(Updated) AMD Navi GPU to Offer GTX 1080 Class Performance at ~$250 Report Claims

[leadeater]

5 minutes ago, GoldenLag said:

some that has looked at Nvidia in the past consider the 1070/1080 to be mid end and the 1080ti and Titan to be high end. i myself consider the 1080 to be mid end. the Vega 64/liquid is AMD`s high end. In performance you might consider it to be and high end card, but in reality it is just a mid-end card. Then again that wasnt the point i was making. also i knew someone was going to comment on me calling it mid-end. 

As I said enthusiast blinders, if you only look at those levels of cards then it's what ever you place it as personally but in no way is a 1080 anything but high end.

 

Entry-level	GeForce GT 1030
Mid-range	GeForce GTX 1050 GeForce GTX 1050 Ti GeForce GTX 1060
High-end	GeForce GTX 1070 GeForce GTX 1070 Ti GeForce GTX 1080
Enthusiast	Geforce GTX 1080 Ti Nvidia Titan X Nvidia Titan Xp

 

 

And that's not even including the odd ball stuff like 1060 3GB vs 6GB and the new 1030.

[Taf the Ghost]

The one thing to keep in mind is that AMD seem posed to move all of their Consumer SKUs to HBM. Minus the 550-line. HBM actually worked out, but it's been late & expensive (and everyone else wants it). That was a not-small bet that didn't go so great.

[GoldenLag]

6 minutes ago, Taf the Ghost said:

My hope is they've broken down the GCN and rebuilt whatever systems were causing bottlenecks so it scales properly by CUs. Pretty much neither the Vega 56 or 64 are bottlenecked by Clocks or CU, but by the amount of memory throughput they have.  Nvidia has been tweaking things, shrinking things and just adding more SMs, and scaling pretty well from that. AMD makes a brilliant iGPU Architecture that doesn't scale great beyond 32 CUs. That's the big problem.

 

I also kind of hope they can fully split their die designs for Mainstream & Compute. As much as you can't buy a 1080 Ti competitor from AMD, at least they can find a way to fight Nvidia in the really high margin spaces. There's optics but not a lot of money in the high-end GPU space.

I believe some have said that Navi is CGN based. however same as ryzen it is going to connect multiple CGN dies. not to mention the dies are gonna have the number og CU`s that will be at the maximum of the efficiency curve. with low number of CU`s CGN is as efficient as Pascal and Navi is gonna get around the issue with high number of CU`s by connection many small dies together. however AMD might have other things in mind.

[asus killer]

Just now, Taf the Ghost said:

The one thing to keep in mind is that AMD seem posed to move all of their Consumer SKUs to HBM. Minus the 550-line. HBM actually worked out, but it's been late & expensive (and everyone else wants it). That was a not-small bet that didn't go so great.

they actually mention gddr6 for Navi, and gddr wasn't even in Vega.

[laminutederire]

53 minutes ago, ScratchCat said:

*Please don't be another Vega*

 

Everyone should take this with a grain of salt, the hype for Polaris was just as high -"GTX 980 performance at $200".

Hopefully the extra revenue from the CPU department can boost the R&D budget for the GPU sector.

Well it was somewhat up there at launch and it is now in many if most cases.

[GoldenLag]

1 minute ago, leadeater said:

Snip

i didnt come to the conclusion that it was mid end personally, i was however convinced by someone else who gave it a historical perspektive. Though i wont mind others calling it high end or mid-end. again it wasnt my point. 

[leadeater]

4 minutes ago, GoldenLag said:

I believe some have said that Navi is CGN based. however same as ryzen it is going to connect multiple CGN dies. not to mention the dies are gonna have the number og CU`s that will be at the maximum of the efficiency curve. with low number of CU`s CGN is as efficient as Pascal and Navi is gonna get around the issue with high number of CU`s by connection many small dies together. however AMD might have other things in mind.

Myself I doubt the MCM theory, I just don't think the technology, design and fab processing are all collectively at the point to be able to pull it off for a GPU with the bandwidth requirements those have. If you look at the Infinity Fabric as in example it's only 38GB/s and I don't see it jumping from that to the order of 300GB/s+ in the space of time without requiring far too much die space and transistors.

 

We might see a more CCX style but I'm not counting on MCM at all.

[Taf the Ghost]

3 minutes ago, GoldenLag said:

I believe some have said that Navi is CGN based. however same as ryzen it is going to connect multiple CGN dies. not to mention the dies are gonna have the number og CU`s that will be at the maximum of the efficiency curve. with low number of CU`s CGN is as efficient as Pascal and Navi is gonna get around the issue with high number of CU`s by connection many small dies together. however AMD might have other things in mind.

Since Navi showed up on roadmaps, it's been "scalable". We've thought, since AMD is currently selling two very good MCM products, that is what it meant. And, if GCN could be somehow on-package MCM'd, it'd make a lot of sense. I ran the numbers at some point, and if you could stack 4x 560 dies together, you'd get around 1080 performance without any clock increases. Vega actually clocks a good chunk higher on the same process node.

 

The issue is converting GCN over to a full Infinity Fabric approach. Vega already has some aspects built in there, so that's why the MCM assumption made a lot of sense. It's the exact same set of calculations as we see with Threadripper and Epyc. It's slightly worse than a monolithic die approach because of latencies, but they're still decent. And GPU work loads really should be easier to split. The big issue is Drivers and that must be a nightmare.

 

Other thing is that Navi is the last GCN variant. Leaks coming out is that post-GCN, so 2021, will be a completely new Architecture.

[GoldenLag]

1 minute ago, leadeater said:

Myself I doubt the MCM theory, I just don't think the technology, design and fab processing are all collectively at the point to be able to pull it off for a GPU with the bandwidth requirements those have. If you look at the Infinity Fabric as in example it's only 38GB/s and I don't see it jumping from that to the order of 300GB/s+ in the space of time without requiring far too much die space and transistors.

 

We might see a more CCX style but I'm not counting on MCM at all.

I have some doubts of CGN myself. however the theory of the center die in Epyc with Cache might be what they are waiting for to the increase bandwidth. Even Zen will run out of bandwidth if they dont get more . They might be focusing on Zen tech that would be in Zen in the future. It is hard to say how they are going to get around the bandwidth issue

[leadeater]

20 minutes ago, GoldenLag said:

I have some doubts of CGN myself. however the theory of the center die in Epyc with Cache might be what they are waiting for to the increase bandwidth. Even Zen will run out of bandwidth if they dont get more . They might be focusing on Zen tech that would be in Zen in the future. It is hard to say how they are going to get around the bandwidth issue

Zen actually already is out of bandwidth, that's why when you go over 2933Mhz ram the performance doesn't go up. DDR4 2666 dual channel is 42.5GB/s and DDR4 2933 dual channel is 47GB/s.

 

https://www.anandtech.com/show/11551/amds-future-in-servers-new-7000-series-cpus-launched-and-epyc-analysis/2

 

Any memory bandwidth above 38GB/s is going to have very little impact on performance, which incidentally is why Ryzen 2 (Zen+) performance is not actually coming from the faster memory support unless IF bandwidth is also increased.

 

For this reason as well I also doubt the central cache die.

 

[Trixanity]

29 minutes ago, Taf the Ghost said:

I also kind of hope they can fully split their die designs for Mainstream & Compute. 

Lisa Su said that they would; it's inevitable.

[anotherriddle]

I'll continue to discard every rumor that claims to have pricing information of a product this far in the future.

[Taf the Ghost]

14 minutes ago, leadeater said:

Myself I doubt the MCM theory, I just don't think the technology, design and fab processing are all collectively at the point to be able to pull it off for a GPU with the bandwidth requirements those have. If you look at the Infinity Fabric as in example it's only 38GB/s and I don't see it jumping from that to the order of 300GB/s+ in the space of time without requiring far too much die space and transistors.

 

We might see a more CCX style but I'm not counting on MCM at all.

 

5 minutes ago, GoldenLag said:

I have some doubts of CGN myself. however the theory of the center die in Epyc with Cache might be what they are waiting for to the increase bandwidth. Even Zen will run out of bandwidth if they dont get more . They might be focusing on Zen tech that would be in Zen in the future. It is hard to say how they are going to get around the bandwidth issue

Inner working of GPUs isn't my strong suit, but do the CUs have that much bandwidth between themselves as it stands? Further, it's 42 Gb/s per link, and there's 3 links coming off each die in Epyc. AMD would need to put enough links on the dies to connect enough to cover what actually needs to be shared between the CUs.

 

I'm not saying that a MCM GPU isn't a huge amount of technical hurdles, but I think that one might be less of an issue.

[leadeater]

3 minutes ago, Taf the Ghost said:

Inner working of GPUs isn't my strong suit, but do the CUs have that much bandwidth between themselves as it stands? Further, it's 42 Gb/s per link, and there's 3 links coming off each die in Epyc. AMD would need to put enough links on the dies to connect enough to cover what actually needs to be shared between the CUs.

 

I'm not saying that a MCM GPU isn't a huge amount of technical hurdles, but I think that one might be less of an issue.

No idea either but if inter die bandwidth is required it will be a lot, or it's not going to be required at all.

 

The per link bandwidth is still the limiter unless you are going to have local HBM per die but then that would make producing the final render output image insanely complicated, stitching all that together would be as iffy as crossfire without some special magic like Gen-z.

 

EYPC's requirements are quite different, it's running multiple processes on multiple cores across dies which may or may not require bandwidth over the IF. In the case where it does the multiple links are not additive, if you need something from another die you have 38GB/s and that's it, the total 152GB/s is just a bit of marketing play.

[Okjoek]

52 minutes ago, leadeater said:

What? No it isn't.

Yes it is according to AdoredTV.

Set it to about 13 minutes in where he talks about this.

[Taf the Ghost]

4 minutes ago, leadeater said:

No idea either but if inter die bandwidth is required it will be a lot, or it's not going to be required at all.

 

The per link bandwidth is still the limiter unless you are going to have local HBM per die but then that would make producing the final render output image insanely complicated, stitching all that together would be as iffy as crossfire without some special magic like Gen-z.

 

EYPC's requirements are quite different, it's running multiple processes on multiple cores across dies which may or may not require bandwidth over the IF. In the case where it does the multiple links are not additive, if you need something from another die you have 38GB/s and that's it, the total 152GB/s is just a bit of marketing play.

Right, it's the Splitting part of it that I'm not sure how they'd solve. It strikes me it's more of a Opaque Driver approach. GPUs already operate in a massively parallel approach, the issue is how everything comes out of the MCM setup. 

[Coaxialgamer]

46 minutes ago, leadeater said:

As I said enthusiast blinders, if you only look at those levels of cards then it's what ever you place it as personally but in no way is a 1080 anything but high end.

 

Entry-level	GeForce GT 1030
Mid-range	GeForce GTX 1050 GeForce GTX 1050 Ti GeForce GTX 1060
High-end	GeForce GTX 1070 GeForce GTX 1070 Ti GeForce GTX 1080
Enthusiast	Geforce GTX 1080 Ti Nvidia Titan X Nvidia Titan Xp

 

 

And that's not even including the odd ball stuff like 1060 3GB vs 6GB and the new 1030.

It's priced and classed as a high end card. But it's high end in the same sense that the 680 was. It's a small die. It's not much larger than the polaris 10

[asus killer]

for a company that doesn't have Intel or Nvidia money for research wouldn't it be better to concentrate more on just one product? 

Does it make sense to be developing polaris (rx500x), Vega 7nm and Navi?

[leadeater]

3 minutes ago, Okjoek said:

Yes it is according to AdoredTV.

Aaaaaaand.... not clicking on that .

 

It's not overly complicated though, you just need to look at the product stack (in full) and the performance they give then it's a simple case of group cards in to performance tiers. Really the only thing that actually makes a 1080 look not high end is just how much bigger and faster the GP102 die is and the 3 products cut from that, if you consider just how close those 3 top products actually are to each other (1080Ti, Titan X, Titan Xp) it's hardly worth treating them as separate cards but rather one sane purchase and two status symbols .

 

I could support the idea of calling a 1080 mid range if we're going to class those top 3 as high end but that creates a giant problem for every card below a 1080 and what to call those, is a 1030 low end but then is a 1060 6GB also low end? How can those possibly be in the same class name they are nothing alike. Only way 1080 being mid range works is if we ignore the existance of anything less than 1060 6GB.

[GoldenLag]

11 minutes ago, Okjoek said:

Yes it is according to AdoredTV.

Set it to about 13 minutes in where he talks about this.

This is where i changed my idea of the 1080 being enthusiast to mid-range. Mostly cuz im a cheapskate

[Dylanc1500]

4 minutes ago, leadeater said:

Aaaaaaand.... not clicking on that .

 

It's not overly complicated though, you just need to look at the product stack (in full) and the performance they give then it's a simple case of group cards in to performance tiers. Really the only thing that actually makes a 1080 look not high end is just how much bigger and faster the GP102 die is and the 3 products cut from that, if you consider just how close those 3 top products actually are to each other (1080Ti, Titan X, Titan Xp) it's hardly worth treating them as separate cards but rather one sane purchase and two status symbols .

 

I could support the idea of calling a 1080 mid range if we're going to class those top 3 as high end but that creates a giant problem for every card below a 1080 and what to call those, is a 1030 low end but then is a 1060 6GB also low end? How can those possibly be in the same class name they are nothing alike. Only way 1080 being mid range works is if we ignore the existance of anything less than 1060 6GB.

I tend to agree with you, if product stack is taken into context, then the 1080 comes as high end in my personal opinion. The fact of the matter is that what one would consider "high-end" is entirely based on personal opinion.

 

So if a Titan Xp/X is a status symbol lol, what does that make the Titan V?

[leadeater]

5 minutes ago, Coaxialgamer said:

It's priced and classed as a high end card. But it's high end in the same sense that the 680 was. It's a small die. It's not much larger than the polaris 10

It's not a small Pascal die though, it's the third biggest and at full configuration of that die. There are 6 dies and the largest one is not even in the GeForce lineup at all, if we are going to use die size then none of the GeForce cards are high end only the Tesla P100 and Quadro P100 are since they use the GP100 die.

 

Ignoring the GP100 dies since it's not represented in GeForce that brings it up to second largest die of 5.

 

I don't see how a 1080 could be called mid range on performance or die size.

[leadeater]

3 minutes ago, Dylanc1500 said:

So if a Titan Xp/X is a status symbol lol, what does that make the Titan V? 

Dreams.... .

 

Edit:

Err edited that joke, too far lol

[Taf the Ghost]

3 minutes ago, leadeater said:

Aaaaaaand.... not clicking on that .

 

It's not overly complicated though, you just need to look at the product stack (in full) and the performance they give then it's a simple case of group cards in to performance tiers. Really the only thing that actually makes a 1080 look not high end is just how much bigger and faster the GP102 die is and the 3 products cut from that, if you consider just how close those 3 top products actually are to each other (1080Ti, Titan X, Titan Xp) it's hardly worth treating them as separate cards but rather one sane purchase and two status symbols .

 

I could support the idea of calling a 1080 mid range if we're going to class those top 3 as high end but that creates a giant problem for every card below a 1080 and what to call those, is a 1030 low end but then is a 1060 6GB also low end? How can those possibly be in the same class name they are nothing alike. Only way 1080 being mid range works is if we ignore the existance of anything less than 1060 6GB.

1070, 1070 Ti & 1080 are all the same die. It's just that the top-end of the GPU was relatively much stronger, so Nvidia charged more. And keeps charging more. The die itself is only slightly bigger than the 680/770 die. Pascal & the node it's on just happened to turn out really, really good.

[Sauron]

1 hour ago, asus killer said:

i don't get this, AMD already has Vega 64 to compete with 1080 kind of, so what would be the point to be developing a card for a year to do the exact same thing? Yes it would be more power efficient with 7nm, still seems like a lot of work and money for nothing.

I don't believe they are developing Navi at 7nm just to come up with a Vega 64 performance clone. If it ends up to be true it's because Navi was a failure and couldn't be pushed further.

That's a bit like saying that there's no point in making a gtx 1060 when the gtx 780 exists.