Speculation: AMD's 7nm processors will all be APUs

[Vattila]

Having some kind of small IGP that takes up as little die space as possible opens up a new market for the die, and is a small added bonus / convenience everywhere else.

Ideally, all AMD offerings should at least beat Intel on IGP, as integrated graphics is AMD's undisputed market-leading feature. No smaller. It is somewhat absurd that Intel now wins on this account, from a feature set perspective, in important parts of the desktop market, especially business, where the IGP-less Ryzen is up against Core i3-i7 with IGP.

PS. Curiously, a GCX with 11 Vega CUs is about the size of a CCX.

 

[Vattila]

HSA would be great if we had useful developer tools for it.

Agree. C++ AMP and ROCm (HCC) are promising.

 

[Topweasel]

PS. Curiously, a GCX with 11 Vega CUs is about the size of a CCX.

Probably not a coincidence.

 

[ksec]

Given Navi is suppose to be modular, may be they will simply have 2 - 3 CPU die size along with 2 - 3 Navi Die Size, then mix and match them with Infinity Fabric.

So yes I agree apart from the wording. There wont be APU "Die", there will be APU with CPU die and a GPU die inside the same package. And likely, some day these APU may even include HBM as well.

 

[Jan Olšan]

Given Navi is suppose to be modular

Is there actually any source for this? I haven't seen anything, it could be the case where somebody probably just started to circulate it a s a speculation and mpw nobody remembers where it is from so it is hard to label it as such.

 

[TimCh]

Is there actually any source for this? I haven't seen anything, it could be the case where somebody probably just started to circulate it a s a speculation and mpw nobody remembers where it is from so it is hard to label it as such.

Navi "scalability" is from a 2016 capsaicin event slide. Some people interpret the "scalability" as modular in reality nobody seems to know what it means.

 

[Vattila]

So yes I agree apart from the wording. There wont be APU "Die", there will be APU with CPU die and a GPU die inside the same package.

Yes, multi-die and packaging is a promising solution enabled by Infinity Fabric. I guess it comes down to cost. When 2.5D or the cheaper 2.1D packaging solutions are ready for the high volume mainstream, producing separate CPU and GPU dies ("chiplets") on their respective optimal manufacturing processes may be the way to go.

However, I guess the Zen 2 and 3 generations, which are targeted for AM4, will aim for a single die and simple packaging in the mainstream.

 

[wahdangun]

the IGP is pretty big and useless for too many applications, I think they are going to try replicating the current strategy.

No it will be so good for VDI deployment

 

[scannall]

Yes, multi-die and packaging is a promising solution enabled by Infinity Fabric. I guess it comes down to cost. When 2.5D or the cheaper 2.1D packaging solutions are ready for the high volume mainstream, producing separate CPU and GPU dies ("chiplets") on their respective optimal manufacturing processes may be the way to go.

However, I guess the Zen 2 and 3 generations, which are targeted for AM4, will aim for a single die and simple packaging in the mainstream.

Only Zen+ and Zen 2 are for certain on the AM4 platform. PCie4 and possibly DDR5 should be out about the time Zen 3 comes along.

 

[Vattila]

it will be so good for VDI deployment

Virtual Desktop Infrastructure is an excellent use-case for the hypothetical on-die GCX. With support for virtualisation, as you would presume, power-efficient GPU resources can be allocated to virtual machines. Also, the on-die multi-media accelerators can be put to good use in a data centre setting, e.g. for power-efficient transcoding of video. With 4 hypothetical GCXs per EPYC processor, i.e. 48 threads of x86 general compute plus 44 threads of GPU compute at your disposal, it is not difficult to come up with more use-cases in server, e.g. power-efficient rendering farms with excellent compute density (compared to bulky PCIe cards).

 

[Yotsugi]

Virtual Desktop Infrastructure is an excellent use-case for the hypothetical on-die GCX.

You're suggesting wasting precious die area (and thus margins) on something so niche? There's hundreds of other tasks performed in data center that give zero shits about GPUs.

(compared to bulky PCIe cards)

SMX2 wants to talk to you.

 

[Vattila]

Only Zen+ and Zen 2 are for certain on the AM4 platform.

Yes, that could be. My point though was about packaging. For mainstream products, I suspect AMD will stick to single die and simple packaging on 7nm. And Zen 2 and Zen 3 will be on 7nm.

 

[wahdangun]

You're suggesting wasting precious die area (and thus margins) on something so niche? There's hundreds of other tasks performed in data center that give zero shits about GPUs.

SMX2 wants to talk to you.

No, it will not be niche situation anymore especially with virtualize everything mentality that happening right now, my company right now in the transition to full VDI, and right now all our server have been virtualized from router to VOIP.

 

[Vattila]

My thoughts: [...] three 4-core CCX modules on a flexible die for use in TR, Epyc and standard desktop Ryzen with no GCX [plus a later] APU that is two 4-core CCX modules with a GCX.

Yes, that's the obvious continuation of the current solution and the likely alternative to my hypothetical all-in-one die — and ever more likely, if that leaked purported roadmap slide with Matisse and Picasso has root in reality. But, if so, AMD needs to get the 8-core APU out ASAP to not lose OEM designs to Intel in the mainstream desktop market, now that Intel has moved beyond 4 cores. I fear that the IGP-less Ryzen will have less success with the OEMs than in the DIY channel.

Personally, I am an HSA enthusiast, and would like to see APUs become widespread in the marketplace, so that developers start to program for HSA.

 

[beginner99]

I presume that an HSA-compliant SoC/SiP will have much better compute efficiency and density, and that this is the reason AMD have spent all those resources on developing HSA and HSA-compliant APUs.

Ryzen 3,5,7 all are identical to the server chips. So no it makes 0 sense to add a GPU as we should not forget that most of these servers will run some run of the mill Java Business apps and/or a database. Nothing that profits from a GPU (or any type of AVX).

You're suggesting wasting die area on something that is going to be useless 90% of the time?

I agree but realistically intel is doing the same thing. No one will use the igpu in a 8700k. OK, they will also go in laptops but I guess 90% of the 6-cores will end up in desktops.

 

[Yotsugi]

No, it will not be niche situation anymore especially with virtualize everything mentality that happening right now, my company right now in the transition to full VDI, and right now all our server have been virtualized from router to VOIP.

99% of the hyperscale couldn't give a single fuck about that and they are the biggest consumers of server hardware out there. Besides, margins.

No one will use the igpu in a 8700k.

But that's a consumer die and not a server one like Zeppelin.

 

[wahdangun]

99% of the hyperscale couldn't give a single fuck about that and they are the biggest consumers of server hardware out there. Besides, margins.

But that's a consumer die and not a server one like Zeppelin.

And that chip is not targeted at that market,

If VDI is niche why the hell Microsoft double down on it in their windows server line.

 

[Yotsugi]

And that chip is not targeted at that market,

The successor of Zeppelin is not targeted at hyperscale?

 

[wahdangun]

The successor of Zeppelin is not targeted at hyperscale?

, I'm not talking about zen2 EPYC, but and RR EPYC like configuration, with 704 *4 shader configuration, it will be great for VDI cluster (where you can allocating more symmetrical resources), especially if it can get professional driver.

 

[Yotsugi]

, I'm not talking about zen2 EPYC, but and RR EPYC like configuration, with 704 *4 shader configuration, it will be great for VDI cluster (where you can allocating more symmetrical resources), especially if it can get professional driver.

In what way this will be better than bog standard 2U rack with 1P EPYC + GPUs?

 

[Topweasel]

Ryzen 3,5,7 all are identical to the server chips. So no it makes 0 sense to add a GPU as we should not forget that most of these servers will run some run of the mill Java Business apps and/or a database. Nothing that profits from a GPU (or any type of AVX).



I agree but realistically intel is doing the same thing. No one will use the igpu in a 8700k. OK, they will also go in laptops but I guess 90% of the 6-cores will end up in desktops.

The 8700k. Is a rebrand of a CPU designed primarily for mobile usage but even if it's a high clocked version of it for desktop users it's still a mobile CPU. There is nothing shared between this and the server version. Specially starting with SL where the server chip has been really really retweeked.

Ryzen on the other hand is a desktop offering of a die that is used all the way up to the biggest server chips. We get Ryzen the way it is because of it's use on the server. If there was enough use for an IGP for server work it would be one thing. But AMD isn't wasting a ton of die space on a useless function just to make the desktop versions more flexible. Specially since they have a dedicate die and product line just for that.

 

[el etro]

It came to me an idea.

With 7nm cell sizes for the caches being so low, i see a chance that all 7nm Ryzens come with iGPU. The shrink promised is around 55%, so there's hope for it. Even a 10CU Navi IGP for the higher end models would be fine, just to avoid spending money for a dGPU unecessarily.

 

[LightningZ71]

Don't discount the possibility of a third die. On 7nm, its possible to have 4 CCXs on a die that's roughly the same size with roughly the same ccx L3. That could give you a HEDT/server die family that is 4 CCXs, a Desktop Die that is 3 CCXs and a GCX/iGPU, and a smaller mobile Die that is 2 CCXs and a GCX/iGPU, or the same sized die with 2 CCXs, a 32-64MB L4 cache, and a GCX/iGPU. With better revenue streams, a modular design, and an existing design that's being tweaked instead of being built from the ground up, this isn't entirely unreasonable.

 

[raghu78]

Don't discount the possibility of a third die. On 7nm, its possible to have 4 CCXs on a die that's roughly the same size with roughly the same ccx L3. That could give you a HEDT/server die family that is 4 CCXs, a Desktop Die that is 3 CCXs and a GCX/iGPU, and a smaller mobile Die that is 2 CCXs and a GCX/iGPU, or the same sized die with 2 CCXs, a 32-64MB L4 cache, and a GCX/iGPU. With better revenue streams, a modular design, and an existing design that's being tweaked instead of being built from the ground up, this isn't entirely unreasonable.

This is exactly what I think will happen. With higher revenues and profits more resources are now available to AMD. I think that we could see a 4 CCX server die using 7SoC, a 3CCX die for desktop (though i don't see iGfx) using 7HPC and a 2 CCX notebook die with a massive iGPU (24-32CU) using 7SoC. This is very much possible given that 7SoC 6T is a 70% shrink from 14LPP 9T and is more than sufficient for servers and notebooks. 7HPC 9T is a 55% shrink from 14LPP 9T and will be required to hit 5 Ghz speeds.

 

[DigDog]

I AM A MORON.

why were onboard graphics not actually put onboard - i.e. on the motherboard - instead of on the CPU? surely having on-die integrated graphics will be a waste of resources, or at least a design complication, plus the obvious added heat. If we really want to provide people with an alternative to a gaming GPU, and there's new technology that allows it, you could simply design a small chip that fits on the mobo.
this gives the buyer the choice to have it or not, and simplifies the design - of what is arguably the most complex component in the whole system.
I do get that having the GPU on-die makes it more efficient .. slightly. Not sure this is a valid argument for on-die graphics, though.

i'm talking Intel here, and AMD second.