AMD Raven Ridge 'Zen APU' Thread

[Valantar]

From the cost/benefits point of view it makes little sense (maybe only in High-end Ultrabooks) to pair a RR with HBM. Costs can be higher than a discrete low-end GPU solution with no advantage in performance. It may have more sense to couple a full Ryzen CPU with a mainstream Vega and one HBM2 stack dedicated to SFF PCs, i.e. for All-in-one or Apple.

I don't quite agree there. Power savings are king in mobile, and a single 1-4GB HBM2 stack would easily suffice for a low-end to midrange (i)GPU, while consuming 1/2 the power (or less) of a 128-bit GDDR5 setup. Not to mention the RAM area savings = PCB area savings = smaller, thinner, lighter devices. Both of those are powerful incentives in mobile. Still, I don't think it'll happen this generation.

 

[Atari2600]

What would be the point in adding a low perf GPU and pairing it with HBM? What market does it serve and why does it need HBM?

It wouldn't be low perf (i.e. not the ~11 CU that RR will now ship with). More like 20+ CU.


It'd be for use (1) in areas which benefit from limited acceleration across GPUs, so presents a more balanced solution that dGPUs and (2) use as a high end APU for consumers/prosumers.


HBM is obviously needed to feed instance (2), and in terms of (1), having a local on-die area that can act as a... I'll not say cache as the pure latency isn't there for that... but more a second storage using separate memory "channels" can lift the bandwidth load off the main memory access - and that would apply to both x86 and GPU.

To MCM a mobile Vega 10 (interposer only for the GPU+HBM) and 8-16 CPU cores might make some sense.

Erm... just had a thought.

With the way infinity fabric is... isn't that virtually the same as a CCX+GPU? Would the HBCC in Vega allow for seamless integration into IF I wonder.

Hmmmm....

 

[moinmoin]

With the way infinity fabric is... isn't that virtually the same as a CCX+GPU? Would the HBCC in Vega allow for seamless integration into IF I wonder.

I thought that's exactly what the long rant about infinity fabric at the AMD Financial Analyst Day was about, IF being usable for mixing heterogeneous modules.

 

[leoneazzurro]

I don't quite agree there. Power savings are king in mobile, and a single 1-4GB HBM2 stack would easily suffice for a low-end to midrange (i)GPU, while consuming 1/2 the power (or less) of a 128-bit GDDR5 setup. Not to mention the RAM area savings = PCB area savings = smaller, thinner, lighter devices. Both of those are powerful incentives in mobile. Still, I don't think it'll happen this generation.

Yes, but if you read I said it would make sense for high end ultrabooks (where the user pays a price premium for those power savings) but I think AMD is not ready yet for that segment, as the brand must recover a lot due to past years... For mass (i.e. cheap) notebooks a DDR4 RR is more than enough probably to leave in the dust most of Intel APU offerings from the graphical point of view.

 

[imported_jjj]

It wouldn't be low perf (i.e. not the ~11 CU that RR will now ship with). More like 20+ CU.

It'd be for use (1) in areas which benefit from limited acceleration across GPUs, so presents a more balanced solution that dGPUs and (2) use as a high end APU for consumers/prosumers.

HBM is obviously needed to feed instance (2), and in terms of (1), having a local on-die area that can act as a... I'll not say cache as the pure latency isn't there for that... but more a second storage using separate memory "channels" can lift the bandwidth load off the main memory access - and that would apply to both x86 and GPU.

Erm... just had a thought.

With the way infinity fabric is... isn't that virtually the same as a CCX+GPU? Would the HBCC in Vega allow for seamless integration into IF I wonder.

Hmmmm....

I assumed it's a given that AMD would not invest in another die.
They have very limited resources, they couldn't even do a native quad core Summit Ridge or a very low power mobile SoC this year - RR is said to go down to 12W. Vega 11 is likely for next year as they just did the Polaris refresh.
So for a MCM APU, it has to be an existing die. Vega 10 or RR with a slight chance that some Polaris die is equipped to handle this.
AMD had 1 billion for R&D last year ,they can afford a bit more this year but not much. They are not gonna invest much in a small niche.

As for HBCC. i believe they need to achieve near perfect scaling with GPU chiplets this decade. Likely with Navi but we'll see how close Vega is. I am looking forward to the rumored dual Vega.
For the APU, no idea how close to ideal it could get but it would be better than discrete- aside from the TDP limitation.

 

[raghu78]

Looks like no HBM2 for Raven Ridge, just as I predicted oh so long ago...

AMD demoed a 15W Raven Ridge (RR) BGA chip (FP5) for ultrathin notebooks. HBM2 does not make any sense for that segment as the GPU is clocked at very low speeds.

https://videocardz.com/69428/amd-snowy-owl-naples-starship-grey-hawk-river-hawk-great-horned-owl

Great Horned Owl , the embedded equivalent version of RR is expected to have 512 cores for 12-15W SKUs. I expect the same for RR. My guess is the Vega GPU in 12-15w RR SKUs clocks around 850-1000 Mhz. For the same core count as Bristol Ridge, RR will have significantly higher DDR4 memory speeds and vastly improved color compression and a next gen pixel engine with tiled rasterizer which should be much more efficient in bandwidth utilization. Vega sports a new cache architecture and it would not surprise me to see RR sport more on chip GPU L2 cache than Bristol Ridge to further reduce bandwidth requirements.

RR for desktops is a completely different case altogether. The TDP is 65W and the GPU will have 4 times (or more) the power budget of the 12-15W RR SKUs. The RR desktop SKU has 704 sp which could very well clock above 1500 - 1600 Mhz (given that Vega is designed for those clock speeds). You are talking of a 2 - 2.2 TF GPU at those speeds. HBM2 should allow RR desktop to perform to its full potential . HBM2 makes a lot of sense for RR desktop and we would not know until Q1 2018 if the AM4 desktop version has HBM2 or not. If there is a Kabylake G with AMD GPU and HBM2 I would bet that we will see AMD Raven Ridge with HBM2.

 

[imported_jjj]

HBM2 makes a lot of sense for RR desktop and we would not know until Q1 2018 if the AM4 desktop version has HBM2 or not. If there is a Kabylake G with AMD GPU and HBM2 I would bet that we will see AMD Raven Ridge with HBM2.

Lets see why not so much:
- costs more than double even with just 2GB HBM and nobody needs that APU at 500$
- for a 65W TPD APU , the TDP added by the HBM kills any benefit
- AMD is not insane, not that insane anyway

For HBM to be viable it needs a high ASP product and to be the best solution for that product.
Even the console guys can't afford HBM and those are pretty big chips that go into a system so the cost of HBM could be masked.
When AMD has a much cheaper HBM solution, they can use it in some more products, if it makes sense and desktop APU is the very last product that needs HBM as the main benefit of HBM is lower power vs GDDR. Desktop isn't using GDDR and is much less TDP constrained than laptop or server.

 

[Topweasel]

Lets see why not so much:
- costs more than double even with just 2GB HBM and nobody needs that APU at 500$
- for a 65W TPD APU , the TDP added by the HBM kills any benefit
- AMD is not insane, not that insane anyway

For HBM to be viable it needs a high ASP product and to be the best solution for that product.
Even the console guys can't afford HBM and those are pretty big chips that go into a system so the cost of HBM could be masked.
When AMD has a much cheaper HBM solution, they can use it in some more products, if it makes sense and desktop APU is the very last product that needs HBM as the main benefit of HBM is lower power vs GDDR. Desktop isn't using GDDR and is much less TDP constrained than laptop or server.

What if the the use of HBM removes the need for system memory? This would be a big plus for Ultrabooks/Surface like products. A $200 increase in price for 8GB of HBM might be worth not having include RAM since that would be sourced by AMD. There is a reason products such as the Iris lineup exist.

 

[mtcn77]

What if the the use of HBM removes the need for system memory? This would be a big plus for Ultrabooks/Surface like products. A $200 increase in price for 8GB of HBM might be worth not having include RAM since that would be sourced by AMD. There is a reason products such as the Iris lineup exist.

If only Iris lineup was used to accelerate L4 memory...

 

[otinane]

So is it true? RR for desktops will come in 2018 and OEM only like A12-9800?

 

[nix_zero]

So is it true? RR for desktops will come in 2018 and OEM only like A12-9800?

if you were amd chief of marketing what market you would exploit first, premium mobile where amd has zero penetration and its still growing and being more actractive every day or stagnationg desktop low end apu one?

 

[clocks11]

So is it true? RR for desktops will come in 2018 and OEM only like A12-9800?

Gosh I hope not. I need a new system this year, and want Ravin Ridge. I guess I may have go to plan B.

 

[DisEnchantment]

Gosh I hope not. I need a new system this year, and want Ravin Ridge. I guess I may have go to plan B.

I think Raven Ridge for AIO desktops could be very interesting, probably slim too.

 

[Arachnotronic]

AMD demoed a 15W Raven Ridge (RR) BGA chip (FP5) for ultrathin notebooks. HBM2 does not make any sense for that segment as the GPU is clocked at very low speeds.

https://videocardz.com/69428/amd-snowy-owl-naples-starship-grey-hawk-river-hawk-great-horned-owl

Great Horned Owl , the embedded equivalent version of RR is expected to have 512 cores for 12-15W SKUs. I expect the same for RR. My guess is the Vega GPU in 12-15w RR SKUs clocks around 850-1000 Mhz. For the same core count as Bristol Ridge, RR will have significantly higher DDR4 memory speeds and vastly improved color compression and a next gen pixel engine with tiled rasterizer which should be much more efficient in bandwidth utilization. Vega sports a new cache architecture and it would not surprise me to see RR sport more on chip GPU L2 cache than Bristol Ridge to further reduce bandwidth requirements.

RR for desktops is a completely different case altogether. The TDP is 65W and the GPU will have 4 times (or more) the power budget of the 12-15W RR SKUs. The RR desktop SKU has 704 sp which could very well clock above 1500 - 1600 Mhz (given that Vega is designed for those clock speeds). You are talking of a 2 - 2.2 TF GPU at those speeds. HBM2 should allow RR desktop to perform to its full potential . HBM2 makes a lot of sense for RR desktop and we would not know until Q1 2018 if the AM4 desktop version has HBM2 or not. If there is a Kabylake G with AMD GPU and HBM2 I would bet that we will see AMD Raven Ridge with HBM2.

yeah it won't have HBM2, mark this post.

 

[Bouowmx]

AMD Raven Ridge for PGA (socket AM4) in 2018 (images of the same road map in this thread have been removed, pls no h8 for me for using site) http://wccftech.com/amd-pinnacle-ridge-raven-ridge-processors-confirmed/

No confirmation of OEM-only availability, but kind of a waste to put video outputs on most motherboards for DIY.

 

[Topweasel]

AMD Raven Ridge for PGA (socket AM4) in 2018 (images of the same road map in this thread have been removed, pls no h8 for me for using site) http://wccftech.com/amd-pinnacle-ridge-raven-ridge-processors-confirmed/

No confirmation of OEM-only availability, but kind of a waste to put video outputs on most motherboards for DIY.

Well let's keep in mind that AMD is really trying to press ZEN into as many markets as quickly as possible and we don't know what production allotments are going to be like. They did the VGA support because AM4 is supposed to support their lineup for a while going forward. So even if they don't offer it right away eventually retail AM4 APU's will hit the street.

What's more likely is that the RR AM4 launch will be restricted to OEMs for a bit as the try to push out more OEM design wins and as production ramps up offer retail desktop components. Ryzen's important market is enthusiasts and it helps get word of mouth out. The APU's strong point is going to be in OEM sales regardless of Retail availability.

 

[imported_jjj]

What if the the use of HBM removes the need for system memory? This would be a big plus for Ultrabooks/Surface like products. A $200 increase in price for 8GB of HBM might be worth not having include RAM since that would be sourced by AMD. There is a reason products such as the Iris lineup exist.

That's even worse.
From a cost perspective HBM has a premium over DDR , the more you add , the worse it gets (then there is the interposer, yileds and other costs).
More importantly you add way too much heat to the package for an ultrabook , drastically decreasing the TDP available for the CPUGPU so end up with a much much slower solution at a higher cost. You don't need more BW to begin with at such a low TDP so the idea is ridiculous from the start.
Intel's Iris is an entirely different solution , not even close to being comparable in costs and TDP - not that it's a good solution or one that is competitive.Intel did it because there was no competition

A large discrete GPU with HBP ,maybe even paired with a SoC in a single package ( the SoC not on the interposer) does make sense from a form factor perspective but at much higher TDP and price. Achieves a similar goal to Nvidia's downclocked garbage that was just announced but does it better. The problem with that is the low volumes and they would need to use already existing products and just pack them together.

A much cheaper HBM with no interposer or a much cheaper interposer would enable for HBM to be used in more budget friendly products - starting with discrete GPUs that are not 500$+ but those kind of solution are just emerging.

 

[imported_jjj]

Erm... just had a thought.

With the way infinity fabric is... isn't that virtually the same as a CCX+GPU? Would the HBCC in Vega allow for seamless integration into IF I wonder.

Hmmmm....

Remembered that Raja Koduri said something on the topic during the recent Reddit AMA

"We haven't mentioned any multi GPU designs on a single ASIC like Epyc, but the capability is possible with Infinity Fabric."

 

[raghu78]

yeah it won't have HBM2, mark this post.

So you expect to see Intel Kabylake G with HBM2 but not Raven Ridge with HBM2. We will see. AMD is not going to let themselves be beaten by Intel in gaming. Thats my belief.

 

[jpiniero]

So you expect to see Intel Kabylake G with HBM2 but not Raven Ridge with HBM2. We will see. AMD is not going to let themselves be beaten by Intel in gaming. Thats my belief.

EMIB saves a lot of the cost. Plus, it's really a discrete GPU in all but name only.

 

[raghu78]

EMIB saves a lot of the cost. Plus, it's really a discrete GPU in all but name only.

I don't see why a 2GB HBM2 cache is not possible with RR . Providing 128 GB/s bandwidth would allow the integrated Vega GPU to perform to full potential. AMD could command the price of a USD 100 GPU + a USD 200 CPU easily when it would match or beat the discrete GPU while consuming lower power and providing huge form factor benefits. 704 sp Vega with higher IPC and higher clocks than Polaris would likely match or even exceed a full RX 560 / GTX 1050 in performance. A 210 sq mm chip at GF 14LPP is not going to be more than USD 20-25 (the manufacturing cost of Ryzen). We have seen Polaris 10 chips with a 232 sq mm die and well built AIB partner boards with 8GB of GDDR5 sell for USD 230 which includes AIB margins that AMD won't have to account for when selling its own APUs. A single 2 Hi HBM2 stack should be possible to integrate given that AMD can definitely command a price of USD 300 - USD 325 for such a powerful APU. Such an APU would be the perfect chip for AIO PCs, HTPC and small form factor desktops.

 

[mtcn77]

Samsung had an announcement that they didn't need the interposer layer for value oriented chips. They can easily couple that with 2GB single-stack HBM pile. 4GB if HBM2.

 

[Valantar]

I don't see why a 2GB HBM2 cache is not possible with RR .

Well, a HBM2 controller and interface on the chip would certainly help. We have no indication of this existing, at least not that has been mentioned by a credible source.

Samsung had an announcement that they didn't need the interposer layer for value oriented chips. They can easily couple that with 2GB single-stack HBM pile. 4GB if HBM2.

Are you talking about Samsung's "low cost HBM" initiative? I never got the sense that they were skipping the interposer, just that they were working to lower costs. Are you proposing that they stack the HBM on top of the CPU? That would be a horrible idea, at least.

 

[raghu78]

Well, a HBM2 controller and interface on the chip would certainly help. We have no indication of this existing, at least not that has been mentioned by a credible source.

We do not yet have a die shot of Raven Ridge. So we can still speculate until we see a die shot. If AMD has a designed a 2TF APU (easily better than PS4 given Vega's next gen GCN architecture) and paired it with dual channel DDR4 3200 the chief designer of the RR APU should be fired for wasting precious die space (as the GPU will not be able to perform to its potential) and increasing power / cost and reducing yields.

 

[jpiniero]

AMD can definitely command a price of USD 300 - USD 325 for such a powerful APU.

No they can't. That's the issue. Even Kaby-G is probably an Apple semi-exclusive.

Are you proposing that they stack the HBM on top of the CPU?

That's what I am predicting AMD is doing with Navi.