Vega Rumors

[Magic Hate Ball]

Yes, i had exactly 290X in mind when i mentioned that 300W blowers do not really exist. 290X was not even a 300W card (furmarks aside) and it's blower already failed hard.

~270W

You could also override power limit, disable thermal throttling and have it work full clock on 0 rpm fan. See my point?

Nice, you lost your argument now you're just flailing.

Titan X Pascal OC'd pulling 300 watts sustained at 88 degrees (high, yes, but sustainable and only 4 degrees above stock) as a blower cooler.

You.

Lost.

 

[N4CR]

Found out that the reason the RX580 clocks better is tighter/upgraded voltage controller.. it's literally the only difference
74Mhz isn't much really for base clocks. It uses more power, means more volts.
They OC slightly higher on average it appears but having trouble finding an OC results compilation, maybe on a mining forum this can be found. This could be down to finer voltage control from a better controller which is the only known major component change...

Blower coolers eh, here's my 10c;
7970 reference 210W, what do you think happens when you run it at 1.3V and 1.3GHz? It's easy 300W+ range, just like a barely OC'd 290x which, like the 1080 will also thermal throttle unless you run the fan at a high rate 70-90% depending on your ambient temps. Funniest part is it's also not much slower than a stock 290x at those speeds. I would run that 7970 @ 80-90 degrees with the blower screaming sometimes. Merciless overclocking.

Anything with heavy OC in the 200W+ range will be approaching 300W or exceeding it - yes it's possible too on a blower cooler.
Even Quadro P6000 is 250W and a blower cooler. 300W is well within expected design margins.

 

[Ieldra]

On a side note, as we mentioned in our architectural breakdown, the amount of power this card draws will depend on its temperature. 408W at the wall at 65C is only 388W at the wall at 40C, as current leakage scales with GPU temperature. Ultimately the R9 Fury X will trend towards 65C, but it means that early readings can be a bit misleading.

Taken from the anandtech fury x review.

The TDP is what you need to dissipate in order to keep the chip within operating conditions.

 

[lolfail9001]

Titan X Pascal OC'd pulling 300 watts sustained at 88 degrees (high, yes, but sustainable and only 4 degrees above stock) as a blower cooler.

88 degrees means it is actually throttling itself at all costs (so basically turbo goes out of the window) unless you raise temp limit as well. You are literally claiming that thermally throttling GPU is cooled "just fine".
So, you lost too.

 

[Magic Hate Ball]

88 degrees means it is actually throttling itself at all costs (so basically turbo goes out of the window) unless you raise temp limit as well. You are literally claiming that thermally throttling GPU is cooled "just fine".
So, you lost too.

Wrong.

See: http://www.guru3d.com/articles_pages/nvidia_titan_x_(pascal)_overclock_guide,4.html

Two points directly from the article where I got the Titan X Pascal power draw.

• The card's temperature under heavy game stress stabilized at roughly 88 Degrees C. We note down the hottest GPU reading, not the average.
  
• We achieve 88 Degrees C max, and thus stay under the 90 Degrees C thermal limit. This maximized the boost clock (at the cost of a lot of noise of course).

 

[lolfail9001]

Wrong.

Check list:

> Raised thermal limit above default throttle point (that article states is actually at 80 degrees, amusing isn't it).

> Raised fan speed to a borderline banshee level

> Done it in the best environment for blower card other than, well, being run in Siberia wilderness.

So, you are still looking to prove my point about blowers not being suitable for 300W TDP conditions? Sure, i am not right to claim it is impossible to use them, but they are anything but usable at this level.

 

[Magic Hate Ball]

Check list:

> Raised thermal limit above default throttle point (that article states is actually at 80 degrees, amusing isn't it).

So, you are still looking to prove my point about blowers not being suitable for 300W TDP conditions? Sure, i am not right to claim it is impossible to use them, but they are anything but usable at this level.

I thought we were on Hardcore Overclocking Forums not Mild Salsa Overclocking Forums here?

Also, it says 90 degrees, right there for max. Hell the stock clocks hit 84 degees!

 

[Ieldra]

I thought we were on Hardcore Overclocking Forums not Mild Salsa Overclocking Forums here?

Also, it says 90 degrees, right there for max. Hell the stock clocks hit 84 degees!

You seem to missing the point. Let's says for a 2.1ghz Titan Xp you need to dissipate 350w to keep it a temperature it won't throttle.

If you now raise the temperature limit the TDP goes down, because your thermal design reference allows for higher temperatures, which reduce the operating clocks.

What exactly is your point then? Throttling is a relative term, throttling relative to what? The base clocks? Boost clocks?

At 88 degrees it is thermally limited, the temperature of the chip is dictating the sustainable clocks, and the temperature limit was raised

 

[Magic Hate Ball]

You seem to missing the point. Let's says for a 2.1ghz Titan Xp you need to dissipate 350w to keep it a temperature it won't throttle.

If you now raise the temperature limit the TDP goes down, because your thermal design reference allows for higher temperatures, which reduce the operating clocks.

What exactly is your point then? Throttling is a relative term, throttling relative to what? The base clocks? Boost clocks?

At 88 degrees it is thermally limited, the temperature of the chip is dictating the sustainable clocks, and the temperature limit was raised

Hi, welcome to the argument.

They were arguing that it's impossible for a Vega FE blower to reach their stated TFlops rating because it'd need water cooling to hit 300watts. But the Titan X Pascal (which isn't even the full Titan Xp) does.... with a blower.

That's all we're talking here.

EDIT: And that's IF IF IF IF it actually pulls 300 watts.

 

[Ieldra]

Hi, welcome to the argument.

They were arguing that it's impossible for a Vega FE blower to reach their stated TFlops rating because it'd need water cooling to hit 300watts. But the Titan X Pascal (which isn't even the full Titan Xp) does.... with a blower.

That's all we're talking here.

It doesn't need watercooling to hit 300w... It needs watercooling to stay within operating temperature at a given clock and voltage, otherwise it would not have watercooling.

Somethings gotta give, watercooled one will be like 30c cooler on the gpu die and that alone is 30w+ difference, more power = more heat = higher TDP needed for same clocks/voltage. It's inevitably going to clock lower.

I don't understand your point regarding the Titan X, it is thermally limited in the review you posted, you're just confirming what they are saying

To put this another way, if watercooling didn't affect your achievable clockrate why do Overclockers watercool their hardware... Have you ever heard of 2100mhz stable operation Pascal cards on air? How many?

 

[Ieldra]

Found out that the reason the RX580 clocks better is tighter/upgraded voltage controller.. it's literally the only difference
74Mhz isn't much really for base clocks. It uses more power, means more volts.
They OC slightly higher on average it appears but having trouble finding an OC results compilation, maybe on a mining forum this can be found. This could be down to finer voltage control from a better controller which is the only known major component change...

Blower coolers eh, here's my 10c;
7970 reference 210W, what do you think happens when you run it at 1.3V and 1.3GHz? It's easy 300W+ range, just like a barely OC'd 290x which, like the 1080 will also thermal throttle unless you run the fan at a high rate 70-90% depending on your ambient temps. Funniest part is it's also not much slower than a stock 290x at those speeds. I would run that 7970 @ 80-90 degrees with the blower screaming sometimes. Merciless overclocking.

Anything with heavy OC in the 200W+ range will be approaching 300W or exceeding it - yes it's possible too on a blower cooler.
Even Quadro P6000 is 250W and a blower cooler. 300W is well within expected design margins.

Y'all keep talking about the TDP as if it's a hard limit in terms of power dissipated.

An Intel stock cooler can dissipate 50,000W - yes that's 50kW.

The problem is that at these levels of power it will be dissipate that heat primarily by melting.

It doesn't matter how you try to rationalize this, whatever results you're getting on a blower cooler with 300w power draw can be improved upon massively with watercooling and lower temperatures. Power consumption will drop, clocks will go up, they will be sustained better etc.

 

[Magic Hate Ball]

It doesn't need watercooling to hit 300w... It needs watercooling to stay within operating temperature at a given clock and voltage, otherwise it would not have watercooling.

Somethings gotta give, watercooled one will be like 30c cooler on the gpu die and that alone is 30w+ difference, more power = more heat = higher TDP needed for same clocks/voltage. It's inevitably going to clock lower.

I don't understand your point regarding the Titan X, it is thermally limited in the review you posted, you're just confirming what they are saying

To put this another way, if watercooling didn't affect your achievable clockrate why do Overclockers watercool their hardware... Have you ever heard of 2100mhz stable operation Pascal cards on air? How many?

I'm not saying going over 300 watts, which is what you're implying with the article I linked. I'm just saying stock Vega FE should be able to be cooled by a blower, that's all. Water is obviously better, but it's physically possible to use a blower and cool this thing.

Secondly, they have a Titan X Pascal hitting 2000+mhz consistently on air here, look at the graph. The mhz stays flat when the load is on the card:

 

[CSI_PC]

Per the review http://www.anandtech.com/show/7457/the-radeon-r9-290x-review

However at the same time how power consumption is being defined is getting far murkier: AMD doesn’t list the power consumption of the 290X in any of their documentation or specifications, and after asking them directly we’re only being told that the “average gaming scenario power” is 250W. We’ll dive into this more when we do a breakdown of the changes to PowerTune on 290X, but in short AMD is likely underreporting the 290X’s power consumption. Based on our test results we’re seeing 290X draw more power than any other “250W” card in our collection, and in reality the TDP of the card is almost certainly closer to 300W. There are limits to how long the card can sustain that level of power draw due to cooling requirements, but given sufficient cooling the power limit of the card appears to be around 300W, and for the moment we’re labeling it as such.

Didn't the 290x have 2 modes though, one with hurricane level of fan and one not quite hurricane level
I love how you did not highlight the section I have in red that is directly between the two sentences you did, meaning the cooling is not up to the job even with uber fan mode.

I guess it also comes down to in all this arguing whether having fans running at nuts levels is acceptable, although the red section suggests the cooling is still not up to it in the Anandtech review for this specific GPU.
Because we might as well say most reference designs are fine from both manufacturers if we just set the fan to near its maximum 24/7.
Cheers

 

[Ieldra]

I'm not saying going over 300 watts, which is what you're implying with the article I linked. I'm just saying stock Vega FE should be able to be cooled by a blower, that's all. Water is obviously better, but it's physically possible to use a blower and cool this thing.

Secondly, they have a Titan X Pascal hitting 2000+mhz consistently on air here, look at the graph. The mhz stays flat when the load is on the card:

Stock Vega can be cooled by a 2x2 inch heatsink because they dictate what 'stock' is.

As for the Titan X at 2ghz, what does that mean? What does it prove? Nothing... It's just a number, stick a waterblock on that card and it will clock higher and run cooler and draw less power... Which is the point everyone has been making. Why should Vega be any different

 

[Ieldra]

This is just like those forum posts where one guy says 'you can dissipate 200w from a single 140mm radiator' and then other people say 80w per 140mm rad. It depends entirely on the conditions you set a 360mm radiator may be good for 800w with a 15c delta but only 50w with a 1c delta

 

[Magic Hate Ball]

Stock Vega can be cooled by a 2x2 inch heatsink because they dictate what 'stock' is.

As for the Titan X at 2ghz, what does that mean? What does it prove? Nothing... It's just a number, stick a waterblock on that card and it will clock higher and run cooler and draw less power... Which is the point everyone has been making. Why should Vega be any different

AMD has listed TFlops for Vega FE. In order to meet that consistently it'll have to stay cool and maintain clocks. They say it won't with a blower cooler and that it'll REQUIRE water cooling. I am saying that it could as long as you can deal with some noise and here are all these other cards dissipating that level of heat with a blower.

Pretty straightforward.

 

[Ieldra]

AMD has listed TFlops for Vega FE. In order to meet that consistently it'll have to stay cool and maintain clocks. They say it won't with a blower cooler and that it'll REQUIRE water cooling. I am saying that it could as long as you can deal with some noise and here are all these other cards dissipating that level of heat with a blower.

Pretty straightforward.

AMD has indeed listed specs for Vega FE, precisely for the reasons outlined above.... Do you think they based the specs on the best cooled version or the worst cooled version ?

Do you think there is no difference between the two?

 

[Magic Hate Ball]

AMD has indeed listed specs for Vega FE, precisely for the reasons outlined above.... Do you think they based the specs on the best cooled version or the worst cooled version ?

Do you think there is no difference between the two?

Assumptions are the mother of all fuck ups. There could be a small clock disparity at stock that could widen with OC headroom. But as I keep fucking repeating, the blower will be able to cool to that level of TFlop performance.

 

[Ieldra]

Assumptions are the mother of all fuck ups. There could be a small clock disparity at stock that could widen with OC headroom. But as I keep fucking repeating, the blower will be able to cool to that level of TFlop performance.

The blower type cooler being able to sustain the 1525mhz needed for the rated 12.5tflops is also an assumption.

If the blower and CLC cards have the same TDP then somethings gotta give... You can't have your cake, eat it too, then expect to have more cake left over

 

[lolfail9001]

AMD has listed TFlops for Vega FE. In order to meet that consistently it'll have to stay cool and maintain clocks.

AMD is under no obligation for their cards to deliver that performance in real life, see Radeon WX7100 and it's faux teraflops, see Macbook Pro Radeons and their faux teraflops.. Do you not know what "up to" means?

For all we know, the golden version is the only one to actually keep that boost clock running kind of consistently, while the blower version runs at 1200Mhz in real life (that, by the way, would explain the underwhelming results leaked benches show).

 

[Magic Hate Ball]

The blower type cooler being able to sustain the 1525mhz needed for the rated 12.5tflops is also an assumption.

If the blower and CLC cards have the same TDP then somethings gotta give... You can't have your cake, eat it too, then expect to have more cake left over

The thing about blowers is you can crank them to 11 if you want. They never ship with anything above 50-60% fan speed, there's lots of airflow available to those with headphones.

 

[Ieldra]

To give a concrete example. I have a windforce 980ti with a stock power limit of 300w which I have modified through bios to 400w.

With my 24/7 1480mh @ 1.187v I peak at around 80% of the power limit so 320w with memory bumped up to 7.9ghz. If I max out my fans (4k RPM) I end up with roughly 40c delta (over ambient). The cooler on the card is rated for 600w by gigabyte (written on the box).

The hotter the core runs, the higher the voltage I need for a given frequency. If I can keep it under 60c I can do 1500mhz with 1.193v, if it goes over 60c it crashes, but will run at 1.21v, until it hits 68c or so, then crash again. Past that temperature and voltage it just goes downhill and won't even do stable 1480mhz at 1. 23v past like 75c.

This is behavior will only become more and pronounced with each node shrink.

If I watercooled my card the delta T would drop to something in the range of 15c, and I would be able to clock higher and with lower voltages.

Now it *is* possible that the Vega FEs are running at 'efficient' clocks/voltage that are relatively low compared to what it is capable of, but it's a 300w card, if that's the 'efficiency profile' then what the hell kind of power requirement will it have when you push it past that comfortable spot on the V/f curve

Edit:

For context consider a ~170w rx480 at ~1300mhz (around 6 tflops) vs 240 1430mhz (around 6.6tflops)

A 10% performance increase with a 40% power increase... At 300w we are talking 400w+ which is nuts. This affects Pascal as well by the way, the main difference being that the equivalent cards draw far less power when they are clocked within that high efficiency window.

1080ti go up to about 300w draw on air before the temperatures catching up to the clock increases, the Pascal cards are just amazing example of this behavior

 

[Singularity_Survivor]

What is the expected release date for Vega?

 

[Ieldra]

What is the expected release date for Vega?

End of July is the launch, availability presumably in August

 

[TaintedSquirrel]

Are there any new AMD GPUs planned for 2018? low/mid Navi maybe?
Vega is going to have a very low shelf-life depending on Volta.

 

[Ieldra]

Are there any new AMD GPUs planned for 2018? low/mid Navi maybe?
Vega is going to have a very low shelf-life depending on Volta.

Vega 11 which should be a lower end Vega part, then Vega 20 should be an HPC oriented double precision chip

 

[razor1]

yep Vega 11 is going to be announced when they launch Vega 10, but won't be out till end of year (rumored). Vega 20 might be later though that is supposed to be 7nm right?

 

[N4CR]

If the blower and CLC cards have the same TDP then somethings gotta give... You can't have your cake, eat it too, then expect to have more cake left over

You are clutching at straws dude.
Even the Flounders Edition coolers could cool OC/sustained boost clocks when you bump the air volume/pressure up.
>Same with the 290X reference edition at 275W TDP stock.
>Same with the 7970 heavily OC'd well beyond 300W use.

Both companies go for slightly quieter running hence the throttling. Nvidia's own thermi caused this knee jerk because it got crucified on blower noise..

Do you think a professional compute card has users that are looking primarily for noise reasons? AMD could just bump the RPM up if they do use a stock blower and call it a day.
The main reason the Fury-X used water is the size of the card. HBM makes for a more compact design which needs water or a nano type config to take advantage of. Clearly not for pure OC (yeah inb4 overclockers dream) or the few extra shaders or whatever, it's like saying a 290 can't clock as high as a 290x because the thermal load and cooler is massively different (it's such a little difference they use the same cooler designs).

But yes if it's a 300W card with a 300W blower cooler, it's not going to OC very well, nor is it likely there is so much headroom left on the table. That is unless AMD is misleading. They've been playing the game different this time around so who knows

yep Vega 11 is going to be announced when they launch Vega 10, but won't be out till end of year (rumored). Vega 20 might be later though that is supposed to be 7nm right?

Edit: Vega 20 is 2018 allegedly. I reckon late q3-4 lmao typical AMD. But it will reduce the nvidia process gap quite a bit. 5nm is already being worked on with IBM at GF so AMD will get the scraps from that later.

Vega 10x2 is just dual Vega 10 at '14nm' or whatever it is this time around. It is rumoured they may share interposer/die level infinity fabric type design to make a true MCM dGPU. This is the one I'm most interested in as if they do this, it's likely a test/proof of concept for Navi and could be what keeps AMD competitive in higher end GPUs (outside of power use). Just like Fury was almost more of a production scaling test than a card launch (IMO). AMD/Raja knows GCN is a bit of a lemon hence Navi getting a full revamp. That said, I wonder if it's too much effort doing this on Vega so they will make a normal dual GPU card. It doesn't make sense they double the product number though...

 

[razor1]

hmm dual GPU graphics cards bad idea lol

 

[Gideon]

hmm dual GPU graphics cards bad idea lol

Well a gpu is mostly a bunch of processors working together. A infinity fabric setup with multiple smaller gpu dies could be interesting. It would be a totally new way of doing things on a video card, but it's always a bit of a gamble doing something for the first time.

 

[razor1]

infinity fabric is just an extension of hyper transport and thus similar to a pci-e bus, nothing more, depending on how many lanes between the two GPU's is what matters. I don't think they will be putting a butt load of lanes between the two GPU's, just doesn't make economic sense right now at least not for a gaming card, professional cards possibly. There needs to be enough bandwidth to pool the memory otherwise it will hit the same problems and mGPU is the only way to go if that happens.

 

[Anarchist4000]

Vega 20 is just dual Vega 10 at '14nm' or whatever it is this time around. It is rumoured they may share interposer/die level infinity fabric type design to make a true MCM dGPU. This is the one I'm most interested in as if they do this, it's likely a test/proof of concept for Navi and could be what keeps AMD competitive in higher end GPUs (outside of power use). Just like Fury was almost more of a production scaling test than a card launch (IMO). AMD/Raja knows GCN is a bit of a lemon hence Navi getting a full revamp. That said, I wonder if it's too much effort doing this on Vega so they will make a normal dual GPU card. It doesn't make sense they double the product number though...

I thought Vega20 was the 7nm design with FP64 support? A dual Vega 11 might make more sense for an economical MCM. Assuming it's around 200mm2, that would be on par with Ryzen for even a 4 way ThreadRipper configuration. What I'm curious about is that Infinity Fabric had two interconnect schemes. The MCM approach like ThreadRipper and the dual socket plus MCM like Naples. It seems likely Vega has a whole lot of PCIE lanes on board. They could use a PLX, but more lanes would make sense for some configurations. SSG could easily use 4/8/16 lanes and SAN controllers can be 16x as well. Having a 32/64 lane chip could be reasonable, although a lot of IO. That leaves open the possibility of a dual Vega10 without a PLX and a whole lot of bandwidth between them. As well as large APUs or SAN configurations. Probably one of AMDs biggest advantages here is HBCC/unified memory on x86 and Infinity on their CPUs as a backplane. It would seem silly not to press that advantage. Nvidia won't be able to match that unless they can convince Intel to give them an x86 license.

Personally I'm of the opinion Vega goes the dual socket Infinity approach and Navi brings the MCM approach. MCM with a NUMA model makes more sense when async compute and graphics with DX12/Vulkan become more commonplace. Mapping multiple queues to multiple hardware units is far simpler than a single queue to multiple units.

 

[Gideon]

infinity fabric is just an extension of hyper transport and thus similar to a pci-e bus, nothing more, depending on how many lanes between the two GPU's is what matters. I don't think they will be putting a butt load of lanes between the two GPU's, just doesn't make economic sense right now at least not for a gaming card, professional cards possibly. There needs to be enough bandwidth to pool the memory otherwise it will hit the same problems and mGPU is the only way to go if that happens.

We cant limit our self in thinking on how cards are done now and that is how it would work in this setup. Perhaps you can have a gpu dedicated to geometry, 1 to physics and so on and have them all on this fabric to communicate the final image. But who knows it could just be the same as gluing two old gpu's together like normal using the fabric to sync it. Will be a bit before we know what they have up their sleeve.

 

[razor1]

We cant limit our self in thinking on how cards are done now and that is how it would work in this setup. Perhaps you can have a gpu dedicated to geometry, 1 to physics and so on and have them all on this fabric to communicate the final image. But who knows it could just be the same as gluing two old gpu's together like normal using the fabric to sync it. Will be a bit before we know what they have up their sleeve.

That is only if there is enough bandwidth to pool the memory. There is a reason why nvlink hasn't been used on lower end quadros. Cost is prohibitive. Infinity fabric isn't much different than Intel's fabric either. So over all, the same limiting factors will be there to scale with price of the final product.

 

[N4CR]

I thought Vega20 was the 7nm design with FP64 support? A dual Vega 11 might make more sense for an economical MCM. Assuming it's around 200mm2, that would be on par with Ryzen for even a 4 way ThreadRipper configuration. What I'm curious about is that Infinity Fabric had two interconnect schemes. The MCM approach like ThreadRipper and the dual socket plus MCM like Naples. It seems likely Vega has a whole lot of PCIE lanes on board. They could use a PLX, but more lanes would make sense for some configurations. SSG could easily use 4/8/16 lanes and SAN controllers can be 16x as well. Having a 32/64 lane chip could be reasonable, although a lot of IO. That leaves open the possibility of a dual Vega10 without a PLX and a whole lot of bandwidth between them. As well as large APUs or SAN configurations. Probably one of AMDs biggest advantages here is HBCC/unified memory on x86 and Infinity on their CPUs as a backplane. It would seem silly not to press that advantage. Nvidia won't be able to match that unless they can convince Intel to give them an x86 license.

Personally I'm of the opinion Vega goes the dual socket Infinity approach and Navi brings the MCM approach. MCM with a NUMA model makes more sense when async compute and graphics with DX12/Vulkan become more commonplace. Mapping multiple queues to multiple hardware units is far simpler than a single queue to multiple units.

Sorry, I should not morning post. Got mixed with Vega 10x2 >_< You're right re:Vega20.
WIth infinity fabric/HBCC/unified etc approach they have their own solution for die-die but the rest not sure how. You're totally right they need pcie lanes for SSG already so this might be the basis for it all... how do those lanes work. Wonder if can dig up a shot of the current SSG cards (unlikely)?

Thing is if they can make infinity fabric work on CPUs, then logically it should scale to GPU applications well too. I'd bet they'll do a TR/Infinity fabric style solution as they know it has real world success. It solves the memory access problem too - AMD builds a large die, slightly slower overall vs nvidia because automated routing + process disadvantage (for now) and more infinity fabric/pcie circuitry needed... But the beauty of this is more linear scaling. So Nvidia spends megabucks on R&D to reinvent the wheel every two years, while AMD just slowly process bumps and scales an existing design (like they do currently..) with more time between node changes. Wonder with navi if they'll try get as close as possible to a SOC mGPU MCM arrangement. Shit is going to get interesting in next few years and I can't wait..
As I've maintained for quite some time now, the future of GPUs is going to be mGPU, similar to CPUs. It will get to a point where new technology at substrate level is needed to progress much further, or we make mGPU work at hardware level for big gains. Eventually one will be cheaper than the other.

 

[noko]

WTF is FE Gold? Thanks

 

[Ieldra]

WTF is FE Gold? Thanks

The watercooled Vega card

 

[JustReason]

I am very interested in the infinity fabric. I would assume it will use pci-e x16 like the cpu, however what will be the speed? Will it be tied to ram speed? Not sure how much correlates with the CPU.

 

[Simplyfun]

I think AMD is about to make mGPU a "thing" that is invisible to whatever is making the calls. Given their past history, I think it will work, but ver 1 will have quirks we don't like.

 

[N4CR]

I am very interested in the infinity fabric. I would assume it will use pci-e x16 like the cpu, however what will be the speed? Will it be tied to ram speed? Not sure how much correlates with the CPU.

Ram speed is where it gets interesting with the differences of HBM design. This is why everyones unsure about it. No doubt if they are/have done it, will be some logical way
But that also leads to Simplyfuns' post which is probably bang on the money.

I think AMD is about to make mGPU a "thing" that is invisible to whatever is making the calls. Given their past history, I think it will work, but ver 1 will have quirks we don't like.

This sounds about right. E.g. 50% scaling only, or some sort of frame time/pacing issue.

This is RTGs only real ticket to high end competition for next few years though and could be similarly disruptive as Naples could potentially be.

E.g. we may see as following
Vega 10: OCd 1080 to Ti or faster in some games.
Gv104 ~Ti speed range
Nvidia is still tied with AMD for overall performance until they get Gv102 out later. Of course the power use shilling will be in overdrive during this time.
Gv102 launches 2018, usual +50-60% with a much bigger die than Gp102, clock bump, some process refinements as usual, Titan, ten gorillion dollars.
AMD counters with Vega10x2, matches or takes crown back (I doubt it will be ready this year but please prove me wrong AMD) costs a little less.
>PowerUseShillingIntensifies.webm
late 2018 dec Nvidia releases Volta Ti 'Titan watered down edition because we can't sell any more for HPC first'
AMD releases Navi, drawing nearly even on process this time (7nm gloflo)
Nvidia releases Ti 'watered down edition Titan with a cost cut and less ram' and some 'even less watered down edition' Titan with more ram than the last one and a clock bump
MCM process repeats (note how they expect Vega10x2 much sooner than the usual year between single and dual GPU launches...) as they now make MCM/mGPU cards part of key R&D
High end is now muddied. AMD offers higher performance and slightly higher power usage.
Epeen upgraditis types now confused 'cus muh Nvidia is not so stronk all of a sudden and 'AMD SUXX LOLOLOL' can't be so easily played when they're competitive or better on performance.


Next few years are going to be a great time for GPU consumers... 4k, 8k, VR, mGPU/MCM bring it on!