AMD's Vega-based Cards to Reportedly Launch in May 2017 - Leak

[Cybrnook2002]

Looking forward to seeing what Vega has. Always like AMD's top gpu's.

 

[Totally]

On the other hand, if you pull all the data from the Titan X review we see that both manufacturers see decreasing performance per flop as they go up the scale and it's worse for AMD:
View attachment 82981

So if we plot out all of these cards as Gflops vs Performance and fit a basic trend line we get that big Vega would be around 80% of Titan performance at 12 TFlops. (this puts it even with the 1080).


View attachment 82980

So for AMD to be at 1080TI levels, they'll need to have improved their card efficiency by 10 - 15 percent for this architecture.

Given the number of changes they've talked about with this architecture, I don't think that's infeasible but it is a hurdle to overcome.

Your conclusion based on the data is wrong. You need to break the data into their proper components. you need to look at the 9XX, 10XX, 3XX, and 4XX separately since they are all different arc, when lump them together like that you are hiding the fact that the scaling of the 10XX(1060->1080) is pretty bad and being propped up by the 9XX when lumped together as AMD vs. Nvidia.

 

[Prima.Vera]

1 word: VAPORWARE

 

[blued]

Not really these are targeted at the 1080Ti not 1080. This would be about average for AMD vs Nvidia release schedule. I do however find it to be a bit annoying on the hype train as per usual. I was expecting a Jan-Feb release.

What has been released vs the 1080 in the last year? Would you call that average?

 

[owen10578]

Right ok..just was saying what i've seen a lot of people and publications has said. Vega=490/Fury2. I guess we'll see what it will be.

 

[renz496]

Branch, there's a reason why Volta was pushed back (and Pascal didn't exist on engineering slides until a year and a bit before release). Vulkan\DX12 caught Nvidia with their pants down, so Pascal ended up being a Maxwell+ arch to use as a stepping whilst Volta is rearchitected.

Vulkan/DX12 is not the reason why pascal exist. also pascal is a bit more complicated than being a simple "maxwell+". nvidia did not want to repeat the same problem they have with kepler so nvidia actually end up making two version of pascal; compute pascal (GP100) and gaming pascal (GP102 and the rest). kepler excel at GPGPU related work especially DP but as a gaming architecture not so much. maxwell design probably the best design nvidia can come up with right now for gaming purpose.

 

[Blueberries]

So if we plot out all of these cards as Gflops vs Performance and fit a basic trend line we get that big Vega would be around 80% of Titan performance at 12 TFlops. (this puts it even with the 1080).

What are you defining "Performance" as? TFLOPS is in itself performance so these graphs make no sense to me.

 

[Kanan]

I'm sure Vega will be great and a success. Why, because AMD changed a lot and they are surrounded by success rather than failure like before. If Ryzen can be a success, Vega can be too.

And yes, those drivers aren't even remotely optimal, with 10% faster speed than 1080 in Doom, and further optimizations and higher clockspeed (it was probably thermal throtteling in that closed cage and with barraged fans that don't help to get the heat out the case) it is indeed possible that it could rival Titan XP / GM102.

 

[Ivaroeines]

A launch from AMD is when AMD ships their chips to manufacturers, then the card-makers have to produce the cards, with the reference designs that should start quite soon after launch. If the purchase start is a coordinated and worldwide event, we may see/buy cards in early autumn(probably some time later than this).

It feels like the time between a AMD product announcement and first purchase date, is the same as a generation for Nvidia and Intel. My guess is that promotion of their products are much more important to AMD than product development.

 

[cdawall]

What has been released vs the 1080 in the last year? Would you call that average?

September 2014 was the release of the 980, June 2015 was the release of the Ti, October 2015 was the release of the fury.

The Vega cards replace the fury's not the 290/390 SKU's from what I have seen in the road map.

I honestly think AMD banked on Polaris clocking a good bit higher and being competitive with the 1070, but were horribly let down by GloFo. AMD has a weird gap yes, but I do not think it was Vega I think it was Polaris failing them.

 

[mtcn77]

'Stupendous', lol.

 

[Totally]

What are you defining "Performance" as? TFLOPS is in itself performance so these graphs make no sense to me.

Good catch I didn't see till you pointed that out I had assumed that it was relative performance(%) between the best performing card and all the others but seeing the top performing card at 98.4 indicates that is not the case.

The Vega cards replace the fury's not the 290/390 SKU's from what I have seen in the road map.

I honestly think AMD banked on Polaris clocking a good bit higher and being competitive with the 1070, but were horribly let down by GloFo. AMD has a weird gap yes, but I do not think it was Vega I think it was Polaris failing them.

But considering they aren't doing anything to address the issue and are awfully quiet about it, seems like a self-defeating strat to me. I'm not interested in paying $$$$ for a Fury replacement nor in a 480 as it falls short of what I need from a card.

 

[cdawall]

But considering they aren't doing anything to address the issue and are awfully quiet about it, seems like a self-defeating to me. I'm not interested in paying $$$$ for a Fury replacement nor in a 480 as it falls short of what I need from a card.

They can't exactly dump glofo. Vega 10 will be your fury and Vega 11 should be the "490"

 

[robert3892]

If AMD releases their flagship GPU in May NVIDIA will already have a head start on them should NVIDIA announce the 1080Ti in March. That isn't very wise marketing in my opinion.

 

[R0H1T]

Vulkan/DX12 is not the reason why pascal exist. also pascal is a bit more complicated than being a simple "maxwell+". nvidia did not want to repeat the same problem they have with kepler so nvidia actually end up making two version of pascal; compute pascal (GP100) and gaming pascal (GP102 and the rest). kepler excel at GPGPU related work especially DP but as a gaming architecture not so much. maxwell design probably the best design nvidia can come up with right now for gaming purpose.

They've always had a gaming flagship & a separate (compute) flagship ever since the days of Fermi. That they've neutered DP on subsequent Titan's is something entirely different, the original Titan had excellent DP capabilities but every card that's followed had DP cut down massively & yet many call it a workstation card.

The GP102 & GP100 are different because of HBM2, Nvidia probably felt that they didn't need "nexgen" memory for their gaming flagship that or saving a few more $ was a better idea i.e. with GDDR5x & the single card cannot support these two competing memory technologies.

 

[Totally]

They've always had a gaming flagship & a separate (compute) flagship ever since the days of Fermi. That they've neutered DP on subsequent Titan's is something entirely different, the original Titan had excellent DP capabilities but every card that's followed had DP cut down massively & yet many call it a workstation card.

The GP102 & GP100 are different because of HBM2, Nvidia probably felt that they didn't need "nexgen" memory for their gaming flagship that or saving a few more $ was a better idea i.e. with GDDR5x & the single card cannot support these two competing memory technologies.

The reason they crippled it was that was Titan cutting into their own workstation graphics business. People aren't going to give up their hard earned when they don't have to and the original Titan presented a very viable alternative to their Quadro line, so Titan in that form had to go. I say it was more a self-preservation move than cost-cutting.

 

[R0H1T]

The reason they crippled it was that was Titan cutting into their own workstation graphics business. People aren't going to give up their hard earned when they don't have to and the original Titan presented a very viable alternative to their Quadro line, so Titan in that form had to go. I say it's self-preservation and not cost cutting.

No one said it wasn't, but they could've gone the route of AMD & given 16/32 GB of HBM2 to 1080Ti/Titan & yet they didn't & IMO that's down a lot to costs.
The Titan is still marketed as a workstation card, I do wonder why?

 

[KrisCo]

I honestly think AMD banked on Polaris clocking a good bit higher and being competitive with the 1070, but were horribly let down by GloFo. AMD has a weird gap yes, but I do not think it was Vega I think it was Polaris failing them.

I just chalked up Polaris to be a stop gap for AMD's real show, same with Pascal for Nvidia. Seems kinda reminiscent of the kepler days. Although I may just be bitter at paying flagship prices for my 680s only to discover they were only the midrange cards a very short time later.

 

[RejZoR]

Only way I can think of this HBC (High Bandwidth Cache) to work is that driver would analyze per-game memory behavior and adapt operation accordingly, meaning the game performance would improve as you play it. Or via game profiles, much like the ones for CrossfireX. This way it would know what gets shuffled around regularly and what only needs rare access. This way they could stuff framebuffer and frequently accessed data in HBM2, less frequent in DDR4 and even less frequent on SSD. Question is, how efficiently can this be manipulated by driver without the need to predefine all this during game design...

 

[Brusfantomet]

Only way I can think of this HBC (High Bandwidth Cache) to work is that driver would analyze per-game memory behavior and adapt operation accordingly, meaning the game performance would improve as you play it. Or via game profiles, much like the ones for CrossfireX. This way it would know what gets shuffled around regularly and what only needs rare access. This way they could stuff framebuffer and frequently accessed data in HBM2, less frequent in DDR4 and even less frequent on SSD. Question is, how efficiently can this be manipulated by driver without the need to predefine all this during game design...

What you are describing here is Cache prefetching. witch has been done in hardware for over 15 years by both AMD and Intel. It could aslo be an hybrid approach with both hardware and software, and this is the part of the Vega drivers thats was not done during the doom Vega demo.

 

[RejZoR]

Not necessarely. I'm talking direct memory access, not prefetching. Prefetching still relies on storing whatever data into main GPU memory based on some sort of prefetching algorithm (HBM2 in this case). HBC, from what I understand AMD's slides is mentioning direct, seamless access to these resources. Not sure how, but apparently it can be done. We'll know more when they actually release this thing.

We've seen similar with HyperMemory and TurboCache on low end cards where system memory seamlesly behaved like on-board graphic memory. It saved costs while delivering nearly the same performance. Having systems with quad channel DDR4 RAM, I can see that as a viable game data storage memory.

 

[Brusfantomet]

The hardware prefetching uses DMA, the software one does not need to (since it already runs on the cpu).

this is not the same as Hyper memory/TurboCache those used system memory as an extended frame buffer to increase the frame buffer of the card. The dynamic part was only changing the size of the system ram frame buffer.

Quad channel DDR 4 is not viable for high end gaming, your 2400 MHz ram has a theoretical bandwith of 76,8 GB/s witch is about the same as a R7-250 at 73,6 GB/s.

 

[Slizzo]

The slide below is from GTC2014

The slide below is from GTC 2015

In 2013, Volta was present but it's HMC design stalled somewhat.

In 2014 (that would be almost 3 years ago, Volta had disappeared from the road map.

In 2015, Volta reappeared and it seems very much for a late 2017 release, though it's far enough to be 2018.

So really, Vega, DX12 etc has got nothing to do with Volta. The memory arrangement affected his position.

Please bear in mind that the only reason Titan X is £1100 is because AMD have NOTHING to touch it with. Nvidia couldn't care less about DX12 and Vulkan for it's GFX cards - their own mid range GP104 (not GP102 and not GP100) is still top dog.

Nvidia by your own definition (a revamped Maxwell masquerading as Pascal) don't even need to try to stay King of Cards.

Please keep in mind that Volta's release has been pulled up for a few reasons. This was covered by our very own @btarunr here: https://www.techpowerup.com/224413/nvidia-accelerates-volta-to-may-2017

It's sticking to 16nm, and will not be moving to 10nm. Also, there's rumors that they needed to pull it up to meet a government contract.

 

[Prima.Vera]

I'm sure Vega will be great and a success. Why, because AMD changed a lot and they are surrounded by success rather than failure like before. If Ryzen can be a success, Vega can be too.

And yes, those drivers aren't even remotely optimal, with 10% faster speed than 1080 in Doom, and further optimizations and higher clockspeed (it was probably thermal throtteling in that closed cage and with barraged fans that don't help to get the heat out the case) it is indeed possible that it could rival Titan XP / GM102.

A lot of "IF"s and "Could"s in there....

 

[efikkan]

Only way I can think of this HBC (High Bandwidth Cache) to work is that driver would analyze per-game memory behavior and adapt operation accordingly, meaning the game performance would improve as you play it. Or via game profiles, much like the ones for CrossfireX. This way it would know what gets shuffled around regularly and what only needs rare access. This way they could stuff framebuffer and frequently accessed data in HBM2, less frequent in DDR4 and even less frequent on SSD. Question is, how efficiently can this be manipulated by driver without the need to predefine all this during game design...

The algorithm you describe is not possible. In games, the part of the allocated memory that's not used in every frame would be the game would outside the bounds of the camera. But it's impossible for the GPU to know which parts will be used next.

Not necessarely. I'm talking direct memory access, not prefetching. Prefetching still relies on storing whatever data into main GPU memory based on some sort of prefetching algorithm (HBM2 in this case). HBC, from what I understand AMD's slides is mentioning direct, seamless access to these resources. Not sure how, but apparently it can be done. We'll know more when they actually release this thing.

The purpose of caching is to hide latency of a larger storage pool. The two basic prinicples are manual and automatic prefetching. Manual prefetching would require implementation in every game, but AMD has indicated they are talking about automatic prefetching. Automatic prefetching can only work if it's able to detect (linear) patterns in memory accesses. This can work well for certain compute workloads, where the data processed is just a long linear stream of data. It is however impossible to do with random accesses, like rendering. If they try to do this it will result in either unstable performance or popping resources, depending on how they handle missing data on the hardware side.