Another 6 months before VEGA?

[NKD]

Not really, they are showing their best they can do in particular tests, clocks could have been lower, but they also down clocked Intel's chip to compensate. When AMD was able to get clocks up, they didn't down clock Intel's chip anymore.

And we still don't know the end results cause those tests could be very specific to the best Ryzen can do in particular applications.

I am talking about clocks there. They never showed clocks I think they held back the real capabilities of clock speeds on ryzen. They started official demos at 3.2 and now we know of chips running at 3.6ghz stock.

 

[razor1]

yeah and when they showed their 3.0 I think, not 3.2 ghz, with blender they showed it against a 3.0 ghz Broadwell.


The second time they showed it Ryzen was at 3.4 ghz without boost, and the Intel chip was stock 3.2 with boost (boosting up to 3.5 ghz).

 

[NKD]

and now we have ryzen running at 3.6 stock and 3.9 boost. So they pretty much never let the tru clock speeds out.

 

[razor1]

and now we have ryzen running at 3.6 stock and 3.9 boost. So they pretty much never let the tru clock speeds out.

where did ya see that?

The 3.4 was at CES.

Either way its not that big of deal, They are showing the best they can at the time they had in their hand.

 

[Anarchist4000]

The Titan X (TXP) is almost, and sometimes is, twice as fast as AMD's best GPU (Fury X). That would be a hell of a jump to match the TXP never mind beat it.

Polaris 10 and GP104 are roughly the same size and performance. The Vega die and GP102 are about the same size, and that doesn't account for a likely smaller memory controller and other enhancements. If it is a new architecture I'd certainly expect it to outperform TXP. Hype isn't always good, but it's hard to imagine performance not being comparable unless they really laid an egg with the design.

Core counts are pretty much know now. Also going by Polaris's die size, Vega's die size adds up , we even discussed this prior to seeing Vega's die, it was going to be bigger than GP102 is exactly what I stated with 4096 cores..... the GP102 has a die size of 471 mm^2, so Vega coming just north of that, was to be expected with a core count of 4096, which is the die that Raja showed Linus and others.

Doesn't look like it's much bigger than GP102 at all. Polaris 10 is roughly half the die size of GP102 with 2304 cores. Double that and you're looking at 4608 cores and doesn't account for components that should improve the scaling. Fairly sure we've been over that before. Larger caches might account for the increase, but that's a good deal of cache that got added. The design I was speculating about may have actually shrunk the core sizes, and Carrizo for example was making use of high density logic for many parts. You also said AMD doesn't have the funds to make any major changes to the architecture, yet here we are with what looks like a ground up new architecture with some radical new features. Features that seem likely to get standardized on the next major DX release as devs have been writing papers on techniques.

Cores have been speculated on, but nobody has solid counts. Nor do we know what is actually inside the cores. We just know the die got substantially larger when it should have scaled well in addition to AMD claiming higher IPC and clocks. I'd have to assume Raja's percentage tweet was in response to that 40-70% less MRF accesses not being very productive I pulled out of that paper. Or he's just taking everyone for a ride and hopefully has a nice one after that stock bump. Regardless, the MRF stat has some obvious implications for core design that nobody is talking about. Definitely makes things interesting and that core count gets a bit misleading.

They are, I have no doubt they are, cause Vega doesn't have a full set of DP units either, so there might be a Vega with all of its DP units, which will account for a chip to go against the GP100, but that is only if it can muster up performance to it.

The FP64 is probably on one of the APUs and not the discrete boards. Where it has better access to system memory for the larger supercomputer workloads. Most supercomputers are CPU driven anyways, GPUs being a bit of a niche in that segment. It matches up with their high performance APU in server market strategy much better. Especially since APUs and CPUs are basically the same form factor. That strategy likely puts it in a higher performance tier than P100. Memory bottlenecks aren't uncommon in a supercomputer and that would easily surpass the bandwidth and latency NVLink could provide for workloads likely to use FP64. Even without higher peak compute performance. No reason to believe it wouldn't be faster either as AMD traditionally packs in more compute performance than Nvidia.

I'm still not ruling out Vega 10 having Zen cores and being an APU despite being on a discrete card.

 

[noko]

I will be interested in what options are opened up to AIBs - as in if some will put in extra memory either SSD style Flash or DDR memory connected to the card ? Will the drivers automatically map out the available faster ram available? Stuff that is truly usable and works right off the bat. Then comes into play CFX tech - with the increase performance how is AMD not going to bottleneck just using XDMA and not having CFX bridges?

VR performance at this stage is also key and should be in any real review of Vega. Has Vega improve VR performance and is on par or better then Nvidia performance? Then getting to know when the newer stuff will actually get used would be nice, what development is going on by what developers to use it?

Clearly at this stage, Vega needs to basically smash Pascal in most respects but mainly performance overall. Otherwise AMD better be ready to sell some rather cheaper but good perf/$ cards.

 

[Shintai]

I really don't think amd is demoing a full vega chip. I think they are being sneaky here, everything is too tight lipped.

Sure, sure, it was what, a 128SP part and ran at 200Mhz? You search for excuses.

Polaris was also to be entirely different than it became, because it could by no means be what it was.

 

[Shintai]

I will be interested in what options are opened up to AIBs - as in if some will put in extra memory either SSD style Flash or DDR memory connected to the card ?

Sure, and sell them for 3000$. Also I doubt it matter a single bit for gaming.

 

[SixFootDuo]

It's very possible they are making last minute improvements on both the GPU and CPU and that's a good thing for the potential customer out there. I personally do not waiting an additional 90 days or whatever the case may be to give them time to pull out all the performance they can.


I know initial impressions from people that had a close look or had the aptitude to truly understand the numbers, capabilities and performance weren't overly impressed. That was for both the GPU and CPU. At least that was my take on it from everything I've read. Sorry, the Quake Vulcan demo was still pretty much below the 1080 and that was with them using Vulcan that's much faster than DX11 / 12 .... right?


Let's hope it all gets improved at the 11th hour before release.

 

[CSI_PC]

I really don't think amd is demoing a full vega chip. I think they are being sneaky here, everything is too tight lipped.

It is the full chip according to Raja in an interview but would not say which model, but it could be argued that is not being utlised to its full capability in terms of its integrated functions.
Some of these IMO are at least 1 generation too early (they will not start to have any impact for at least another 12-14months) while others are yet to be seen fully functioning.
Still to be though if there is a cut core and a full version that will be used in the Pro market rather than consumer.

How much more is to come from the chip is to be seen, but if the argument is that it is still early product with much to do on the software-firmware side, then the worry for me is this could be really close to a June product if everything goes well.
Against that though (from a later product and also much more expected from software-firmware side) was it being very stable while it was shown with the next version of Star Wars Battlefront and also Doom Vulkan.
Remember how much excitement there was about Async Compute with the architecture of GCN, but in reality it provides at most on PC around 10% (still a nice figure) and usually anywhere 2% to 8% in various games.
Cheers

 

[CSI_PC]

LOL they did kinda with ryzen!

Different teams and different approach even with marketing and product presentation and what they say is achievable.
Cheers

 

[Dayaks]

Polaris 10 and GP104 are roughly the same size and performance. The Vega die and GP102 are about the same size, and that doesn't account for a likely smaller memory controller and other enhancements. If it is a new architecture I'd certainly expect it to outperform TXP. Hype isn't always good, but it's hard to imagine performance not being comparable unless they really laid an egg with the design.


Doesn't look like it's much bigger than GP102 at all. Polaris 10 is roughly half the die size of GP102 with 2304 cores. Double that and you're looking at 4608 cores and doesn't account for components that should improve the scaling. Fairly sure we've been over that before. Larger caches might account for the increase, but that's a good deal of cache that got added. The design I was speculating about may have actually shrunk the core sizes, and Carrizo for example was making use of high density logic for many parts. You also said AMD doesn't have the funds to make any major changes to the architecture, yet here we are with what looks like a ground up new architecture with some radical new features. Features that seem likely to get standardized on the next major DX release as devs have been writing papers on techniques.

Cores have been speculated on, but nobody has solid counts. Nor do we know what is actually inside the cores. We just know the die got substantially larger when it should have scaled well in addition to AMD claiming higher IPC and clocks. I'd have to assume Raja's percentage tweet was in response to that 40-70% less MRF accesses not being very productive I pulled out of that paper. Or he's just taking everyone for a ride and hopefully has a nice one after that stock bump. Regardless, the MRF stat has some obvious implications for core design that nobody is talking about. Definitely makes things interesting and that core count gets a bit misleading.


The FP64 is probably on one of the APUs and not the discrete boards. Where it has better access to system memory for the larger supercomputer workloads. Most supercomputers are CPU driven anyways, GPUs being a bit of a niche in that segment. It matches up with their high performance APU in server market strategy much better. Especially since APUs and CPUs are basically the same form factor. That strategy likely puts it in a higher performance tier than P100. Memory bottlenecks aren't uncommon in a supercomputer and that would easily surpass the bandwidth and latency NVLink could provide for workloads likely to use FP64. Even without higher peak compute performance. No reason to believe it wouldn't be faster either as AMD traditionally packs in more compute performance than Nvidia.

I'm still not ruling out Vega 10 having Zen cores and being an APU despite being on a discrete card.

That's true, that's probably a better data point. 1060 is 200mm^2 and rx480 is 232mm^2. With a few slight improvements it's not out of the realm of possibility AMD could lead for once.

 

[razor1]

Polaris 10 and GP104 are roughly the same size and performance. The Vega die and GP102 are about the same size, and that doesn't account for a likely smaller memory controller and other enhancements. If it is a new architecture I'd certainly expect it to outperform TXP. Hype isn't always good, but it's hard to imagine performance not being comparable unless they really laid an egg with the design.

P10 and GP104 aren't in the same performance tier, you are hyping it right there, they are in the same power consumption bracket though. Think you mean GP106, which has a smaller die size than P10

Doesn't look like it's much bigger than GP102 at all. Polaris 10 is roughly half the die size of GP102 with 2304 cores. Double that and you're looking at 4608 cores and doesn't account for components that should improve the scaling. Fairly sure we've been over that before. Larger caches might account for the increase, but that's a good deal of cache that got added. The design I was speculating about may have actually shrunk the core sizes, and Carrizo for example was making use of high density logic for many parts. You also said AMD doesn't have the funds to make any major changes to the architecture, yet here we are with what looks like a ground up new architecture with some radical new features. Features that seem likely to get standardized on the next major DX release as devs have been writing papers on techniques.

And in GPU design, die size always doubles if you double ALU's? There are redundant areas and using HBM reduces the bus die needed. You mentioned it in the previous quote yet here you didn't take that into account? They don't have the funds that is why its coming out later, not only that, I don't expect it to be much different than Polaris outside of its primitive shaders, which seem to be everything AMD is hyping for its new memory and tiled rasterizer. What does that tell us? Yeah they look to have gotten hurt because of lack of funds. There is no way AMD wasn't able to do these things one year later than their competition, while their competition didn't need a specifically new pipeline which dev's have to account for to eek out every bit of performance.

This is what happens when AMD hypes the shit out there "new features" in Polaris, and it comes out as a dud and only competes with last gen Maxwell products in performance/watt, you expect Vega to do something differently in the 6 months of time they had extra to get 50% more perf/watt? You think AMD has some magic fairy dust that they can sprinkle when Raja already stated they have less engineering talent?

Cores have been speculated on, but nobody has solid counts. Nor do we know what is actually inside the cores. We just know the die got substantially larger when it should have scaled well in addition to AMD claiming higher IPC and clocks. I'd have to assume Raja's percentage tweet was in response to that 40-70% less MRF accesses not being very productive I pulled out of that paper. Or he's just taking everyone for a ride and hopefully has a nice one after that stock bump. Regardless, the MRF stat has some obvious implications for core design that nobody is talking about. Definitely makes things interesting and that core count gets a bit misleading.

You maybe guessing, I am not lol. Look IPC BS we know that, GPU's aren't based on IPC, they are based on through put come on man, you have been around the block on this one. These aren't CPU's.

Clocks, err, we kinda know where the boost clocks will end up too, at least the range, clocks are just a bit higher then Polaris, 10%-15% higher, that is it, this is probably due to less transistor density nothing more. We also know the power consumption something AMD stated <300, which giving AMD benefit of the doubt after its recent debacle with P10, lets say its at 275 watts, which keeps it in line with previous high end cards from them.

The FP64 is probably on one of the APUs and not the discrete boards. Where it has better access to system memory for the larger supercomputer workloads. Most supercomputers are CPU driven anyways, GPUs being a bit of a niche in that segment. It matches up with their high performance APU in server market strategy much better. Especially since APUs and CPUs are basically the same form factor. That strategy likely puts it in a higher performance tier than P100. Memory bottlenecks aren't uncommon in a supercomputer and that would easily surpass the bandwidth and latency NVLink could provide for workloads likely to use FP64. Even without higher peak compute performance. No reason to believe it wouldn't be faster either as AMD traditionally packs in more compute performance than Nvidia.

I'm still not ruling out Vega 10 having Zen cores and being an APU despite being on a discrete card.

At this point you can start ruling it out, at least it won't be coming till much much later.

 

[Ieldra]

P10 is 15% larger than GP106 and draws 33% more power.

This isn't even accounting for feature size advantage gained by using Glofo 14nm

 

[CSI_PC]

Polaris 10 and GP104 are roughly the same size and performance. The Vega die and GP102 are about the same size, and that doesn't account for a likely smaller memory controller and other enhancements. If it is a new architecture I'd certainly expect it to outperform TXP. Hype isn't always good, but it's hard to imagine performance not being comparable unless they really laid an egg with the design.

The closest comparison you can get with Vega is the MI25 against the P40 (full core GP102).
Vega is mentioned to be around 520mm2, the P40 is 471mm2, with the MI25 having about 0.5 FP32 TFLOPS theoretically more than the P40, and of course the MI25 has FP16.
The TDP for the P40 is 250W.
But importantly the P40 is using traditional GDDR5 7.2gbps 384bit bus memory 24GB, and this has an impact on the whole design/TDP-TBP.

I agree Vega has features missing from Pascal, but only a few of those can be brought to effect before Volta has a more complete lineup in 2018 (products up and down the whole tier range and sectors).
Personally I think they focused on some too far forward thinking tech within Vega, but for its size its performance should be higher than what we have currently seen in Doom Vulkan so there is still a bit to come but just how much is to be seen; like I said some of its advantages will not start (so not established) to be truly utilised for at least 12-14 months (that will also require Polaris to be replaced to make it economical for devs on the retail side) and it is competing against a complete product lined up Volta by then.
Still not convinced though the consumer retail Vega will match a Pascal Titan in games optimised for both manufacturers, but that is a tricky discussion to quantify regarding games and their future optimisations/porting/custom IHV implemented tech/etc, although AMD controlling key home consoles helps.
It will be interesting to see just how well it does perform in various games compared to the 1080 and Titan.
Cheers

 

[Ieldra]

The closest comparison you can get with Vega is the MI25 against the P40 (full core GP102).
Vega is mentioned to be around 520mm2, the P40 is 471mm2, with the MI25 having about 0.5 FP32 TFLOPS theoretically more than the P40, and of course the MI25 has FP16.
The TDP for the P40 is 250W.
But importantly the P40 is using traditional GDDR5 7.2gbps 384bit bus memory 24GB, and this has an impact on the whole design/TDP-TBP.

I agree Vega has features missing from Pascal, but only a few of those can be brought to effect before Volta has a more complete lineup in 2018 (products up and down the whole tier range and sectors).
Personally I think they focused on some too far forward thinking tech within Vega, but for its size its performance should be higher than what we have currently seen in Doom Vulkan so there is still a bit to come but just how much is to be seen; like I said some of its advantages will not start (so not established) to be truly utilised for at least 12-14 months (that will also require Polaris to be replaced to make it economical for devs on the retail side) and it is competing against a complete product lined up Volta by then.
Still not convinced though the consumer retail Vega will match a Pascal Titan in games optimised for both manufacturers, but that is a tricky discussion to quantify regarding games and their future optimisations/porting/custom IHV implemented tech/etc, although AMD controlling key home consoles helps.
It will be interesting to see just how well it does perform in various games compared to the 1080 and Titan.
Cheers

It could also be that Vega dedicates less die area to the EUs in favor of better distributed resources within each shader engine.

I know we're all assuming 4096 ALU but was that confirmed in the end? 4608 ALU could reasonably be organized into six SEs

 

[Anarchist4000]

P10 and GP104 aren't in the same performance tier, you are hyping it right there, they are in the same power consumption bracket though. Think you mean GP106, which has a smaller die size than P10

Sorry, GP106.

And in GPU design, die size always doubles if you double ALU's? There are redundant areas and using HBM reduces the bus die needed. You mentioned it in the previous quote yet here you didn't take that into account? They don't have the funds that is why its coming out later, not only that, I don't expect it to be much different than Polaris outside of its primitive shaders, which seem to be everything AMD is hyping for its new memory and tiled rasterizer. What does that tell us? Yeah they look to have gotten hurt because of lack of funds. There is no way AMD wasn't able to do these things one year later than their competition, while their competition didn't need a specifically new pipeline which dev's have to account for to eek out every bit of performance.

Which is why I mentioned parts that won't scale in the response. If this is one year late, then they apparently brought next gen DX features years earlier than the competition with limited funding. Primitive shaders are nice, but not much more than a specialized compute shader from what I've seen. Cool feature, but far from what I'd consider the architecture to be providing. The CU makeup and HBCC seem far more influential. Considering they're sharing the fabric with Zen, it's not surprising they're releasing together. Outside of AMD and some devs I doubt anyone knows how much work is required to optimize for Vega. Most of the major gains look to be automatic.

This is what happens when AMD hypes the shit out there "new features" in Polaris, and it comes out as a dud and only competes with last gen Maxwell products in performance/watt, you expect Vega to do something differently in the 6 months of time they had extra to get 50% more perf/watt? You think AMD has some magic fairy dust that they can sprinkle when Raja already stated they have less engineering talent?

Considering they're different architectures yeah. P10 is a pretty good card and seems to be selling well enough.

You maybe guessing, I am not lol. Look IPC BS we know that, GPU's aren't based on IPC, they are based on through put come on man, you have been around the block on this one. These aren't CPU's.

So you know for certain that no pipelining is occuring within Vega? AMD said higher clocks and IPC. While IPC doesn't normally apply to GPUs(1 instruction per clock everywhere, maybe 2 for packed math) that doesn't mean the design wasn't changed where it's relevant. Different clock domains and/or hyperthreading would accomplish that. So IPC may very well be relevant. If they designed around that operand cache, they could be executing 3-4 waves on a single core without packed math. Packed may in fact double that to 6-8 waves on a single core if they're doing what I think.

P10 is 15% larger than GP106 and draws 33% more power.

What about after equalizing performance. Take a relatively well optimized game like doom, equalize performance, then what kind of power draw do you have. Even 5% higher performance could translate into significantly higher power draw. In a properly optimized game, I'm guessing the difference is minimal.

Vega is mentioned to be around 520mm2, the P40 is 471mm2

I'm not entirely sure it's actually larger than 500mm2. There were the pictures, but I've seen/heard everything from slightly above and below the 500mm2 figure. That still leaves it no more than 10-15%, if that.

 

[razor1]

Which is why I mentioned parts that won't scale in the response. If this is one year late, then they apparently brought next gen DX features years earlier than the competition with limited funding. Primitive shaders are nice, but not much more than a specialized compute shader from what I've seen. Cool feature, but far from what I'd consider the architecture to be providing. The CU makeup and HBCC seem far more influential. Considering they're sharing the fabric with Zen, it's not surprising they're releasing together. Outside of AMD and some devs I doubt anyone knows how much work is required to optimize for Vega. Most of the major gains look to be automatic.

Yes and if you break down gpu component sizes based on polaris, it ends up close to what Vega is at size wise.

Who cares about next gen DX features when the next DX API is a ways off? You think developers will make use of 27% of the market? Yeah consoles are there, but those games have to be backward compatible to older tech, so that too is kinda not going to be on the forefront of development. You say its a specialized compute unit, I don't think its specialized at all, I think they made tweaks to something they knew they had problems with the GCN architecture which couldn't be solved in a way that would have been automatic. What stopped AMD from implementing their features they have been talking about where nV has them already without developer intervention? Best case numbers AMD gave for Vega in those areas would equalize their stance with nV, which they have stated over multiple times those performance metrics are app dependent.

So now you are saying Vega will only show its true merit with Zen because of its HBCC? Sounds a bit underwhelming when you put it that way. I don't see that happening.

Considering they're different architectures yeah. P10 is a pretty good card and seems to be selling well enough.

Is it really selling well? Where did you see that, it sold less % wise when looking at its first full Q sales figures where AMD is till flogging old tech at bargin bin prices. Fury X, r3xx lines still are being sold by OEM's and system builders. That doesn't sound like selling well.

So you know for certain that no pipelining is occuring within Vega? AMD said higher clocks and IPC. While IPC doesn't normally apply to GPUs(1 instruction per clock everywhere, maybe 2 for packed math) that doesn't mean the design wasn't changed where it's relevant. Different clock domains and/or hyperthreading would accomplish that. So IPC may very well be relevant. If they designed around that operand cache, they could be executing 3-4 waves on a single core without packed math. Packed may in fact double that to 6-8 waves on a single core if they're doing what I think.

IPC the only thing will be higher is for half rate, since it wasn't there before.

Packed math, nah, we won't see that being influential in Vega's life time.

And all of this will be done through developer intervention and this is why it won't show up in Vega's life time, nor will be even conceivable in the short term to help sales.

What about after equalizing performance. Take a relatively well optimized game like doom, equalize performance, then what kind of power draw do you have. Even 5% higher performance could translate into significantly higher power draw. In a properly optimized game, I'm guessing the difference is minimal.

So you are saying developer's will optimize for AMD Vega only? Were AMD has how much marketshare again? And when is Vega coming to consoles? All of these things don't add up, and if that is what AMD is betting on, they are screwing themselves over, again, because of improper timing.

To sum it up, everything AMD has showed so far, they are betting console Vega will help drive PC development, guess what, console Vega is and end of this year product, which doesn't bode well for any of their game dev initiatives.

 

[CSI_PC]

I'm not entirely sure it's actually larger than 500mm2. There were the pictures, but I've seen/heard everything from slightly above and below the 500mm2 figure. That still leaves it no more than 10-15%, if that.

WhyCry has managed to be pretty accurate from photos so far on all previous dGPUs, he suggests 520mm2 (lowest estimation), some other rumours have it even larger.
Lowest mentioned I heard is 500mm2 rumour quite awhile ago, but I would go by WhyCry who has been near spot on so far.
Still the best comparison is MI25 vs P40 with the factors I mentioned earlier.
Cheers

 

[Shintai]

Sorry, GP106.

No they are not. Polaris 10 needs 37% more power, 33% more bandwidth and 30% more transistors to deliver the same. If you isolate the GPUs the power difference will only increase. Hence why Pascal s lead only increases in higher SKUs.

What about after equalizing performance. Take a relatively well optimized game like doom, equalize performance, then what kind of power draw do you have. Even 5% higher performance could translate into significantly higher power draw. In a properly optimized game, I'm guessing the difference is minimal.

It only get worse for Polaris 10 then. That you even have to mention Doom says it all. What else, Hitman?

 

[razor1]

No they are not. Polaris 10 needs 37% more power, 33% more bandwidth and 30% more transistors to deliver the same. If you isolate the GPUs the power difference will only increase. Hence why Pascal s lead only increases in higher SKUs.

Going by unit count and specs, just shows us how far AMD has fallen behind in the past 2 generations, now even possibly even three gens, ever since AMD started cutting R&D cutting on the GPU side of things.

I don't even see why people try to compare polaris to pascal its not even a fair fight.........

All this hub bub of AMD's long drawn out process of building up hype for months before launch, never seems to work out well for them. For them to set expectations, then wait for months on end, and actually at this time they didn't even set up expectations, outside of Doom Vulkan and Battlefront, which aren't mind blowing examples for going up against pascal with a 500mm2 + chip w/HBM2 that seems to be around nV's 471mm2 chip. So those are the expectations I will hold AMD on, just wondering if its going to be another p11 perf/watt showing, for now I'll give AMD benefit of the doubt, cause everything they have stated so far about features haven't given me anything useful in terms of short term AMD needs.

 

[Shintai]

Going by unit count and specs, just shows us how far AMD has fallen behind in the past 2 generations, now even possibly even three gens, ever since AMD started cutting R&D cutting on the GPU side of things.

I don't even see why people try to compare polaris to pascal its not even a fair fight.........

All this hub bub of AMD's long drawn out process of building up hype for months before launch, never seems to work out well for them. For them to set expectations, then wait for months on end, and actually at this time they didn't even set up expectations, outside of Doom Vulkan and Battlefront, which aren't mind blowing examples for going up against pascal with a 500mm2 + chip w/HBM2 that seems to be around nV's 471mm2 chip. So those are the expectations I will hold AMD on, just wondering if its going to be another p11 perf/watt showing, for now I'll give AMD benefit of the doubt, cause everything they have stated so far about features haven't given me anything useful in terms of short term AMD needs.

With Volta the perf/watt delta between may hit as much as 2½x.

Where are all the Polaris 10 and 11 gaming laptops? Oh right, its all 1080, 1070, 1060, 1050. Polaris pretty much only went into the MacBook that flopped.

With Vega 10 its clear RTG choose one single area to focus and its HPC. Even the "7nm" Vega 20 is the same 4096SP, it just got FP64 and a few other things added like GMI and twice the HBM2 stacks.

 

[CSI_PC]

It could also be that Vega dedicates less die area to the EUs in favor of better distributed resources within each shader engine.

I know we're all assuming 4096 ALU but was that confirmed in the end? 4608 ALU could reasonably be organized into six SEs

I originally thought it having up to 25% more ALUs (one more SE), unfortunately I would need to find it as a journalist mentioned even though they can now go above 4 SE they currently have not for this Vega *shrug*.

Talking about the SE, is this how they have managed to increase the FP32 TFLOPs for Scorpio console as recently Microsoft inferred it was still a Polaris tech, Scorpio SoC is quite a bit higher than a dGPU 480 and would be surprised if that was all clocks; compare that to PS4 Pro that is lower than the dGPU 480 but also Polaris design.

Cheers

 

[NKD]

It could also be that Vega dedicates less die area to the EUs in favor of better distributed resources within each shader engine.

I know we're all assuming 4096 ALU but was that confirmed in the end? 4608 ALU could reasonably be organized into six SEs

I have been thinking this too. I am thinking they are demoing may be not the highest end chip. Reports have been there would be a line up of these cards so I am suspecting they can possibly have top end card with 4608

 

[CSI_PC]

I have been thinking this too. I am thinking they are demoing may be not the highest end chip. Reports have been there would be a line up of these cards so I am suspecting they can possibly have top end card with 4608

Well a few of us was expecting you would have a cut core for retail consumer and the worst yield more expensive full core as the pro MI25 model due to going large die, similar to what Nvidia did with P40 and the Titan.
If that is the case you would still end up probably with a 4096 core as the consumer or baby Vega, but with AMD and their positioning of Vega products it is not clear.

Cheers

 

[razor1]

Well a few of us was expecting you would have a cut core for retail consumer and the worst yield more expensive full core as the pro MI25 model due to going large die, similar to what Nvidia did with P40 and the Titan.
If that is the case you would still end up probably with a 4096 core as the consumer or baby Vega.

Cheers

nV and AMD in the past have designed their cores differently, where nV didn't have redundancy in units and AMD did because they knew the possibility of having bad areas in the dies. Not sure if that is the case here cause it has more to do with node maturity if I remember correctly.

Having said that, just looking at the instinct card numbers everything they have stated about flops and what not, everything points to a full chip having 4096 cores with an increased clock speed that is boosting up to 1500mhz. And as you stated that will be the full Vega, why would they put in a cut down version of Vega into a what 5k + card and put in the full version of Vega into a 600 buck card? Doesn't make much sense do it, not only that we have never seen that being done by either company when it comes to gaming GPU's vs professional GPU's because a full die of this size will always have less turn out per wafer so demand for the gaming market won't be easily filled but for a market such as deep learning or HPC, demand is much less and less quantity higher price make sense.

And also a couple of reviewers/journalists have stated 4096 units is the full Vega, so unless they got their info wrong, or instinct cards aren't full chips, that is the only way Vega is going to be different. I don't see so many sources being wrong.

 

[CSI_PC]

nV and AMD in the past have designed their cores differently, where nV didn't have redundancy in units and AMD did because they knew the possibility of having bad areas in the dies. Not sure if that is the case here cause it has more to do with node maturity if I remember correctly..

I thought it was the other way round where Nvidia had greater flexibility in what they could disable in the past (especially since Maxwell), but I thought AMD started catch up since Fiji.

Crude example, for the China market did Sapphire re-enable the 460 to full CU, with der8auer showing he could do the same as well with custom BIOS?
Cheers

 

[razor1]

I thought it was the other way round where Nvidia had greater flexibility in what they could disable in the past (especially since Maxwell), but I thought AMD started catch up since Fiji.
Cheers

Hmm no this has been like this for many years now, since the 68xx and ATi x8xx lines. The latest gens I can think of nV was still doing this was with Fermi. Not sure after that though.

Edit, this was because nV was maxing out the die size pretty much ever gen till Maxwell, so they didn't have the luxury of redundancy.

 

[CSI_PC]

Hmm no this has been like this for many years now, since the 68xx and ATi x8xx lines. The latest gens I can think of nV was still doing this was with Fermi. Not sure after that though.

Edit, this was because nV was maxing out the die size pretty much ever gen till Maxwell, so they didn't have the luxury of redundancy.

Why I mentioned Maxwell
I think we are a bit at cross wires, as I am coming from both die harvesting and also more advanced flexibility for disabling SM/CU/cores-ALU/etc along with yields and costs.
Cheers

 

[razor1]

Why I mentioned Maxwell
Cheers

Ah ok I see what you are saying read what ya typed wrong lol.

 

[CSI_PC]

Ah ok I see what you are saying read what ya typed wrong lol.

Late here so I blame my wording
Cheers

 

[noko]

So after all this discussion . . . Folks here still don't have a clue where Vega will fall on performance compared to a 1080/TitanXP.

 

[Algrim]

I have been thinking this too. I am thinking they are demoing may be not the highest end chip. Reports have been there would be a line up of these cards so I am suspecting they can possibly have top end card with 4608

They called it "Flagship Vega". English language, though quite flexible, still doesn't allow you to attach the flagship moniker to a tugboat.

 

[noko]

They called it "Flagship Vega". English language, though quite flexible, still doesn't allow you to attach the flagship moniker to a tugboat.

Hey, in America anything is possible:
http://www.timesunion.com/local/article/Flagship-tugboat-marks-110-years-1422242.php

 

[Anarchist4000]

No they are not. Polaris 10 needs 37% more power, 33% more bandwidth and 30% more transistors to deliver the same. If you isolate the GPUs the power difference will only increase. Hence why Pascal s lead only increases in higher SKUs.

It only get worse for Polaris 10 then. That you even have to mention Doom says it all. What else, Hitman?

You don't think benchmarks on well optimized games using the latest APIs are valid? Sticking to older games that favor Nvidia the most then? A 480 and 1060 are pretty much equal in terms of performance, with 480 favored on newer games. Those power and bandwidth figures only occur with poorly optimized deferred renderers. So gamers get higher fps at the expense of additional input lag. Why the deferred renderers have been sucking with VR.

So you are saying developer's will optimize for AMD Vega only? Were AMD has how much marketshare again? And when is Vega coming to consoles? All of these things don't add up, and if that is what AMD is betting on, they are screwing themselves over, again, because of improper timing.

Where did I say that? In the case of Doom they optimized for both, and the result is Nvidia pundits seem to think it's invalid because AMD is winning. Or is it still invalid because it uses intrinsics that will be included with SM6?

So now you are saying Vega will only show its true merit with Zen because of its HBCC? Sounds a bit underwhelming when you put it that way. I don't see that happening.

No, only that they share the same technology. The APUs should be really interesting though.

Who cares about next gen DX features when the next DX API is a ways off? You think developers will make use of 27% of the market? Yeah consoles are there, but those games have to be backward compatible to older tech, so that too is kinda not going to be on the forefront of development. You say its a specialized compute unit, I don't think its specialized at all, I think they made tweaks to something they knew they had problems with the GCN architecture which couldn't be solved in a way that would have been automatic. What stopped AMD from implementing their features they have been talking about where nV has them already without developer intervention? Best case numbers AMD gave for Vega in those areas would equalize their stance with nV, which they have stated over multiple times those performance metrics are app dependent.

Developers care. Or should gamers just wait a little longer for Nvidia to catch up before enjoying the features? AMD has the majority of the gaming market with consoles thrown in. So yeah I'd expect developers that target the bulk of the market (consoles) to focus on consoles. With scorpio games should finally be close to directly portable to PC.

 

[ChosenUno]

Well optimized games are not the norm, thus the use of them in arguments is cherry picking. How many games so far have been "well optimized" versus how many games haven't?

Plus, DOOM is an invalid data point because they're using vendor-specific performance optimization. It's an outlier and should not be considered "the norm".

In fact I'm inclined to believe there will be no more than 2 games released in 2017 that will have that level of AMD favoritism.

 

[noko]

You don't think benchmarks on well optimized games using the latest APIs are valid? Sticking to older games that favor Nvidia the most then? A 480 and 1060 are pretty much equal in terms of performance, with 480 favored on newer games. Those power and bandwidth figures only occur with poorly optimized deferred renderers. So gamers get higher fps at the expense of additional input lag. Why the deferred renderers have been sucking with VR.


Where did I say that? In the case of Doom they optimized for both, and the result is Nvidia pundits seem to think it's invalid because AMD is winning. Or is it still invalid because it uses intrinsics that will be included with SM6?


No, only that they share the same technology. The APUs should be really interesting though.


Developers care. Or should gamers just wait a little longer for Nvidia to catch up before enjoying the features? AMD has the majority of the gaming market with consoles thrown in. So yeah I'd expect developers that target the bulk of the market (consoles) to focus on consoles. With scorpio games should finally be close to directly portable to PC.

No, AMD just has to tug Nvidia along as gaming techniques improve while Nvidia optimizes for current and older technologies.

 

[noko]

Well optimized games are not the norm, thus the use of them in arguments is cherry picking. How many games so far have been "well optimized" versus how many games haven't?

Plus, DOOM is an invalid data point because they're using vendor-specific performance optimization. It's an outlier and should not be considered "the norm".

In fact I'm inclined to believe there will be no more than 2 games released in 2017 that will have that level of AMD favoritism.

So then all games that use Nvidia specific optimizations and black box code should also be invalid data points . Doom was optimize for both hardware, Nvidia and AMD - AMD hardware is more capable is the real answer - AMD software holds their GPUs back most of the time with some design flaws as well. If AMD improves performance over time with their drivers with nothing done with the game programming or hardware - the hardware was indeed being held back. Hopefully Vega will remedy some of the hardware issues (Rasterization). Doom is one of the best games to use to show hardware max capability - using a badily coded game just because it gives low numbers and thus ahhhhh it is so graphically intensive is not always the case.

 

[NKD]

Well optimized games are not the norm, thus the use of them in arguments is cherry picking. How many games so far have been "well optimized" versus how many games haven't?

Plus, DOOM is an invalid data point because they're using vendor-specific performance optimization. It's an outlier and should not be considered "the norm".

In fact I'm inclined to believe there will be no more than 2 games released in 2017 that will have that level of AMD favoritism.

Lol. Doom takes advantage of a feature that boosts performance and nvidia doesn't have it. So a company should give its users less optimized experience? Nvidia is going to perform the way it's gonna perform why hold the other users back?

That's just stupid argument. They are squeezing the most performance out of both hardware, why would you want a developer to hold back? So now amd has primitive shaders she developers apparently prefer. They should stop innovating. When vega consoles or even before than we might see developers actually use primitive shaders. Looks like a more future proof hardware to me.

 

[Shintai]

You don't think benchmarks on well optimized games using the latest APIs are valid? Sticking to older games that favor Nvidia the most then? A 480 and 1060 are pretty much equal in terms of performance, with 480 favored on newer games. Those power and bandwidth figures only occur with poorly optimized deferred renderers. So gamers get higher fps at the expense of additional input lag. Why the deferred renderers have been sucking with VR.

This is the point where you have to prove it.

If anything AMD power consumption goes up with DX12 titles etc due to higher utilization rates. 480 doesn't favour newer games, it favours heaily sponsored titles. And your VR excuse is hilarious.

Just wait as always right?