AMD Vega 10, Vega 11, Vega 12 and Vega 20 Rumors and Discussion

I know about OEMs, but what share of the very high-end market do they really represent?

 

Looks like there might be up to seven Vega SKUs, though that probably includes Radeon Instincts, Radeon Pros, etc.
https://videocardz.com/67524/amd-linux-drivers-lists-seven-vega-10-ids

 

As far as high end boutique builders, the mentioned Falcon Northwest, alone, would make up a healthy chunk of overall volume.

The biggest issue I see is multi-gpu. These boutiques will definitely support that. Demanding six expansion slots to pull it off is a tall order. It's obviously possible, but it's surly not a change that the oems would appreciate.

Isn't full fat Vega 10 expected to put off 225W?

That's closer to the 250W 1080 Ti than the 180W 1080. It's one thing to bin a chip for a 20ish% drop in power to get down to a laptop-worthy 140-150W. But you're bit getting 225W down that far. There's a reason we haven't seen a mobile GP102 yet.

I'm wondering if there's a Nano successor in there. It's likely that a good number of those models are pro-related parts, but it still leaves some left over.

 

It's undeniably possible.

We all know that you can take a big chip and underclock it to easily beat smaller moderately clocked chips in perf/watt. It happens all the time on the pro side (the P4 has GP104 at 75W).

But it's expensive. That cost can be swallowed by pro margins, but consumer margins are tighter.

And remember that AMD was in a similar position when the Nano came out. Hawaii wasn't going to fit on an MXM-B card, so AMD was ceding the entire high performance laptop market to Nvidia and the mobile 980. To me, it would've made sense to put fiji in mobile, but that never happened despite the existence of the Nano.

Maybe it was a size issue (~600 mm2, 4 stacks of hbm, etc), but I'm no engineer so I can't point to a smoking gun that explains why Fiji never rescued amd's mobile predicament.

But we know that it didn't happen. So why will Vega 10 be different? The MXM-B standard hasn't changed to my knowledge. Vega 10's die might be a little smaller, but hbm2 die require over twice as much area, so the overall interposer size is probably comparable.

 

What's that about HMB2 stacks being twice as large? I don't remember reading that before.

That's what the interposer looks like, in any case: http://wccftech.com/amd-vega-gpu-pictures-hbm2-official/

There seems to be way less wasted space than there was for Fiji. (Edit: I guess the stacks themselves do look larger, but thanks to better packing, I think the package is smaller for Vega, but I'm too lazy to check.)

 

I thought the same thing, but was corrected a couple months ago.

http://www.anandtech.com/show/9969/jedec-publishes-hbm2-specification

 

Larger OEM's like Dell, HP, Lenovo, they own a large amount of the high end too.

HP have been selling 980's, up to the Ti since launch, they also now have been selling 1070's, and these are their gaming PC's, its there top of the line regular consumer PC's.

Three slot coolers as a reference design will also give bad press right off the bat. Also I would imagine the cost of the structural components will need to go up too.

Just look at the FX series, the two slot design didn't do anything for it even before it was reviewed, so prior to performance being revealed, people were poking at it.

 

Thanks! I wonder whether that's part of the spec or just what SK Hynix happens to be manufacturing at the moment. If the latter is true, perhaps we'll get smaller packages later.

I see. Still, I think there's nothing wrong with splitting the lineup: RX Vega with a dual-slot blower for OEMs (and Crossfire, and people who want something a bit more reasonable and affordable) and RX Vega Ultimate FX Black Edition XTX XXX with a 4-slot cooler and a 14cm fan.

 

yeah having those specialty cards are always good, just a specific card for specific needs.

Dual slot coolers are pretty much standard now so once they were done and people started to look at it as its needed, no big deal.

 

Anecdotally, I remember reading that only the thickness was defined in the spec and the other dimensions were left up to those implementing it.

And that Hynix was the only one to release full dimensions for their hbm2. That's how the Vega 10 die size estimates were possible when Koduri waved it around - folks knew how big the hbm2 was.

 

I seem to recall similar concerns when the first 2-slot cards came out

Would almost certainly need better support though.

 

Thinking about that, I just looked at all the AM4 boards that released and all I checked appear to have 3 slot clearance. Even mATX that only allows 4 slots by spec; two x16 slots, but one is in the 4th an opposed to 3rd position. Anecdotal, but AM3 was more variable in the products so 3 slot clearance could be a specification for upcoming parts. The space should help cooling 2 slot designs without blowers, but it would also apply to larger 3 slot cards. The smaller footprint with HBM might wraith more reasonable, and as mentioned below, there could be a use for it.

What about a Pro Duo style device? That Vega10x2, originally on a slide, with a 3 slot blower wouldn't be unwarranted. More compact for someone looking at Crossfire/SLI from the onset. Naples and Infinity Fabric does two sockets, what if Vega was doing the same? With HBM/HBM2 the cards are certainly short enough to make that more practical than in the past. HBCC and the "Intelligent Workgroup Distributors" I'd think could make mGPU appear as a single device.

 

Honestly, I'd be absolutely amazed if AMD didn't put a clc on a "Vega 10X2" part.

The Titan Z wasn't a resounding success with its 3-slot cooler, so I doubt that AMD would abandon their multi-year clc relationships/experience to go for something that will surely cause their product to become less popular.

Admittedly, a water cooler is more expensive, but that expense is made palatable when your product is going for over a thousand bucks (maybe higher).

But I must agree that IF AMD went for an air cooler on their inevitable 300+W Vega 10X2 part, then it would need a 3-slot cooler.

 

Yes!
This is the way forward. With increasing cost of making gigantic monolithic dies, it seems logical to make smaller more cost-effective dies put together on a interposer. On the upside, compared to the dual ccx-modules in the Zen architecture, a dual-die gpu wouldn't suffer much from the extra latency penalty through the infinity fabric (or whatever they would use for crosstalk). No more need for buggy micro-stuttery crossfire drivers!

 

There has been some speculation on this possibility before, but could the "Scalability" under Navi in the AMD Capsaicin roadmap be indicate a multi die + interposer approach?

 

Well, the term "scalability" is largely used by AMD since a long time, and so far was mostly refer to an "high end to low end" scalability of the core. ( something who was not Polaris ).
Now we know that, AMD, but too Nvidia, seem goes to an approach for hexascale system with an die to to die approach, basically it is the possibility to set up way more gpu's on a "node" than before. ( see it as put pile of GPU's "dies" side by side ) and who are interconnected..

So i dont know if it is something as Zen with the CCX Fabric interconnect, but without more information, i think it refer more to the "coherent interconnect fabric".
http://wccftech.com/amd-coherent-interconnect-fabric-gpus-cpus-apus/
Im sorry to use this site for the example ( Jan 2016 ) but this was the first article who have pop-up..

Vega will surely be the first to introduct it, with Navi who will push it higher on scalability.

Some more reference on it
http://www.bitsandchips.it/9-hardware/5858-amd-exascale-heterogeneous-processor

It is something "coherent" for AMD to use this same interconnect "inside and outside" the dies, capable of communicate between every piece there ( cpu cores, APU cores+gpu, GPU cores to cores, storage ( Dram,Nvram,Nand etc)..

 

Will that bring any tangible improvements ? I mean the way I see it, this will only replace PCI-E comunnication between the two chips. We'll still get AFR or whatever DX12multi-gpu rendering technique is fashionable these days

 

Still the issue of mounting the radiator which can be problematic for many cases. Titan Z had issues because of SLI scaling. Three slot coolers for devices with potentially double the resources. Naples for example at 4x~200mm2 chips. That's a third larger than Fiji or P100, so it would seem feasible.