Radeon VII

[LePhuronn]

@Davedree of course Navi exists. Just because nobody's held a card up doesn't mean it doesn't exist. In fact, many AMD people have said "we'll talk about Navi later in the year". But you seem to have missed my point. I've always said Navi will replace Vega and Polaris, not surpass them. Nobody has ever even alluded to Navi being any more than a midrange part. Navi may be the last iteration of GCN with all the limitations that come with it, but Navi is going to be a dedicated gaming architecture, which is the focus that AMD have been missing for a while. AMD is a profitable company again so they have the money, resources and cohesion to start their road back. It is part of a full reset to get AMD's graphics game back in order.

Navi will be the start of that journey with a small, cheap, frugal gaming-focussed arch that aims to disrupt the GPU market and hopefully bring some sanity to prices and expectations. GTX 1060 performance on 75W for $130? Hell yes please. But the proper work will be Arcturus: a unified AMD under Lisa Su (i.e. no more RTG and no more in-fighting) with actual money now to invest and built on a cutting-edge process (TSMC 7nm+). And no more GCN.

 

[nashathedog]

A page back posted the answer from the real article, not WCCFTech extract.

I don't usually go to WCCFtech for info so I wouldn't know, although I did find myself on an article of there's linked here at OCUK not too long ago which was written as if they were the interviewer, My post today was a response to what I read on an OC3D net article linked a couple of posts before mine, While I do try to glean info from Chinese & Korean etc articles like you linked a page back the poor machine translations often leave plenty of room for incorrect interpretations so I prefer to wait until we get more reliable translations for info.

The difference between DLSS & DirectML is that the latter is in DirectX (like DXR) while the former is Nvidia custom tech like PhysX

I thought I'd clarified this with what I quoted & highlighted,

Radeon VII would support DirectML, Microsoft's Machine Learning (ML) add-on to DirectX 12

I wonder if Vulkan will be able to use or get these types of updates anytime soon?

 

[D.P.]

A page back posted the answer from the real article, not WCCFTech extract.



The difference between DLSS & DirectML is that the latter is in DirectX (like DXR) while the former is Nvidia custom tech like PhysX

ehh, DLSS and DirectmL have nothing in common, regardless of what AMD marketing might claim.

DLSS is a specific application that uses deep learning accelerated by Turing's unique tensor cores. DirectML is a middleware sitting on top of DirectCompute. AMD GPus lack the tensor cores DL inference is vastly slower, over 12x slower in fact as it relies on a software implementation within the CUs.

 

[D.P.]

@Davedree of course Navi exists. Just because nobody's held a card up doesn't mean it doesn't exist. In fact, many AMD people have said "we'll talk about Navi later in the year". But you seem to have missed my point. I've always said Navi will replace Vega and Polaris, not surpass them. Nobody has ever even alluded to Navi being any more than a midrange part. Navi may be the last iteration of GCN with all the limitations that come with it, but Navi is going to be a dedicated gaming architecture, which is the focus that AMD have been missing for a while. AMD is a profitable company again so they have the money, resources and cohesion to start their road back. It is part of a full reset to get AMD's graphics game back in order.

Navi will be the start of that journey with a small, cheap, frugal gaming-focussed arch that aims to disrupt the GPU market and hopefully bring some sanity to prices and expectations. GTX 1060 performance on 75W for $130? Hell yes please. But the proper work will be Arcturus: a unified AMD under Lisa Su (i.e. no more RTG and no more in-fighting) with actual money now to invest and built on a cutting-edge process (TSMC 7nm+). And no more GCN.

Navi wont have anywhere near the efficiency or pricing of the BS rumours.
There is a misplaced rumour that Vwga10 was solely designed for compute, despite the fact it was missing critical compute functionality and almost all the new features were aimed at gaming such as the new primitive shaders and culling technology.

And even if Vega10 was designed for compute, there isn't a massive amount of compute related functionality that you could strip away. They will liekly go back to GDDr6 which will simplify the memory interface and lower costs but will come with increased power.
The 7nm process is only 25% faster, or 40-50% of the power for same performance, as boen out with Vega20/R7. IF AMD stick to maximizing power gains then Vega10 performance will be about 150-160w. A gtx 1060 level card form AMD on TSMC 7nm will probably still draw 100w, compared to 120 for the 1060 on 16nm.

The prices rumours are liekly the die costs as sold to AIB, who have to produce the board, HSF, VRMS, packaging. Once you add shipping and margins you have to at least double those costs.
7nm is a very expensive process, even when yields are high costs for equal sized dies are at least 3x 16nm costs.

Navi will definitely be an improvement over Polaris and Vega, might equire slightly less power than equivalent Pascal or Turing cards despite the node advantage, and AMD can likely price them a little lower than the Nvidia equivalent and still have a healthy profit. But they wont be anything revolutionary. AMD's architecture is just too far behind,Navu wont be a significiant enough departure, and the 7nm process doesn't provide that much free lunch.

 

[Illuminist]

Navi wont have anywhere near the efficiency or pricing of the BS rumours.
There is a misplaced rumour that Vwga10 was solely designed for compute, despite the fact it was missing critical compute functionality and almost all the new features were aimed at gaming such as the new primitive shaders and culling technology.

And even if Vega10 was designed for compute, there isn't a massive amount of compute related functionality that you could strip away. They will liekly go back to GDDr6 which will simplify the memory interface and lower costs but will come with increased power.
The 7nm process is only 25% faster, or 40-50% of the power for same performance, as boen out with Vega20/R7. IF AMD stick to maximizing power gains then Vega10 performance will be about 150-160w. A gtx 1060 level card form AMD on TSMC 7nm will probably still draw 100w, compared to 120 for the 1060 on 16nm.

The prices rumours are liekly the die costs as sold to AIB, who have to produce the board, HSF, VRMS, packaging. Once you add shipping and margins you have to at least double those costs.
7nm is a very expensive process, even when yields are high costs for equal sized dies are at least 3x 16nm costs.

Navi will definitely be an improvement over Polaris and Vega, might equire slightly less power than equivalent Pascal or Turing cards despite the node advantage, and AMD can likely price them a little lower than the Nvidia equivalent and still have a healthy profit. But they wont be anything revolutionary. AMD's architecture is just too far behind,Navu wont be a significiant enough departure, and the 7nm process doesn't provide that much free lunch.

ehh, DLSS and DirectmL have nothing in common, regardless of what AMD marketing might claim.

DLSS is a specific application that uses deep learning accelerated by Turing's unique tensor cores. DirectML is a middleware sitting on top of DirectCompute. AMD GPus lack the tensor cores DL inference is vastly slower, over 12x slower in fact as it relies on a software implementation within the CUs.

Bleak

 

[melmac]

To be fair I did say the same thing back in dec 2016, abd before that I was stating Fiji had a broken front end, with the fact that Amd hadn't progressed to a x6 wide geometry engine. I was hoping for a 4608/5120 vega, as effectively I predicted vega64 would be cf rx470 performance at 4096. It would only have required another 2 cu's on each geometry engine for 72 cu's, notns lot of work. I was underwhelmed by the die size of vega, with it being a half arsed compute card with no fp64, it really needed a seperate spin off for a leaner gaming Vega.
I think what 4k is saying is that Amd could have cut a lot out of the chip design of Vega10 prior to transposing onto 7nm, therefore having a seperate gaming and compute line.


https://forums.overclockers.co.uk/threads/amd-vega-confirmed-for-2017-h1.18746880/page-105

Agree with what you are saying, but, that's the problem and why would they bother spending all the money on what is basically a throwaway card to make all those changes. Vega isn't the future.

And I am answering your next few posts too Just too lazy to quote them all Again, I sort of agree with all you are said. But, I think it's a cost vs reward situation now. Because I don't think Navi is the future either. It's still going to be GCN. It will still have the same problems and making all the changes that are required to improve it, like you said, just isn't worthwhile. Navi is the last GCN card. Time to move on.

 

[LePhuronn]

@D.P. fully agree the leaked prices for Navi as too low. That could well be AMD's suggested RRP for the reference board, but I expect that to shift up a little even before AIBs have put on their custom coolers, developed custom PCBs, added their own profit margins and everybody else involved in shipping and retail have stuck their percentage on top.

As to the specs I guess we'll just have to see what happens. I don't think there's anything outlandish about them, in fact with all of the tears as of late saying "it's 2019 and AMD still can't beat the 1080 Ti" then the specs are 100% within what everybody believes AMD to be capable of.

I've never been under any illusions that Navi's performance will set the world alight, but I am optimistic/hopeful that the implementation and price of that performance is the big shake up. With the outcry that GTX 1080 performance should be no more than £250 today I'd jump on a Navi RX 3080 like a tramp on a sandwich if it does come out at £250.

 

[shankly1985]

Alright, I finally have the Radeon VII FP64 performance matter sorted out with AMD.

Contrary to earlier statements, it is being throttled. Radeon VII's rate will be 1:8, versus Vega 20's native 1:2 rate. Notably, this is still twice the native FP64 rate of all other Vegas.

https://twitter.com/RyanSmithAT/status/1085680805802733568?s=19

Let's face it they is loads misinformation going around with Vega 7 think it's best people just wait for the official release before making wild claims..

 

[melmac]

@D.P. fully agree the leaked prices for Navi as too low. That could well be AMD's suggested RRP for the reference board, but I expect that to shift up a little even before AIBs have put on their custom coolers, developed custom PCBs, added their own profit margins and everybody else involved in shipping and retail have stuck their percentage on top.

As to the specs I guess we'll just have to see what happens. I don't think there's anything outlandish about them, in fact with all of the tears as of late saying "it's 2019 and AMD still can't beat the 1080 Ti" then the specs are 100% within what everybody believes AMD to be capable of.

I've never been under any illusions that Navi's performance will set the world alight, but I am optimistic/hopeful that the implementation and price of that performance is the big shake up. With the outcry that GTX 1080 performance should be no more than £250 today I'd jump on a Navi RX 3080 like a tramp on a sandwich if it does come out at £250.

It's going to be in a kind of weird position though. Is it replacing both Polaris and Vega or just Polaris? If it's replacing Polaris then it really has to offer 1080 performance for a lot less than £300. The 2060 offers nearly 1080 performance for £329 right now. Navi won't be any shakeup of anything if it's price/performance is any worse than that.

 

[orbitalwalsh]

off topic but hoping the same Skunk Work Team that came up with the Idea for Thread Ripper are mashing together 4 rx580 chiplets made at 7nm and seeing if they can pull off thread ripper for polaris haha

either way, looking forward to seeing how vii does, sure some ultra wide 1440p or 4k owners wouldn't mind increased performance

 

[Kaapstad]

It's going to be in a kind of weird position though. Is it replacing both Polaris and Vega or just Polaris? If it's replacing Polaris then it really has to offer 1080 performance for a lot less than £300. The 2060 offers nearly 1080 performance for £329 right now. Navi won't be any shakeup of anything if it's price/performance is any worse than that.

Or none of them and just goes on the top of AMDs line up.

 

[LePhuronn]

It's going to be in a kind of weird position though. Is it replacing both Polaris and Vega or just Polaris? If it's replacing Polaris then it really has to offer 1080 performance for a lot less than £300. The 2060 offers nearly 1080 performance for £329 right now. Navi won't be any shakeup of anything if it's price/performance is any worse than that.

Based on the the AdoredTV leaks Navi will be replacing Vega and Polaris.

RX 3080 = Vega 64 + 15% at $250 (GTX 1080/RTX 2070)
RX 3070 = Vega 56 at $200 (GTX 1070/RTX 2060)
RX 3060 = RX 580 at $130 (GTX 1060)

And we'll still have Radeon 7 above these 3 to fill the GTX 1080 Ti/RTX 2080 tier.

Ridiculous amounts of salt pinching required, but there's nothing outlandish about these leaks. The specs are easily done, the prices are also very possible but by the time everybody's added their chunk and retailers gouge the hell out of them then the actual prices will probably remove all competitive edge.

 

[Panos]

Alright, I finally have the Radeon VII FP64 performance matter sorted out with AMD.

Contrary to earlier statements, it is being throttled. Radeon VII's rate will be 1:8, versus Vega 20's native 1:2 rate. Notably, this is still twice the native FP64 rate of all other Vegas.

https://twitter.com/RyanSmithAT/status/1085680805802733568?s=19

Let's face it they is loads misinformation going around with Vega 7 think it's best people just wait for the official release before making wild claims..

If it is 1.7TFlops FP64 thats better than the rumoured 764Gflops from yesterday, and miles better than the 420Gflops of the RTX2080Ti (which is 1/32).

 

[Panos]

ehh, DLSS and DirectmL have nothing in common, regardless of what AMD marketing might claim.

DLSS is a specific application that uses deep learning accelerated by Turing's unique tensor cores. DirectML is a middleware sitting on top of DirectCompute. AMD GPus lack the tensor cores DL inference is vastly slower, over 12x slower in fact as it relies on a software implementation within the CUs.

You have written gazillion times over the past year "AMD lacks Tensor cores cannot do XYZ". Yet you completely ignore the fact that AMD doesn't need to have dedicated cores to do such things. For it's design over GPGPU is great.
And the best example is TensorFlow matrix calculations (like CIFAR10) with ROCm vs CUDA. Where the CUDA implementation is using heavily the Tensor Cores on a card like the TV100 (top of the range Volta for business), a humble Vega 64 with it's pitiful 8GB VRAM, keeps up at 90% of it's performance. And wrote humble because it costs 1/8 the price of the TV100.

Stop blindly believing what ever Nvidia marketing says please.

 

[layte]

You have written gazillion times over the past year "AMD lacks Tensor cores cannot do XYZ". Yet you completely ignore the fact that AMD doesn't need to have dedicated cores to do such things. For it's design over GPGPU is great.
And the best example is TensorFlow matrix calculations (like CIFAR10) with ROCm vs CUDA. Where the CUDA implementation is using heavily the Tensor Cores on a card like the TV100 (top of the range Volta for business), a humble Vega 64 with it's pitiful 8GB VRAM, keeps up at 90% of it's performance. And wrote humble because it costs 1/8 the price of the TV100.

Stop blindly believing what ever Nvidia marketing says please.

and all you bring to the table is this single cherry picked benchmark over and over. I'm sure D.P. would be more than happy to furnish you wish a selection of benchmarks showing very different results, it's easy to find huge speedups going from FP32 to FP16 on tensorcore enabled GPUs.

 

[shankly1985]

Adam Kozak of AMD talking to 4Gamer.net

Last year's GDC 2018, Microsoft announced a framework "Windows ML" for Windows 10 platform, and "DirectML" that makes it available from DirectX (October 2018 Update of 10) (related article We are currently experimenting with the evaluation version of SDK of DirectML, but Radeon VII shows excellent results in that experiment.

By the way, Radeon VII scored about 1.62 times the "GeForce RTX 2080" in "Luxmark" which utilizes OpenCL-based GPGPU-like ray tracing renderer. Based on these facts, I think NVIDIA's DLSS-like thing can be done with the GPGPU-like approach for our GPU.

 

[Nasher]

Geforce cards don't get near AMD's when it comes to number crunching, that is pretty common knowledge. They have a pretty narrow scope focusing on just gaming.

 

[Brazo]

Geforce cards don't get near AMD's when it comes to number crunching, that is pretty common knowledge. They have a pretty narrow scope focusing on just gaming.

Your right, AMD basically make their data cards backwards compatible for gamers. It’s a bit like strapping a truck engine onto a skateboard and entering it into an F1 race. Yes it will keep up but you wouldn’t design it that way unless you had no other choice.

 

[dragonslayer]

Based on the the AdoredTV leaks Navi will be replacing Vega and Polaris.

RX 3080 = Vega 64 + 15% at $250 (GTX 1080/RTX 2070)
RX 3070 = Vega 56 at $200 (GTX 1070/RTX 2060)
RX 3060 = RX 580 at $130 (GTX 1060)

And we'll still have Radeon 7 above these 3 to fill the GTX 1080 Ti/RTX 2080 tier.

Ridiculous amounts of salt pinching required, but there's nothing outlandish about these leaks. The specs are easily done, the prices are also very possible but by the time everybody's added their chunk and retailers gouge the hell out of them then the actual prices will probably remove all competitive edge.

I highly doubt that considering Radeon VII is Vega+25% performance there is no way they would sell their own card with only 10% less performance than their top tier for half the price

 

[Nasher]

Your right, AMD basically make their data cards backwards compatible for gamers. It’s a bit like strapping a truck engine onto a skateboard and entering it into an F1 race. Yes it will keep up but you wouldn’t design it that way unless you had no other choice.

And also more future proof with larger and faster memory