VEGA die size confirmed

[Sacriledgy]

Honestly after doing some more research, i think Vega will slightly tops GTX 1080..... but double the power usage.

 

now its all about pricing, if they could price it lower than $400 for lower end Vega and less than $500 for higher end Vega i think they will be in a good shape

[f22luke]

2 hours ago, Brooksie359 said:

Are you making a joke? I don't see why they would be related at all.

 

2 hours ago, leadeater said:

There used to be an old issue where Nvidia drivers performed poorly on Ryzen systems, that issue has been fixed. As for Vega + Ryzen that is very unlikely to see any change at all.

AMD clearly stated that they are moving the data around on Vega very differently than previous GPUs.  The HBCC is supposed to be able to leverage the massive bandwidth inherent to HBM2 is such a way that it can access data elsewhere in the system much faster.  Nobody has talked about how the CPU architecture could affect this design.

[leadeater]

1 hour ago, f22luke said:

AMD clearly stated that they are moving the data around on Vega very differently than previous GPUs.  The HBCC is supposed to be able to leverage the massive bandwidth inherent to HBM2 is such a way that it can access data elsewhere in the system much faster.  Nobody has talked about how the CPU architecture could affect this design.

Yes but the HBCC doesn't require Ryzen for that to function, that is handled by the HBCC. If it had required Ryzen or any Zen architecture CPU for that they would have stated so and then marketed that on both Vega and Ryzen/Zen.

 

I've also asked HPE if EYPC or any other up coming AMD products will support any Gen-z or Gen-z type of memory-semantic communication between SoC/GPU/CPU etc and the answer was no. Any of that type of thing is extremely early on in research and development and won't be coming in this generation of products and architectures from anyone.

 

If you want to read up on where this sort of thing is at have a look at HPE The Machine. Also note it features no AMD products or chips anywhere in it, AMD is part of the working group developing the standards and protocols being used along with many other companies.

https://www.labs.hpe.com/the-machine

https://news.hpe.com/gen-z-looks-to-ignite-it-innovation-with-open-high-performance-interconnect-technology/

http://genzconsortium.org/

https://en.wikichip.org/wiki/amd/infinity_fabric

 

I would be rather careful in giving too much praise/credit to AMD for the Infinity Fabric, it is an idea they have had for a while but not alone, because that was not solely developed by them. If it is indeed primarily based on Gen-z technology which I believe it is.

[f22luke]

1 minute ago, leadeater said:

Yes but the HBCC doesn't require Ryzen for that to function, that is handled by the HBCC. If it had required Ryzen or any Zen architecture CPU for that they would have stated so and then marketed that on both Vega and Ryzen/Zen.

 

I would be rather careful in giving too much praise/credit to AMD for the Infinity Fabric, it is an idea they have had for a while but not alone, because that was not solely developed by them.

It would have been a dumb move to make a feature specific to just Ryzen given their current market share.  I am more interested in weather there is some form of latency advantage on that platform.  Correct me if I'm wrong but isn't the infinity fabric a unified protocol for communication?  I was looking for evidence that faster data transfer between the GPU and the CPU can be achieved, given that they may not have to convert everything on die before it's transferred to the CPU.  This would not mean that the HBCC is a Ryzen specific feature but merely benefits from Ryzen.  AMD has been toting for a while that they are the only company doing both highend CPUs and GPUs but as far as I can tell we have not seen any benefit from the integration.  The heterogeneous system architecture was a flop that was only use in certain AMD backed software.

 

It may be possible but would require lots of software tweaks to get this type of communication working.  I'm thinking that this is a case of Vega bringing features well ahead of it's time that just can't be utilized till much later in the future when the software is more mature.

 

I suspect that we will have to wait for Navi and Zen 2 before we will see this tech in action.

[leadeater]

18 minutes ago, f22luke said:

Correct me if I'm wrong but isn't the infinity fabric a unified protocol for communication?  I was looking for evidence that faster data transfer between the GPU and the CPU can be achieved, given that they may not have to convert everything on die before it's transferred to the CPU.

That would only be the case if the communication between the GPU and CPU was using the Infinity Fabric which it isn't. You are correct if it was then yes there could definitely be a latency and bandwidth advantage. That is the ultimate goal AMD is trying to get to but the only off package implementation of the IF is on EYPC between two CPUs.

 

The perfect world is when the CPU, GPU and Storage (NVMe) are all interconnected using the IF/Gen-z and tasks are just executed where it would be fastest to do so. I could see in that type of world CPUs getting much less complex with less instruction sets and acting more as a gatekeeper/scheduler for on package and off package ASICs (GPUs or something else that may spawn from this technology shift).

 

Right now in the desktop world we really do have more PCIe lanes and slots than we can make use of, being able to actually plug a card in that actually helps with gaming performance beyond a GPU would be great or terrible, who knows .

 

Edit:

Also yes I would very much like someone to test Vega FE on Ryzen as well, I just want to see it be done. Rather frustrating places haven't already bothered to do it.

[Taf the Ghost]

4 hours ago, leadeater said:

There used to be an old issue where Nvidia drivers performed poorly on Ryzen systems, that issue has been fixed. As for Vega + Ryzen that is very unlikely to see any change at all.

It's Ryzen + Nvidia 1080/TI + DX12. That issue doesn't seem to be fixed, and we've already got some Vega FE results coming in 15% over the 1080 currently in that scenario.

 

Further, Nvidia's driver system under DX11 doesn't scale well on multiple cores beyond 4 (6 threads, actually), and the games that are starting to use more are being limited by the Driver already. This only happens in a few games at the moment, but it's still something Nvidia is going to need to deal with.

[leadeater]

Just now, Taf the Ghost said:

It's Ryzen + Nvidia 1080/TI + DX12. That issue doesn't seem to be fixed, and we've already got some Vega FE results coming in 15% over the 1080 currently in that scenario.

 

Further, Nvidia's driver system under DX11 doesn't scale well on multiple cores beyond 4 (6 threads, actually), and the games that are starting to use more are being limited by the Driver already. This only happens in a few games at the moment, but it's still something Nvidia is going to need to deal with.

I thought Nvidia did issue a driver update to address that problem? Did it not work or was that not actually the case?

[Taf the Ghost]

Just now, leadeater said:

I thought Nvidia did issue a driver update to address that problem? Did it not work or was that not actually the case?

I haven't seen any updated testing to say it's been fixed. It was most noticeable in Tomb Raider, which itself got patched. (Seemed like it doubled up the issue.) 

 

 

From a Vega FE review looking at Game vs Pro Mode (and he got the Vega FE to OC). Box standard Vega FE was faster than the 1080 FE in RotTR DX12 at 1080p. Testing was done on a 1700X. In Tomb Raider, DX11 at 1080p, the Nvidia 1080 is roughly 20% ahead. In DX12, the Vega FE non-OC is ahead by about 7%. This was to be expected, as it's an Nvidia issue. If RX Vega does come in right between the 1080 and 1080 Ti (reference cards), then we could see 35% wins in a few DX12 scenarios on Ryzen.

 

I saw a decent argument that it's a coherency issue that crops up. Without the scheduler (either Microsoft's or Nvidia's) being fully aware of the Ryzen topology, it spams calls across the threads at will. (Everything is tuned to the Intel Ring Bus.) With Ryzen having the CCX at a much, much lower latency compared to the other cores, calls to the CCX spanned threads are slower than the intra-CCX ones. This takes a couple of cycles to realign. It's not a huge issue, but 5% can be a lot when the testing difference, across a large slate of games, between the 1800X & 7700k (in OC'd conditions) is around 7%.

[leadeater]

15 hours ago, Taf the Ghost said:

I haven't seen any updated testing to say it's been fixed. It was most noticeable in Tomb Raider, which itself got patched. (Seemed like it doubled up the issue.) 

That must have been what I was thinking of, tbh I give exactly two shits about Nvidia and not any more than that 

[Taf the Ghost]

9 minutes ago, leadeater said:

That must have been what I was thinking of, tbh is give exactly two shits about Nvidia and not any more than that 

For all of the fanboy-ism issues, the only problems Ryzen has is at high-peak scale. Games are tuned to get the most out of the Sandy Bridge architecture, it's the reason the G-series chips game so well. What Ryzen has done is simply show that a lot Bulldozer was bad and everyone was capitalizing on quirks in Sandy Bridge to get the most out of the engines. It's the reason things like Crysis love Ryzen and ARMA 3 doesn't. 

 

Intel would actually be in a really bad place if they weren't bringing Coffee Lake to 6c on the Z-platform. We're already getting a collection of games that utterly max them out. Average frame rates are higher, but their frametimes fall off. (TechDeals was chatting about this recently and why he games on his Ryzen system. Those cores matter going forward.)

[leadeater]

Just now, Taf the Ghost said:

For all of the fanboy-ism issues, the only problems Ryzen has is at high-peak scale. Games are tuned to get the most out of the Sandy Bridge architecture, it's the reason the G-series chips game so well. What Ryzen has done is simply show that a lot Bulldozer was bad and everyone was capitalizing on quirks in Sandy Bridge to get the most out of the engines. It's the reason things like Crysis love Ryzen and ARMA 3 doesn't. 

 

Intel would actually be in a really bad place if they weren't bringing Coffee Lake to 6c on the Z-platform. We're already getting a collection of games that utterly max them out. Average frame rates are higher, but their frametimes fall off. (TechDeals was chatting about this recently and why he games on his Ryzen system. Those cores matter going forward.)

The last mainstream Intel system I used was an E8400 and I got it used off trademe (our ebay), before that I was AMD and after that I've been on Intel X series. I mostly play RTS and TBS games so the whole 4 core gaming limit has really pissed me off for so freakin long, my games can and need more cores god dam it!

[Taf the Ghost]

4 minutes ago, leadeater said:

The last mainstream Intel system I used was an E8400 and I got it used off trademe (our ebay), before that I was AMD and after that I've been on Intel X series. I mostly play RTS and TBS games so the whole 4 core gaming limit has really pissed me off for so freakin long, my games can and need more cores god dam it!

Haha. I game light enough for it not to matter to me, but I care about future upgrades and I give advice to people I know on technical stuff. (And I get the occasional, "Welcome to the Christmas party! While you're here...") But I've been on multi-core parts for a long while, as well, and the ability to do multiple tasks is just so nice. Heck, I've normally been on multiple monitors for a very, very long time.  I got started back when CRTs took up way too much space, haha.

[f22luke]

12 hours ago, leadeater said:

That would only be the case if the communication between the GPU and CPU was using the Infinity Fabric which it isn't. You are correct if it was then yes there could definitely be a latency and bandwidth advantage. That is the ultimate goal AMD is trying to get to but the only off package implementation of the IF is on EYPC between two CPU.

Oh yea I forgot about PCIe encapsulation.  Hmm this makes me interested in Ryzen with integrated graphics.  If it comes with a stack of HBM on die that would be epic.... aw man now we can't use that word like we used to.

[cj09beira]

14 hours ago, Taf the Ghost said:

Haha. I game light enough for it not to matter to me, but I care about future upgrades and I give advice to people I know on technical stuff. (And I get the occasional, "Welcome to the Christmas party! While you're here...") But I've been on multi-core parts for a long while, as well, and the ability to do multiple tasks is just so nice. Heck, I've normally been on multiple monitors for a very, very long time.  I got started back when CRTs took up way too much space, haha.

i had a 1080p and a old crt for a long time,

i have seen the multi-monitor light, and i can't live without it