AMD: Speculation, Rumors, and Discussion (Archive)

[3dilettante]

If the assumption that Vega 10 is the Greenland that has been leaked before, having at least that part of the Vega generation aligned with Zen could make sense.
Vega 10 would have a different or additional set of interfaces on-die compared to the PCIe x16 connections we currently.

There isn't a PCIe x32 or more slot, so I am not sure if a discrete board can do much with what the silicon on both sides might be able to do.
Whether a discrete Vega could detect a Summit Ridge CPU on the far side of a PCIe slot and change protocols, or if that is permitted, is unclear.

On the other hand, having a proprietary communication channel between boards is not unheard of, so maybe that method could have some use in crossfire? It might depend on whether GMI provisions for off-PCB or off-package traffic.

 

[xEx]

If the assumption that Vega 10 is the Greenland that has been leaked before, having at least that part of the Vega generation aligned with Zen could make sense.
Vega 10 would have a different or additional set of interfaces on-die compared to the PCIe x16 connections we currently.

There isn't a PCIe x32 or more slot, so I am not sure if a discrete board can do much with what the silicon on both sides might be able to do.
Whether a discrete Vega could detect a Summit Ridge CPU on the far side of a PCIe slot and change protocols, or if that is permitted, is unclear.

On the other hand, having a proprietary communication channel between boards is not unheard of, so maybe that method could have some use in crossfire? It might depend on whether GMI provisions for off-PCB or off-package traffic.

Did I missed something? I don't remember hearing about AMD having a proprietary protocol to communicate CPU and GPU. Although since their bought of ATI Ive been waiting to see AMD finally and actually taking advantage of its privilege position of being a major company in both sides.

 

[3dilettante]

An older alleged HPC APU slide shows a Zen-based CPU linked to a Greenland GPU.
A later slide shows this APU as an MCM where a 16-core Zen CPU is linked to a Greenland GPU with GMI links.

Depending on the accuracy of those slides and whether the linking of the Greenland name to Vega 10 is accurate, the silicon could have additional interfaces that might include being able to repurpose the PCIe bus.
The purported leaked die shot of AMD's Summit Ridge, if that is the base architecture for the CPU component of the HPC APU, leaves the possibility that GMI can be run over the same interface as PCIe.

 

[Razor1]

another one this time with a comparison video with a heavily overclocked 970.

 

[silent_guy]

Well, videocardz' reasoning is lacking, comparing it to Hawaii (no compression, low GDDR5-clocks) and Fiji (totally different memory subsystem, and almost twice as many CUs contending for bandwidth) for conclusion.

Compression, memory subsystem, and memory clocks are ways to increase the effective BW that's available to the rest of the system, are there ways to increase the effective number of ROPs beyond the physical number of instances? I don't think you can. ROPs seems to be relatively simple beasts that don't allow for SM-like magic tricks such as improving IPC.

So, at best, those memory BW improvements will help the ROPs reach their peak performance, but that's about it.

Hawaii has 44 CUs. Polaris 10 has 36, but clocks and (supposedly) improved efficiency will make the latter better.

Hawaii has a BW of 384GBps, Polaris 10 has 256 (if using GDDR5X.) Seems hard to overcome with compression.

And clock won't be able at all to compensate going from 64 to 32 ROPs.

If true, it's going to be a small miracle if the 480 will outperform a 390X. It would go a long way to explain the low price...

Edit: 480 instead of 380.

 

[xEx]

An older alleged HPC APU slide shows a Zen-based CPU linked to a Greenland GPU.
A later slide shows this APU as an MCM where a 16-core Zen CPU is linked to a Greenland GPU with GMI links.

Depending on the accuracy of those slides and whether the linking of the Greenland name to Vega 10 is accurate, the silicon could have additional interfaces that might include being able to repurpose the PCIe bus.
The purported leaked die shot of AMD's Summit Ridge, if that is the base architecture for the CPU component of the HPC APU, leaves the possibility that GMI can be run over the same interface as PCIe.

ohhh that would be interesting. about the PCI lanes you don't really need 16 lanes, with 8 lanes there is no difference in gaming(about 1% which is within the error margin) so you can use a 8+8 lines without a problem.

Although wouldn't this be an enterprise feature? or could gamers would really get a remarkable improve in performance?

 

[Razor1]

Think ya mean 480 but yep to the rest of it

 

[3dilettante]

ohhh that would be interesting. about the PCI lanes you don't really need 16 lanes, with 8 lanes there is no difference in gaming(about 1% which is within the error margin) so you can use a 8+8 lines without a problem.

There are a lot of unknowns about what Summit Ridge may have or Greenland. However, if GMI can reuse the physical PCIe interface and Vega silicon is Greenland, then it may have a modified dual-mode interface.
If there isn't a dual-mode interface, then there's a standard PCIe x16 interface and additional interfaces on the side.
Whether that capability can be used in a discrete situation is not known, but it would mean that it would physically be there and might have a use.
Some of the possible bandwidth numbers might not allow for 16 lanes of anything on the die to be sufficient.

Other unknowns include the speed and width of a GMI link, for which Greenland is supposedly able to serve as an endpoint for 4 links.
That's an MCM connection, and may not apply in an expansion card scenario, or for a standardized PCIe card slot.
It would most likely be faster than a PCB or connector could sustain, but it might be able to run at a reduced rate. If the PCIe slot cannot play around with the signalling, GPU boards in the past with Xfire could run proprietary connections between each other.

Aligning Zen and Greenland(Vega 10?) could at least allow the marketing of a value-add for the two in a non-gaming niche.

Although wouldn't this be an enterprise feature? or could gamers would really get a remarkable improve in performance?

It could be an HPC feature or possibly useful in a professional context. There could be latency benefits if a desktop system could transparently use the coherence capabilities to implement some of the synchronization and atomic functionality, or possibly a more consistent Xfire method as a prelude to Navi.
Both Zen and Vega are facing strong competition in their respective spheres, so having some kind of value-add if it exists is going to be seriously needed.
It might matter for a mid-gen console update next year, as well.

 

[3dilettante]

Compression, memory subsystem, and memory clocks are ways to increase the effective BW that's available to the rest of the system, are there ways to increase the effective number of ROPs beyond the physical number of instances? I don't think you can. ROPs seems to be relatively simple beasts that don't allow for SM-like magic tricks such as improving IPC.

Between Hawaii and Tonga/Fiji there seems to have been an improvement in the rates for 4x FP16, where it was almost full rate.
http://www.hardware.fr/articles/937-4/performances-theoriques-pixels.html
At least in those scenarios where Hawaii by all rights should have doubled Tonga, it was much closer.

Hawaii has a BW of 384GBps, Polaris 10 has 256 (if using GDDR5X.) Seems hard to overcome with compression.

Tonga was eventually able to overcome a larger proportional deficit with Tahiti, although that was with reasonably close unit counts.
Also, couldn't GDDR5 at 8Gbps be used in lieu of GDDR5X?

 

[ninelven]

another one this time with a comparison video with a heavily overclocked 970.

It isn't even a valid comparison... the 970 has a much better CPU, twice the ram, the 970 is heavily OCd, and the rx480 settings are more demanding... about as stacked a deck as I have ever seen. Really, garbage like that should not be allowed here. B3D should be better than this...

 

[3dilettante]

I'm intrigued with the possibility of 32 ROPs in light of Sony's disclosed mid-gen console refresh.
The rumored specs there make the RX480 sound like an implementation that looks like a hypothetical headless Neo.
It brings to mind a scenario where a discrete implementation had some design parameters influenced by a console, like Bonaire's commonality with its console brethren but even more specific to just one. Since that console is specifically targeting an older semi custom APU, that would have a perverse outcome where parts of Polaris are what they are to keep in line with the original PS4.

If someone else paid for part of what has gone into Polaris, it might be a reason why it is this cheap in both the positive and negative ways it is being discussed.

 

[silent_guy]

Also, couldn't GDDR5 at 8Gbps be used in lieu of GDDR5X?

Right, my BW numbers are wrong. So BW should be quite a bit closer with GDDR5X. (384 vs 320.) Not insurmountable with improved compression. Still a big gap if they use GDDR5.

 

[Esrever]

Is amd cards ever rop bound? They seem to scale better with resolution than nvidia cards. I am guessing Hawaii has more rops than it needs.

 

[Grall]

Is amd cards ever rop bound?

Dunno what precisely in the GPU bounds my R390X, but a very light on the shaders game like WoW: Legion beta can drop in framerate quite noticeably on 1440P rez display when you whack up the view distance in zones with a lot of trees (and thus overdraw.)

 

[Jawed]

That is if the info is right anyway; but no-one still reports if Polaris 10 is an evolution of Tonga or Hawaii as I have asked in the past, or those that do know are under NDA.

What a stupid question. Why would it be an evolution of Hawaii.

 

[Jawed]

Polaris 10 has 256 (if using GDDR5X.)

So, what you're saying is that NVidia with 1070 has access to 256GB/s of memory with GDDR5, but AMD doesn't.

 

[CSI PC]

What a stupid question. Why would it be an evolution of Hawaii.

Sigh,
If you look back I repeatedly mention IMO it is probably Tonga (in fact I was saying that quite awhile ago as well, if it was Fiji that breaks their perf/watt comparison I think), but to be balanced because it has never been clarified what it is an evolution of I opened up to either.
Why?
Because the performance cards apart from Fiji were Hawaii while Tonga was never released as an enthusiast/performance product even when it is newer.
Furthermore nearly every article out there compares Polaris to Hawaii including in specs (PCGamesHardware is one of the ones that does go on about Tonga-Fiji).......
Maybe you should write to tech sites and say they are being stupid, or anyone thinking of comparing 390/390x to 480 without providing full caveats.

Cheers

 

[silent_guy]

So, what you're saying is that NVidia with 1070 has access to 256GB/s of memory with GDDR5, but AMD doesn't.

I'm not saying anything about 1070 at all. Just comparing AMD against AMD. Though we know that Nvidia has traditionally been much better at extracting performance out of a given amount of BW. Maybe AMD finally managed to fix that, but with 32 ROPs I'm not holding my breath.

 

[Jawed]

I'm not saying anything about 1070 at all. Just comparing AMD against AMD. Though we know that Nvidia has traditionally been much better at extracting performance out of a given amount of BW. Maybe AMD finally managed to fix that, but with 32 ROPs I'm not holding my breath.

No, you made a comment about AMD requiring GDDR5X to get to 256GB/s.

ROPs is a whole other question. And, as I posted only last night, I'm pretty sure Polaris 10 has 32 ROPs. In my view it doesn't need 64 to get to 390X performance.

Better geometry/discard + increased clocks will get there.

 

[Jawed]

Sigh,
If you look back I repeatedly mention IMO it is probably Tonga (in fact I was saying that quite awhile ago as well, if it was Fiji that breaks their perf/watt comparison I think), but to be balanced because it has never been clarified what it is an evolution of I opened up to either.
Why?
Because the performance cards apart from Fiji were Hawaii while Tonga was never released as an enthusiast/performance product even when it is newer.
Furthermore nearly every article out there compares Polaris to Hawaii including in specs (PCGamesHardware is one of the ones that does go on about Tonga-Fiji).......
Maybe you should write to tech sites and say they are being stupid, or anyone thinking of comparing 390/390x to 480 without providing full caveats.

Think for a second. Why would AMD ditch, for example, the delta colour compression tech from Tonga in order to make Polaris an evolution of Hawaii. A key component of Tonga that differentiates it from Hawaii. An evolution from Hawaii.

Do you see how stupid your question is?