Vega/Navi Rumors (Updated)

[antihelten]

Vega x2 was leaked before, it exists.

Vega x2 was indeed leaked, but not all leaks are true. Specifically Vega x2 was leaked in a set of slides that were clearly not official (claimed to be internal slides), and most importantly slides that also claimed that we would be seeing Vega on 7nm, even though AMD has just made it clear in their latest presentation that Vega is only 14nm and 14nm+, with Navi being the only product on 7nm.

So I would dare say that this particular leak is looking quite shaky, especially considering that there has been no other corroborating leaks for Vega x2 to my knowledge.

 

[MrTeal]

..... No need to write a book for this queation, please. The voltage/wattage curve is simply not linear. The efficiency does not come from it being dual GPU itself, but rather how they always design them. And please also don't try to convince me with the outliers. It's like saying NVIDIA is a shit company just because of the Titan Z.

Sent from my VTR-L09 using Tapatalk

Ah, so you're saying that since they always run dual GPU cards at lower clocks and thus lower voltages the dual GPU cards will be more efficient? Fair enough, but that's card specific as opposed to being an intrinsic feature of dual GPU cards and you probably should have clarified that in your post calling Valantar out on it.
IE, the 5970 ran much slower and was better binned than the 5870, so it used a lot less power.
OTOH, the 295X2 runs both GPUs at the same speed at the 290X, and it does consume twice the power, as does my 390 Dual-Core.
It's not at all a hard and fast rule.

 

[biostud]

My 7990 is rated 275w which is lower than 2x 7970.

 

[OatisCampbell]

They might have, but that slide is clearly formatted vaguely on purpose. After all, concrete roadmaps look more like this:

The clearly delineated time scale is the key. If a slide/PR image lacks that, I would not take it at face value when it comes to launch times.

The slide AT, and many other sites were mislead by, was either made to mislead or somebody's big error that no one at AMD pointed out for months.

Even in conjunction with that other slide Veradun posted it looks like "Q1 2017 at the latest" when you consider one slide shows a box straddling the line between Q4 2016/Q1 2017 and the other just says 2017. How could anyone have left Capsaicin with any impression OTHER than Vega would be launching very early this year.

EDIT: It could also be "holiday 2016" was the "in a perfect world" prediction, and something just didn't work out. My guess is AND knows how to make an accurate graph.

 

[lobz]

Ah, so you're saying that since they always run dual GPU cards at lower clocks and thus lower voltages the dual GPU cards will be more efficient? Fair enough, but that's card specific as opposed to being an intrinsic feature of dual GPU cards and you probably should have clarified that in your post calling Valantar out on it.
IE, the 5970 ran much slower and was better binned than the 5870, so it used a lot less power.
OTOH, the 295X2 runs both GPUs at the same speed at the 290X, and it does consume twice the power, and it does consume twice the power, as does my 390 Dual-Core.
It's not at all a hard and fast rule.

I agree with you in almost everything. But I'vr tried to make it clear, ghat I wanted to get the conversation back to track again, because w3rd went nuts claiming vega x2 will be the mother of God, then Valantar likely got pissed and dismissed the idea of vega2 because 2x performance would need 500W and it's gotten ridicous enough for me to try to interfere. I'm very excited about this whole topic, the technilacities of it too, regardless of how well Vega will do, and it can get very annoying when we are feeding the trolls.
Mostly: I'm very curious on how AMD will first implement IF on a GPU.

Sent from my VTR-L09 using Tapatalk

 

[IEC]

I would be surprised if IF-linked GPU cores are possible with Vega. I expect that with Navi. The IF is the most exciting piece of news from any major player in a long time because it means the potential for better scalability, great perf/W, and most importantly, lower manufacturing costs. We'll see if AMD can realize all those benefits in the GPU space as they appear to have done in the CPU space.

 

[CatMerc]

Vega x2 was indeed leaked, but not all leaks are true. Specifically Vega x2 was leaked in a set of slides that were clearly not official (claimed to be internal slides), and most importantly slides that also claimed that we would be seeing Vega on 7nm, even though AMD has just made it clear in their latest presentation that Vega is only 14nm and 14nm+, with Navi being the only product on 7nm.

So I would dare say that this particular leak is looking quite shaky, especially considering that there has been no other corroborating leaks for Vega x2 to my knowledge.

You do realize those slides gave us 100% correct details about Vega, right?

While the slides themselves were only released after MI25 was announced with the specs, the information contained in them was leaked weeks before, and all ended up being true. As for 7nm, plans change. Fiji was initially a 20nm product that turned into a 28nm one.

Internal slides tend to be quite a bit uglier than what we're used to.

You don't have to believe anything. AMD already told us all about it and how it works. (connect the dots)

No they haven't. We barely have any technical information about Infinity Fabric. We know it's a standard control and communication protocol used for all AMD products after Ryzen, but beyond that exact details are near non existant.

Also, why do you believe "a GPU die" limits what AMD can do, differently than a CPU, in regards to their fabric..? Knowing AMD is giving us ThreadRipper... explain technically (show us your technical doubts), why AMD can't produce TitanRipper?

In laymen's terms, why does the interposer care if a GPU, or CPU is sitting on it?

Because the issues involved are different. Ryzen's CCX's can be thought of like a very fast multi-socket system. In a multi-socket system, as long as there isn't too much communication between dies, processing still works perfectly fine. Ryzen is somewhat like that, but the penalty is MUCH lesser due to both CCX's sitting on the same die/being connected through a quick link on interposer with Threadripper/Epyc.

GPU's on the other hand, don't work quite as well when you bring a "multi-socket" (CF/SLI) configturation to them. That's due to various issues involving bandwidth and latency, needing to construct a final image, temporal algorithms, and having to intelligently split the workload. Having both GPU's sit on an interposer doesn't solve the issues, it only minimizes them as you now have more bandwidth and less latency involved in communication.

Secondly, AMD's Vega x2 design, doesn't use any more memory.

The GPUs share a unified space and can each draw from high bandwidth memory what they need, independently from each other, or as a single unit. There is no AFR, and the cores (sitting on fabric) can actually assist each other. Unlike any x-fire/sli design, which were interdependent on each other, while also working separately, from one another.

Not only do you have 100% scaling with fabric, you actually technically achieve more performance because unused GPU resources can be decoupled and virtualized/used for sub processes. (Kind of like partitioning(shaders) on a HD). There is so much, this fabric technology brings.

The question is, can killer gamers wait until x2 arrives..?

Good luck splitting that workload to get anywhere near 100% scaling lol

 

[antihelten]

You do realize those slides gave us 100% correct details about Vega, right?

While the slides themselves were only released after MI25 was announced with the specs, the information contained in them was leaked weeks before, and all ended up being true. As for 7nm, plans change. Fiji was initially a 20nm product that turned into a 28nm one.

Internal slides tend to be quite a bit uglier than what we're used to.

100% correct, except for the stuff that was wrong huh (not to mention all of the stuff that is still unconfirmed which would be about 50% of the info in those slides)

And yes plans change, but you don't seriously believe that AMD didn't know which architectures were going on which nodes a mere 6 months ago?

 

[tential]

My experience is that anytime someone uses this statement, they really can't explain their theory or belief. No problem however.

No problem at all. I'm not giving you a budget 150k+ worth of degrees. You want to learn more about the material, go get your finance, economics, and accounting degree, then get your Masters then and additional certifications and come back to the topic.
I find that people too lazy to even learn the basics of a subject they want to engage on (in this case finance) aren't worth writing a full couple years of course work to explain basic concepts.
No problem however, I'm moving beyond having financial discussions with people who lack credentials to properly engage me.

It's a tech forum. So we'll stick with the tech, which is still lackluster from amd. I'm still lost as to how people think a top end competing card to Nvidia titan class gpus being available 1 full year later is ok. Nvidia got 2 titans before amd could even get out the gate.

 

[OatisCampbell]

No problem at all. I'm not giving you a budget 150k+ worth of degrees. You want to learn more about the material, go get your finance, economics, and accounting degree, then get your Masters then and additional certifications and come back to the topic.
I find that people too lazy to even learn the basics of a subject they want to engage on (in this case finance) aren't worth writing a full couple years of course work to explain basic concepts.
No problem however, I'm moving beyond having financial discussions with people who lack credentials to properly engage me.

It's a tech forum. So we'll stick with the tech, which is still lackluster from amd. I'm still lost as to how people think a top end competing card to Nvidia titan class gpus being available 1 full year later is ok. Nvidia got 2 titans before amd could even get out the gate.

E-peen for education?

AMD has far less R&D capital, I'd think one with all of your economic training might intuit they might actually be over achieving to more or less compete with NVIDIA. (rather than "lackluster")

Whether we think it's "OK" or not is irrelevant, most of us are just glad they're creating parts that give us choices.

 

[richaron]

I'm still lost as to how people think a top end competing card to Nvidia titan class gpus being available 1 full year later is ok. Nvidia got 2 titans before amd could even get out the gate.

I'm at a loss as to how you missed all that talk about how Polaris being mainstream chips. You know, how everyone knew AMD wasn't aiming for high end until 2017? But somehow you missed it.

Now you seem to think the sky is falling down because AMD is a year later to release an enthusiast class chip, even though it all appears pretty much on schedule. And it's like you think GP102 cards are suddenly meaningless now, since their first iteration was released 10 months ago? Because you seem to be claiming it's not "ok" to release anything which competes with GP102.

The halo effect is essentially meaningless when comparing the merits of individual cards in different price brackets. But I'm sure someone with such a wealth of knowledge in the business and finance sectors would be much more interested in the halo effect.

 

[Guru]

The cards are coming out late, AMD knows this, they even hinted at possible late Q1 release, their CEO Su hinted in her presentations, I think they were hoping late Q1, early Q2, and it will end up late Q2.

Nvidia did release the 1080ti in the meantime and they are announcing some monster Volta chip in late Q4, which means this year they will have another monster performer.

The thing though is that that designed looked extremely close to the current polaris design, just optimized in a way for computing and stuff like that. The issue with their design is that Polaris is not very good at Vulkan or DX12, they are clearly losing to AMD there, there are cases where the FuryX actually beats the 1070 in DX12, in others it comes close, even in BF1 the difference between OC RX 580 at 1450MHz(which most cards can reach) and a stock 1070 at DX12 is almost non existent.

So they will need to significantly improve their architecture for the new API's, otherwise AMD's Vega GPU's will be competitive against Volta. I mean the Titan xp and 1080TI do have a ton of raw horse power that even though they are okayish only at DX12/Vulkan they still get huge performance either way, but I expect Vega big chip to easily beat the 1080ti at DX12 and Vulkan.

 

[CatMerc]

100% correct, except for the stuff that was wrong huh (not to mention all of the stuff that is still unconfirmed which would be about 50% of the info in those slides)

And yes plans change, but you don't seriously believe that AMD didn't know which architectures were going on which nodes a mere 6 months ago?

For a GPU that is late 2018-early 2019? Yes, node plans could have changed a mere 6 months ago.

The slides gave too much accurate information to dismiss them.

 

[antihelten]

For a GPU that is late 2018-early 2019? Yes, node plans could have changed a mere 6 months ago.

The slides gave too much accurate information to dismiss them.

No, node plans do not change that shortly before release, since node is one of the first things that are locked down when planning a new design (not to mention that Vega 20 on 14nm makes zero sense, since then it would just be identical to Vega 10). At best the slides were correct about Vega 20 on 7nm, but Vega 20 has since then been canceled altogether, or at worst the slides were fake and Vega 20 was never planned for 7nm (and quite possibly doesn't exist at all).

And what accurate info are you talking about exactly. Most of the info was either already known (64CUs, HBM2, packed math), or is still unconfirmed (DP rate, Vega 20, Vega 10 x2).

The only thing that wasn't known at the time was the inclusion of hardware page management support (aka HBCC), however it is interesting to note that this wasn't mentioned at all back in september 2016 when Videocardz first reported on the slides, instead they waited until the end of december 2016 for some reason. In other words Videocardz had to wait until AMD had already announced the existence of HBC and the HBCC, before they themselves could reveal that the slides had apparently mentioned this all along (even though Videocardz made zero mention of this back in september).

 

[w3rd]

You do realize those slides gave us 100% correct details about Vega, right?

While the slides themselves were only released after MI25 was announced with the specs, the information contained in them was leaked weeks before, and all ended up being true. As for 7nm, plans change. Fiji was initially a 20nm product that turned into a 28nm one.

Internal slides tend to be quite a bit uglier than what we're used to.


No they haven't. We barely have any technical information about Infinity Fabric. We know it's a standard control and communication protocol used for all AMD products after Ryzen, but beyond that exact details are near non existant.

Because the issues involved are different. Ryzen's CCX's can be thought of like a very fast multi-socket system. In a multi-socket system, as long as there isn't too much communication between dies, processing still works perfectly fine. Ryzen is somewhat like that, but the penalty is MUCH lesser due to both CCX's sitting on the same die/being connected through a quick link on interposer with Threadripper/Epyc.

GPU's on the other hand, don't work quite as well when you bring a "multi-socket" (CF/SLI) configturation to them. That's due to various issues involving bandwidth and latency, needing to construct a final image, temporal algorithms, and having to intelligently split the workload. Having both GPU's sit on an interposer doesn't solve the issues, it only minimizes them as you now have more bandwidth and less latency involved in communication.

Good luck splitting that workload to get anywhere near 100% scaling lol

Yeah, it is a connect the dots type of thing, & you are missing a few dots.
The info is out there, you have to just listen.

Secondly, You are talking about "multi-socket" and using what you know and trying to apply to Vega. Which is not the same as using Infinity Fabric. Vega x2 is one chip, just like ThreadRipper is one chip, there is no CF/SLI drivers. Oddly, the things you keep mentioning are software rendering issues, not technical hardware issues.

GPU love parallelism more than CPUs do. Not to mention, a GPU doesn't utilize fabric like a CPU does, your questions are flawed premise surrounding the CCX. GPUs are an output device and you could never get two CPUs to talk over SLI/Crossfire, because they do central processing. Yet AMD has done exactly this with "fabric", and you are suggesting they can't do it with dumb GPUs..?

Even though infinity fabric has 100x the bandwidth than X-Fire and SLI.

 

[swilli89]

Yeah, it is a connect the dots type of thing, & you are missing a few dots.
The info is out there, you have to just listen.

Secondly, You are talking about "multi-socket" and using what you know and trying to apply to Vega. Which is not the same as using Infinity Fabric. Vega x2 is one chip, just like ThreadRipper is one chip, there is no CF/SLI drivers. Oddly, the things you keep mentioning are software rendering issues, not technical hardware issues.

GPU love parallelism more than CPUs do. Not to mention, a GPU doesn't utilize fabric like a CPU does, your questions are flawed premise surrounding the CCX. GPUs are an output device and you could never get two CPUs to talk over SLI/Crossfire, because they do central processing. Yet AMD has done exactly this with "fabric", and you are suggesting they can't do it with dumb GPUs..?

Even though infinity fabric has 100x the bandwidth than X-Fire and SLI.

I think everyone understands the possibility of infinity fabric but at this point its just derailing from the thread. AMD has given zero information about a multi-gpu on interposer for Vega and has only alluded to GPU scaling for the Navi generation.

If Vega comes in X2 form it will be in the traditional form we are used to seeing.

 

[w3rd]

Caspian & Cream, a lead AMD engineer talk in detail about Infinity Fabric and mentions that APU, GPU or CPU, it doesn't matter...!

Same event, lead software house mentions they are near 100% scalability.. in their own software, but more can be gained, given it's newness. Same even AMD talks about getting many of the software houses on board to their new technology & multi-rendering

Also, over the last 2 months many behind the scenes personalized conversation also took place. That hints at AMD path, which are not so different than Dr Su's key note speeches. That suggest new gaming technology. (We all know SLI/Xfire & plx chips are old tech).

So what you are saying is, that since AMD hasn't released their bombshell, you can't discuss it, or even allow oneself to admit it exists..? Or even discuss it in a Vega/Navi rumor thread..?

Hogwash..

 

[Glo.]

Caspian & Cream, a lead AMD engineer talk in detail about Infinity Fabric and mentions that APU, GPU or CPU, it doesn't matter...!

Capsaicin & Cream*, fixed for ya'.

So then enlighten us all what he was talking about specifically, about Infinity Fabric. You are making out of AMD's offerings and tech god almighty of hardware, which it is not.

 

[w3rd]

LOL...
I just read Lion, Witch & the Wardrobe this weekend to me nephew.

Trolling is not allowed
Markfw
Anandtech Moderator

 

[T1beriu]

Yeah, it is a connect the dots type of thing, & you are missing a few dots.

Please stop about connecting the dots already. You have proven time and time again you're not able to do that yourself. Or your dots live in a parallel universe?

 

[w3rd]

Capsaicin & Cream*, fixed for ya'.

So then enlighten us all what he was talking about specifically, about Infinity Fabric. You are making out of AMD's offerings and tech god almighty of hardware, which it is not.

Simple exercise:
How many "pins" are involved with SLI/Xfire ribbon cables.
Now let your imagination run wild with fabric.

If you don't see it, this Holiday season, you will.

 

[swilli89]

Capsaicin & Cream*, fixed for ya'.

So then enlighten us all what he was talking about specifically, about Infinity Fabric. You are making out of AMD's offerings and tech god almighty of hardware, which it is not.

The unfortunate thing about AMD right now is they simply don't have the resources for that many parallel projects. I think if they had a lot more funding we would have seen a larger Polaris, something like 350-400mm2 that could have come close or equaled the 1080, but they simply didn't have the manpower to scale up their first gen 14nm design beyond what they did.

 

[Qwertilot]

Or, conversely of course, to feed the improvements from Vega back to a polaris sized chip & launch a whole new generation that way.

 

[guskline]

Just give me a single Vega that competes with the GTX1080 to GTX1080TI in the 1440/4k resolutions.

Just sold one of my Radeon RX480s that I had in CF in my Ryzen 1800x rig in anticipation of Vega.

 

[IEC]

Just give me a single Vega that competes with the GTX1080 to GTX1080TI in the 1440/4k resolutions.

Just sold one of my Radeon RX480s that I had in CF in my Ryzen 1800x rig in anticipation of Vega.

Yes, Vega can't come soon enough. I almost bought a 1080 Ti earlier than I had planned. I'll probably still get one, but I plan to pair it with a Skylake-X rig and put Vega in at least one of my AMD rigs.