AMD VEGA confirmed for 2017 H1

[drunkenmaster]

Based on the colour of a nvlink connector, wtf?

Even if it can run fp32 code on the fp64 cores... it would almost instantly run into massive throttling trying to use all the cores at the same time(most likely), which is why it doesn't list FP32 performance at 15 or 20flops, ie the FP32 cores + either 1 or if ultra clever, 2 FP32 instructions bundled up on each FP64 core.

It has less FP32 shaders than available on a GP102 card, NVlink likely takes up a fair amount of power, any off die connection takes power, IO and die space which takes away from actual performance. Gaming doesn't require the bandwidth available on NVlink so would be nothing short of a waste of power and cause more throttling than a basic normal sli link.

There is no reason to believe they'll release GP100 as a gaming card ever, if they felt it was a good option... they wouldn't have spent several tens of millions making the GP102 core if they could just make GP100 for both uses.

NVlink has no practical gaming purposes, the bandwidth available on Nvlink is not designed for gaming loads and only adds cost, power usage and uses up transistors for no benefit.

Could you get decent gaming performance from the same shader count as Titan X, sure, if the drivers work well enough and/or you can install normal drivers for it. But it won't be really any better performance than Titan X, but from a core that costs 20-25% more to make. Gamers should have no interest in it, same gaming performance from Titan X, Nvidia should have no interest in selling it as a Titan, it costs more to make so either they make less profit at the same price, or they charge 25% more for the same margin for a card with the same performance.

 

[Kaapstad]

Based on the colour of a nvlink connector,

Even if it can run fp32 code on the fp64 cores... it would almost instantly run into massive throttling trying to use all the cores at the same time(most likely), which is why it doesn't list FP32 performance at 15 or 20flops, ie the FP32 cores + either 1 or if ultra clever, 2 FP32 instructions bundled up on each FP64 core.

It has less FP32 shaders than available on a GP102 card, NVlink likely takes up a fair amount of power, any off die connection takes power, IO and die space which takes away from actual performance. Gaming doesn't require the bandwidth available on NVlink so would be nothing short of a waste of power and cause more throttling than a basic normal sli link.

There is no reason to believe they'll release GP100 as a gaming card ever, if they felt it was a good option... they wouldn't have spent several tens of millions making the GP102 core if they could just make GP100 for both uses.

NVlink has no practical gaming purposes, the bandwidth available on Nvlink is not designed for gaming loads and only adds cost, power usage and uses up transistors for no benefit.

Could you get decent gaming performance from the same shader count as Titan X, sure, if the drivers work well enough and/or you can install normal drivers for it. But it won't be really any better performance than Titan X, but from a core that costs 20-25% more to make. Gamers should have no interest in it, same gaming performance from Titan X, Nvidia should have no interest in selling it as a Titan, it costs more to make so either they make less profit at the same price, or they charge 25% more for the same margin for a card with the same performance.

You have made a whole load of assumptions and errors in your post.

My post was just about the NVLink and the only pic I had to show one was the Quadro image.

Just for the record both the Pascal Titan and the Quadro above have exactly the same number of active cores (3584).

 

[Mtom]

OMG i somehow managed to get into an Nvidia topic instead of the Vega one i wanted to visit.

 

[Mauller]

It has less FP32 shaders than available on a GP102 card, NVlink likely takes up a fair amount of power, any off die connection takes power, IO and die space which takes away from actual performance. Gaming doesn't require the bandwidth available on NVlink so would be nothing short of a waste of power and cause more throttling than a basic normal sli link.

There is no reason to believe they'll release GP100 as a gaming card ever, if they felt it was a good option... they wouldn't have spent several tens of millions making the GP102 core if they could just make GP100 for both uses.

NVlink has no practical gaming purposes, the bandwidth available on Nvlink is not designed for gaming loads and only adds cost, power usage and uses up transistors for no advantage.

NVlink and AMD's equivalent do have advantages with gaming. The main advantage is that the links are designed to be far lower latency compared to pci-e, which is what nvidias sli links are based on.

So although the bandwidth may be unnecessary,the latency improvements are not.

GP100 also has a fair amount of die space dedicated to 8 Nvlink connections. So only 1/4 are being used on this card. So depending on the amount of silicon required, only a single nvlink would be needed as a replacement for sli on gaming cards for a dual gpu setup.

It makes me wonder if AMDs equivalent to NVlink is one of the unnamed features with vega. Rajas vega cube makes me think that the tech is there.

 

[Geck0]

Every time I come in here it's off topic. Lame.

 

[nashathedog]

whatever he "claims" unless it's made by AMD in an official manner, then it has the same value as a vacuumed dead rat.

I agree, That's why I was asking who knows the channel and what's he like? The fact that such a claims been made and it's not big news is surprising. Panos posted the video so maybe he knows more about the channel.

 

[Geck0]

Just noticed something interesting with the new NVidia Quadro cards.

The NVLink they use looks more like it was styled for their gaming cards (no green colouring), could this be a clue to an upcoming new NVidia gaming card.

How about you go "notice" them in an nVidia thread?

Seriously.

 

[Kaapstad]

How about you go "notice" them in an nVidia thread?

Seriously.

Because this is the Vega thread and it is full of what people think NVidia will do in response.

 

[KillBoY_UK]

You have made a whole load of assumptions and errors in your post.

This thread is 95% assumptions and hope until Vega release day

 

[JediFragger]

Spoiler

Just noticed something interesting with the new NVidia Quadro cards.

The NVLink they use looks more like it was styled for their gaming cards (no green colouring), could this be a clue to an upcoming new NVidia gaming card.

Why the **** is this even in here!???

 

[Kaapstad]

Why the

The reason I posted it is because almost everyone has decided NVidia's response to Vega will be a cut down Pascal Titan in the shape of a 1080 Ti. I just wanted to highlight that NVidia could do something different when Vega turns up.

 

[drunkenmaster]

You have made a whole load of assumptions and errors in your post.

My post was just about the NVLink and the only pic I had to show one was the Quadro image.

Just for the record both the Pascal Titan and the Quadro above have exactly the same number of active cores (3584).

Just for the record, no they don't. First GP100 has 5,379 cores, it's right there in the image you linked. Second, the GP102 Quadro(highest one) has 3840 active cores, Titan has less enabled but there are more FP32 cores on there. So regardless of which Quadro you meant, you're wrong.

Now for the kicker...

it's actually crazy hard to find proof but these two links help.

http://www.nvidia.com/object/sli-certified-systems-and-motherboards.html

First up, certified SLI, not NVlink platforms, P6000 is listed in this, Quadro GP100 is not.


http://images.nvidia.com/content/pd...2-NV-DS-Quadro-P6000-US-12Sept-NV-FNL-WEB.pdf


Data sheet on P6000, no mention of Nvlink at all.


http://www.nvidia.com/object/quadro...tm_medium=referral&utm_campaign=quadro-pascal

bottom right, quadro SLI HB connector... wonder which cards use that and which uses the NVlink connector.

https://arstechnica.co.uk/gadgets/2017/02/nvidia-quadro-gp100-price-details/

but this really is the kicker

The reduction in total memory (24GB on the P6000) may cause problems for users with large data sets, though. Nvidia is hoping to offset the difference with the inclusion of a unique NVLink connector.

Where the Tesla P100 communicates entirely over Nvidia's proprietary NVLink standard—which allows multiple GPUs to connect directly to each other or supporting CPUs at a much higher bandwidth than PCIe 3.0—the Quadro GP100 includes a NVLink connector on top of the card in place of an SLI connector.

GP102-106 do not support NVlink, Gp100 supports NVlink.... so the only possible gaming card to use an nvlink connector would be based on GP100 and as with my previous points, as the FP64 cores would quickly cause throttling and as GP102 is much smaller, cheaper and has more FP32 cores anyway, Gp100 makes zero sense as a gaming card in any way shape or form. It was never designed for gaming, the biggest issue surrounding GP100, is why if it's the DP card for Tesla and the DP card for quadro.... did they not have double the DP cores and zero SP cores frankly, or more likely for power reasons, 50% more DP cores and be 1/3rd smaller than it is.


As said it's actually hard to find proof, but just in the way that AMD have stated 3.4Ghz+ on Ryzen, some people have gone nuts and decided every Ryzen chip will be 3.4Ghz+ but it's AMD talking about the maximum capability of their best/flagship part using Zen, Nvidia announcing Pascal talked about it's maximum capability, not every chip. P6000 isn't designed to go into an HPC box with 8 cards connected by Nvlink, so why waste the transistors and power adding it. NVlink is only present on GP100, thus an NVlink bridge would only work on a card with a 'unique NVlink connector', which again only GP100 has.

 

[Panos]

Talking about NVlink I would like to point out that Vega comes out with similar tech, in addition it will be available in the mainstream making available the option for dual core cards work as a single unit if so.
Sharing also the HBM2 VRAM between them.

And at least we know AMD going to support it on their mainstream products not the ones cost half car.
(north of $5000)

 

[drunkenmaster]

Infinity fabric can potentially enable two cores on one interposer to work nearly seamlessly together, though I don't believe any Vega chips will be like that, I think Navi is exactly intended to work as smaller dies put together to work as a large die.

But Infinity Fabric is entirely different, and not at all similar to NVlink.

Infinity fabric will send a memory coherent signal from CPU to GPU, but across the pci-e bus, Nvlink is meant to replace and/or supplement the PCI-E bus. Infinity Fabric is well, very complex as it can work in multiple ways, it both is internal hardware inside Vega used for inner chip communication(every chip has some kind of internal communication structure, 2900xt had a ringbus, most gpus since and before have used a crossbar system, Intel use ringbus in i3-i7 currently though it's potentially being replaced soon iirc), it will be used inside Zen also, it can also go external and be used as a HT like link, and it can be embedded in other things like pci-e.

It's hard to say Infinity Fabric is internal while NVlink is external... though in reality that is broadly correct, just Infinity fabric is designed to be more than just an internal bus, it's kind of a go anywhere, implement anyway you see fit monumental update to Hypertransport.

It certainly sounds like the implementation on Navi is to implement the many small dies working as one strategy. It should make it easier and probably far more robust than current xfire/sli connections including Nvlink, but there is an inherent problem with latency when going off package and then off card. It's going to be used more for an X2 type card where instead of two separate chips/packages, they put two chips onto one interposer and treat them as one, rather than making sli/xfire setups suddenly work seamlessly as if there is only one chip working.

 

[Rroff]

Yup the value of something like that would be in huge teaming application for specific data crunching - for gaming type use there would simply be too much latency going off card/core atleast with current designs and it doesn't sound like the changes to Vega have any impact on that - Navi maybe.

 

[CAT-THE-FIFTH]

How about you go "notice" them in an nVidia thread?

Seriously.

I have a feeling overall confidence in AMD delivering anything in the GPU and CPU area is quite low by the standards of enthusiasts - its why I do hope both their next launch GPU and CPU launches hit the ground running in terms of performance and quality of the launched products(no buggy GPU coolers for example),since the last few have been rather hit or miss(or have been delayed way too much in terms of actual products being available - Kaveri is an example of this). Even if AMD can't dislodge the best Nvidia cards or the best DIY CPUs Intel can offer,they just need to be able to cover enough of their ranges with decent enough products,and with a competent launch.

I emphasise both since AMD as a brand is negatively affected if either of its ranges don't fire,and I suspect the last few years of their CPUs being a bit meh hasn't helped their perception especially among some of the less techy types(sort of seen people have some connection between the CPUs being meh and their GPUs being meh by connection,etc).

 

[Legend]

I have a feeling overall confidence in AMD delivering anything in the GPU and CPU area is quite low by the standards of enthusiasts - its why I do hope both their next launch GPU and CPU launches hit the ground running in terms of performance and quality of the launched products(no buggy GPU coolers for example),since the last few have been rather hit or miss(or have been delayed way too much in terms of actual products being available - Kaveri is an example of this). Even if AMD can't dislodge the best Nvidia cards or the best DIY CPUs Intel can offer,they just need to be able to cover enough of their ranges with decent enough products,and with a competent launch.

I emphasise both since AMD as a brand is negatively affected if either of its ranges don't fire,and I suspect the last few years of their CPUs being a bit meh hasn't helped their perception especially among some of the less techy types(sort of seen people have some connection between the CPUs being meh and their GPUs being meh by connection,etc).

I think it's worse than expectations for them being low... they've almost become unrecognisable as a viable entity in the CPU sector, and are going to have to work REALLY hard to shake the budget brand perception that is attached to their GPU range. I hope Ryzen is a big success for them though, and the pricing is looking very interesting... let's just hope the performance is there.

They do really need to deliver something special with VEGA though, and which offers superior 1080 performance. I don't expect it to be 1080Ti levels, but that won't matter if they can offer a 1080 rival at an attractive price. VEGA still won't be cheap though, but that is why the performance level needs to be high, otherwise AMD will forever remain in that budget category of GPUs. AMD have gone out of their way to exclaim a desire to recapture some of that high end market share (which means high end pricing), but the only way they're going to do that is if they deliver on the performance. Time will tell.

 

[Rezident]

With a freesync monitor I'm already set on buying vega but at this point I'm keeping my expectations low as to avoid disappointment

Hah! Same as. Well, there is always hope.

 

[Kaapstad]

Just for the record, no they don't. First GP100 has 5,379 cores, it's right there in the image you linked. Second, the GP102 Quadro(highest one) has 3840 active cores, Titan has less enabled but there are more FP32 cores on there. So regardless of which Quadro you meant, you're wrong.

Now for the kicker...

it's actually crazy hard to find proof but these two links help.

http://www.nvidia.com/object/sli-certified-systems-and-motherboards.html

First up, certified SLI, not NVlink platforms, P6000 is listed in this, Quadro GP100 is not.


http://images.nvidia.com/content/pd...2-NV-DS-Quadro-P6000-US-12Sept-NV-FNL-WEB.pdf


Data sheet on P6000, no mention of Nvlink at all.


http://www.nvidia.com/object/quadro...tm_medium=referral&utm_campaign=quadro-pascal

bottom right, quadro SLI HB connector... wonder which cards use that and which uses the NVlink connector.

https://arstechnica.co.uk/gadgets/2017/02/nvidia-quadro-gp100-price-details/

but this really is the kicker



GP102-106 do not support NVlink, Gp100 supports NVlink.... so the only possible gaming card to use an nvlink connector would be based on GP100 and as with my previous points, as the FP64 cores would quickly cause throttling and as GP102 is much smaller, cheaper and has more FP32 cores anyway, Gp100 makes zero sense as a gaming card in any way shape or form. It was never designed for gaming, the biggest issue surrounding GP100, is why if it's the DP card for Tesla and the DP card for quadro.... did they not have double the DP cores and zero SP cores frankly, or more likely for power reasons, 50% more DP cores and be 1/3rd smaller than it is.


As said it's actually hard to find proof, but just in the way that AMD have stated 3.4Ghz+ on Ryzen, some people have gone nuts and decided every Ryzen chip will be 3.4Ghz+ but it's AMD talking about the maximum capability of their best/flagship part using Zen, Nvidia announcing Pascal talked about it's maximum capability, not every chip. P6000 isn't designed to go into an HPC box with 8 cards connected by Nvlink, so why waste the transistors and power adding it. NVlink is only present on GP100, thus an NVlink bridge would only work on a card with a 'unique NVlink connector', which again only GP100 has.

Nope you are still making assumptions.

The pic I linked definitely states 3584 FP32 cores.

I should not have to point out to you that FP64 is no good for gaming.

I also notice that your addition is a bit out if you are adding the two together.

 

[CAT-THE-FIFTH]

I think it's worse than expectations for them being low... they've almost become unrecognisable as a viable entity in the CPU sector, and are going to have to work REALLY hard to shake the budget brand perception that is attached to their GPU range. I hope Ryzen is a big success for them though, and the pricing is looking very interesting... let's just hope the performance is there.

They do really need to deliver something special with VEGA though, and which offers superior 1080 performance. I don't expect it to be 1080Ti levels, but that won't matter if they can offer a 1080 rival at an attractive price. VEGA still won't be cheap though, but that is why the performance level needs to be high, otherwise AMD will forever remain in that budget category of GPUs. AMD have gone out of their way to exclaim a desire to recapture some of that high end market share (which means high end pricing), but the only way they're going to do that is if they deliver on the performance. Time will tell.

This is why I am saying is if they can have a solid Vega launch and a solid Ryzen launch,together people will have more confidence in them.

This is why I hope they have decent cooling for the Vega cards at launch and reasonable pricing for the performance level they provide.

Even look at the Fury launch - if it had launched with coolers which were not buggy and AMD had even launched at £50 lower prices for the Fury X and Fury,it would have given a far better impression. The problem is that they tried to match the pricing(or even slightly exceed the pricing) of the GTX980 and GTX980TI. Yes,it worked to maintain the pricing of the larger volume R9 390 and R9 390X series cards,but it didn't help their perception as much as they should have.