU, but with 30% lower core clock, is still able to be 40% faster on average? That is pretty nice boost.

Im wondering, what they will be able to achieve with higher TDPs, and faster RAM."

Then there is/are:

doable single node liquid cooling on the combined cpu/apu

some hbm cache?

ryzen speed interconnects for cpu-gpu-~memory traffic

hbcc should do a much better job of using system ram as gpu cache.

They may get by with surprisingly few of the 32 lanes on offer, being used for the GPU interconnect, by using better Fabric protocol rather than pcie3 eg., leaving it considerable io elbow room.

 

[msroadkill612]

Would I *like* to see HBM2 on APU's? Yep, sign me up. And it will happen. But not this year. Probably not next year either. Mark Papermaster has said in interviews that the cost isn't there yet. And I gotta believe the CTO of AMD over wishes and wants of people, well people like me. As a wild guess, I'd expect HBM2 and Vega cores to appear first on the server bits. OpenCL, GPGPU etc. Video cards as well, as they have a bit more room in the budget. *After* you can get some volume flowing, and get the cost down you might have room to get HBM2 onto consumer facing APU's.

I think the reason amd went out on a limb a bit w/ hbm2 was its compactness fitted its plans so well w/ ~on die cache for apu gpuS.

the ram market has been unkind, esp. hbm2, so amd has had to back off a bit for now, to meet price points

 

[iwulff]

Raja did mention Vega has Infinity Fabric tech builtin. I'm unsure of the use case, but possibly to easily create a Zen+Vega APU. This might lead to a higher % of functional dies which they can sell. I am curious of how the final die will look like. Infinity Fabric is all about efficiently linking various parts but also more usable dies.

They might also just create a dual Vega GPU which acts as one GPU, but seeing the power requirements this has to be a lower voltage part most likely.

 

[Topweasel]

Raja did mention Vega has Infinity Fabric tech builtin. I'm unsure of the use case, but possibly to easily create a Zen+Vega APU. This might lead to a higher % of functional dies which they can sell. I am curious of how the final die will look like. Infinity Fabric is all about efficiently linking various parts but also more usable dies.

They might also just create a dual Vega GPU which acts as one GPU, but seeing the power requirements this has to be a lower voltage part most likely.

Vega including IF is going to be more about APU's and HBCC than anything else at Vega's power usage. Just don't see a situation for a dual die setup.

 

[ao_ika_red]

I agree. One thing to consider is the single CCX design, so if anything RR should perform a tiny bit better then the split-CCX Ryzen 3/5.

With regards to Vega, we'll see. It might be a half-dud for the high-end*, but it might be more effective in an APU design. We saw something similar with VLIW4, power hungry at the top, but pretty effective and efficient in APU form.

*If you call the rumoured 1080-class performance "bad"...

As a user of piss-poor 9800GTX, I can confirm that calling 1080-class performance bad is rather offensive. I read somewhere that Vega FE has similar IPC with Fiji but with better memory compression. That alone will be helpful to APU users. I know RR will have HBM tech on it, but I don't think it will be cheap. Thanks to DDR4 advancement and latest AGESA update, I better off invest the fastest DRAM I can get.

Raja did mention Vega has Infinity Fabric tech builtin. I'm unsure of the use case, but possibly to easily create a Zen+Vega APU. This might lead to a higher % of functional dies which they can sell. I am curious of how the final die will look like. Infinity Fabric is all about efficiently linking various parts but also more usable dies.

They might also just create a dual Vega GPU which acts as one GPU, but seeing the power requirements this has to be a lower voltage part most likely.

By enabling IF in Zen and Vega, we'll get better performance by using the fastest compatible DDR4 due to IF nature. If their initial performance target uses DDR4-2400 kit as a yardstick, we'll get improved performance even by using DDR4-2666 kit.

As I stated before, IF is AMD's attempt in advancing their HSA concept. It won't have significant performance gain in gaming space, but expect enormous jump in openCL-based application.

 

[iwulff]

What do you expect for the iGPU performance for Raven Ridge on desktop? I hope for a bit more then Polaris RX 550 performance, but that is entirely dependent on Vega performance and how good it's compression is to save bandwidth.

 

[insertcarehere]

What do you expect for the iGPU performance for Raven Ridge on desktop? I hope for a bit more then Polaris RX 550 performance, but that is entirely dependent on Vega performance and how good it's compression is to save bandwidth.

I am thinking AMD needs to aim for ~GTX1050 notebook performance with their top-end desktop APUs, that card is already starting to falter with newer games at 1080p for maintaining > 30fps, it'd be a hard sell as an entry level gaming all-in-one solution for much below that performance.

 

[Blitzvogel]

I am thinking AMD needs to aim for ~GTX1050 notebook performance with their top-end desktop APUs, that card is already starting to falter with newer games at 1080p for maintaining > 30fps, it'd be a hard sell as an entry level gaming all-in-one solution for much below that performance.

I agree that the RX 550 is already faltering at 1080p, though you could argue that what we expect isn't what AMD was intending (e-sports titles).

 

[jpiniero]

What do you expect for the iGPU performance for Raven Ridge on desktop? I hope for a bit more then Polaris RX 550 performance, but that is entirely dependent on Vega performance and how good it's compression is to save bandwidth.

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

 

[Bouowmx]

For biggest Vega APU, 704 cores, it needs 1357 MHz (close to Vega Frontier Edition's "typical clock") to match NVIDIA GeForce GTX 1050 at minimum boost (not counting GPU Boost). Now reduce ROPs (Vega APU may have 16?), add mobile power restrictions, and low memory bandwidth.

Practically, I still see separating CPU and GPU, to avoid memory bandwidth contention, continuing to be the accepted paradigm.

 

[IntelUser2000]

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

Yea. Notebookcheck's comparisons show 7700HQ with RX 550 Laptop version performs 3-4x A12-9800B graphics. If we assume a generous 50% gain due to the CPU we're still talking about 2-3x.

RX 550 has 7GT GDDR5 with 128-bit width. That's 112GB/s of bandwidth. You'd need a 128GB bandwidth HBM to have any chance of getting that performance. Shared setups are slower so 128GB HBM with exact GPU as the RX 550 will end up with the RX 550 being still faster.

And since they said a generic "40%" gain that could mean in the best case scenario with certain chips. Desktops might end up less at 20%.

 

[PeterScott]

Practically, I still see separating CPU and GPU, to avoid memory bandwidth contention, continuing to be the accepted paradigm.

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

I agree with you both. This is not going to challenge even low end GPU cards, but slot in under them. This is the way the market will remain. Low end GPUs will be slightly above APUs or they wouldn't even get built, so low end GPUs will always be better than APUs.

But it will be a very decent APU for mainstream laptops and lightweight desktops, which is really the biggest market there is.

This is an extremely important part for AMD.

 

[NewFatMike]

You'll probably be able to get 900p medium 30FPS+, which gets you more or less console quality gameplay.

Throw in a Freesync 2 screen on a laptop or what have you and I think a lot of folks will be very happy.

E: especially if tile based rasterization finally starts working.

Sent from my Pixel using Tapatalk

 

[ao_ika_red]

What do you expect for the iGPU performance for Raven Ridge on desktop? I hope for a bit more then Polaris RX 550 performance, but that is entirely dependent on Vega performance and how good it's compression is to save bandwidth.

If we learn from APU history, it won't be faster than RX550. It only be equal with RX550 at its best. But exception for APU+HBM SKU. It should be faster (than a RX550), but it also will be more expensive.

 

[PeterScott]

You'll probably be able to get 900p medium 30FPS+, which gets you more or less console quality gameplay.

Throw in a Freesync 2 screen on a laptop or what have you and I think a lot of folks will be very happy.

E: especially if tile based rasterization finally starts working.

On the lower end of consoles.

CU count:

Raven Ridge: 11
XB1: 12
PS4: 18
PS4P: 36
XB1X: 40

 

[NewFatMike]

They're all GCN 1.0 graphics in consoles, aren't they?

Sent from my Pixel using Tapatalk

 

[USER8000]

They're all GCN 1.0 graphics in consoles, aren't they?

Sent from my Pixel using Tapatalk

Nope - the PS4 and XBox One use GCN1.1,the PS4 PRO is Polaris with additional stuff not found in it,and the Xbox One X has some Vega features.

 

[NewFatMike]

Oh neat! Thanks, I guess I wasn't following as closely as I thought. Hopefully performance will be close on desktop form factors and we won't have to wait for Navi for it.

Sent from my Pixel using Tapatalk

 

[iwulff]

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

It is perhaps unlikely yes. Here is some hoping AMD will improve RAM compatibility, since dual ranked RAM has better performance then single ranked (~2 fps on average) for an iGPU. And high speed RAM also is important to lower the bottleneck current APU's are facing. 3200-3600 speeds might be required for good results. The Vega uArch can get higher clocks then polaris/fiji, so I hope this also reflects the iGPU. Anyway it is possibly and perhaps even likely I am too optimistic, but you can never stop dreaming. In all honesty I still hope RX Vega will surprise us with a good increase in IPC and working new features that can be of good use on an APU.

I am also curious if the CPU will downclock if the iGPU is being used, like how this is done for the current APU's.

 

[Glo.]

Raja did mention Vega has Infinity Fabric tech builtin. I'm unsure of the use case, but possibly to easily create a Zen+Vega APU. This might lead to a higher % of functional dies which they can sell. I am curious of how the final die will look like. Infinity Fabric is all about efficiently linking various parts but also more usable dies.

They might also just create a dual Vega GPU which acts as one GPU, but seeing the power requirements this has to be a lower voltage part most likely.

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".

 

[Seronei]

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".

Do we have any clue of how latency and power consumption compares between HBM2 and DDR4?

 

[PeterScott]

Do we have any clue of how latency and power consumption compares between HBM2 and DDR4?

I think there is very little chance Raven Ridge will get HBM memory. AMD doesn't have resources to spend on tiny niches.

Raven Ridge will be mainstream part with mainstream pricing.

 

[Seronei]

I think there is very little chance Raven Ridge will get HBM memory. AMD doesn't have resources to spend on tiny niches.

Raven Ridge will be mainstream part with mainstream pricing.

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.

 

[PeterScott]

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. .

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.