[Rumor, Tweaktown] AMD to launch next-gen Navi graphics cards at E3

[railven]

Answer hazy, ask again later.

We don't really know. Maybe it is clean sheet, maybe it is not. The general assumption of most people following RTG is that Navi would be the last GCN-related design before they went clean sheet in 2021 (or whenever "generation next" comes out). Navi might be a stepping-stone to that new design. Is anything really clean sheet anymore?

Welps, guess I'll wait for info. AMD is firing good on the CPU section, rattle Intel up. Hopefully they can bring some of that energy over to GPU.

 

[itsmydamnation]

Just because something is clean sheet doesn't mean you dont reuse big parts of something you already have created. It means you start with a fresh design target, some or many of those targets might still come up the same.

But i also think "GCN" gets a bad name because most people have NFI what GCN actually is. Games that run well on Vega hit the GCN part of the GPU harder then the games that run bad on Vega

 

[Dribble]

Said who?

Navi is supposed to replace Vega.

AMD already stopped production of Vega, with the intention of replacing it with Navi.

By contrast, AMD hasn't stopped production of Polaris at all.

Which means Navi's first competitor is Vega - the 56 isn't that expensive these days.

 

[AtenRa]

If AMD releases that at $399, people will yawn. It has to be $329 at most.

I will pass for sure if it is RTX2070 performance at $399 but,
from a market product positioning , this performance at $399 will take sales away from both RTX2070 ($499) and RTX2060 ($349).

 

[tviceman]

I will pass for sure if it is RTX2070 performance at $399 but,
from a market product positioning , this performance at $399 will take sales away from both RTX2070 ($499) and RTX2060 ($349).

True RTX 2070 performance for $399 would be a "good deal" in today's environment, even without hardware ray tracing. I still think it's going to end up between the RTX2060 and 2070 (and closer to the 2060) overall.

With my performance expectations of it being about 10% faster than the RTX 2060, I AMD comes in at $329-349.

 

[Mockingbird]

Ain't that the truth. Everyone who was in the market for Polaris over the last three years is saying the saaaaame thing.

Can we get Navi at $279 or less, please? Preferably $239 or less?

For $279, there is the GeForce GTX 1660 Ti

 

[AtenRa]

True RTX 2070 performance for $399 would be a "good deal" in today's environment, even without hardware ray tracing. I still think it's going to end up between the RTX2060 and 2070 (and closer to the 2060) overall.

With my performance expectations of it being about 10% faster than the RTX 2060, I AMD comes in at $329-349.

yes , +10% over RTX2060 at $349 is also a good deal

 

[Mockingbird]

Unless they're holding back something big, this is a colossal disappoint and a terrible release.

How could I possibly be excited about a gpu about as fast as the GTX 1080 from May 2016 (more than 3 years earlier than its launch date) for $100 less than a GTX 1080 was in May of 2016. If this is all they have to show, its a total failure.

64 CU - AGAIN. What is wrong with their design team??

GeForce RTX 2060 is already almost as fast as the GeForce GTX 1080 for $349

 

[Mockingbird]

Which means Navi's first competitor is Vega - the 56 isn't that expensive these days.

Well, no

Vega is going to be out of stock soon

 

[Mockingbird]

Polaris is in need of retirement. It is older than RX Vega. I get that they needed to stop production of RX Vega because of HBM. Polaris needs to go away, and Navi was supposed to put it out to pasture.

Now we have the RX590 anchoring AMD's midrange lineup for who knows how long. Pathetic.

Replacing Vega is obviously a priority.

Vega is obviously unprofitable, and AMD has already discontinued its production with the intention of replacing it with Navi.

By contrast, Polaris is still profitable

 

[Glo.]

Spot on. This View attachment 6721View attachment 6722 is at least a 256mm^2 chip, if not 275mm^2, as predicted by Ryan;

which makes it 25% smaller than the other 7mm AMD Chip (Vega 20)
@ 25% less than Vega 20 (331mm^2/64CUs) = ~ 250mm^2 / 48CUs

However, @ 275mm^2 or higher, it would be a 56 CU chip
@ 15% less than Vega 20 (331mm^2/64CUs) = ~ 280mm^2 / 56CUs

Also, if this is a 56CU chip and the yields are not good, then AMD can cut it thrice; in 8CU increments (56 / 48 / 40 CU parts)

They can launch the 40 & 48 CU chips now and build up an inventory of full dies. Launch these cut down parts now, where 40 CU competes with Rx2060 and 48CU part competes with Rx2070.

Then, once they've built up a healthy inventory & Nvidia launches a Rx2070Ti, they can simply release the full 56CU chip.

You do realize, that AMD's new CUs could burn a lot of Area? You do realize that more cache could burn area? You do realize that potentially higher clocks could burn area? You do realize that Navi is pretty much incomparable to previous versions of AMD GPUs, because of vast departure from the previous designs?

 

[Dribble]

Well, no

Vega is going to be out of stock soon

But if it replaces Vega with something that isn't better in performance/$ then all you will hear is "well that's no better then vega which has been out for years".

 

[Stuka87]

But if it replaces Vega with something that isn't better in performance/$ then all you will hear is "well that's no better then vega which has been out for years".

Vega is only cheap right now because they want its inventory gone. That price can not be used as a comparison.

 

[french toast]

Spot on. This View attachment 6721View attachment 6722 is at least a 256mm^2 chip, if not 275mm^2, as predicted by Ryan;

which makes it 25% smaller than the other 7mm AMD Chip (Vega 20)
@ 25% less than Vega 20 (331mm^2/64CUs) = ~ 250mm^2 / 48CUs

However, @ 275mm^2 or higher, it would be a 56 CU chip
@ 15% less than Vega 20 (331mm^2/64CUs) = ~ 280mm^2 / 56CUs

Also, if this is a 56CU chip and the yields are not good, then AMD can cut it thrice; in 8CU increments (56 / 48 / 40 CU parts)

They can launch the 40 & 48 CU chips now and build up an inventory of full dies. Launch these cut down parts now, where 40 CU competes with Rx2060 and 48CU part competes with Rx2070.

Then, once they've built up a healthy inventory & Nvidia launches a Rx2070Ti, they can simply release the full 56CU chip.

There has been Alot of changes, 25% IPC and higher clocks, 2x shader engines? more cache?.. Is sure not going to come for free, my guess is new set up CU vs CU is less dense, at the benefit of efficiency, IPC and clocks.

By I stand by the 40 CU is not good enough to compete with 2070 at good efficiency, Navi is a brand new uarch on 7nm, it should in theory be more efficient in traditional raster than Turing.. In theory.

If they skimp on resources and try to clock it past its efficiency spot like is AMD tradition... Then they it might come up short.

5-10% faster than 2070 at $399 with equal power consumption (better efficiency) would be fantastic, how likely though?

 

[gdansk]

I'm skeptical 'RDNA' diverges much from GCN for two reasons:
1. Su said GCN will continue in compute focused applications.
2. AMD would not try to maintain drivers for two different architectures.

If it is very different ISA, I'd be surprised anyway.

 

[exquisitechar]

If it is very different ISA, I'd be surprised anyway.

It isn't.

 

[tviceman]

5-10% faster than 2070 at $399 with equal power consumption (better efficiency) would be fantastic, how likely though?

Not at all likely.

 

[crazzy.heartz]

You do realize, that AMD's new CUs could burn a lot of Area? You do realize that more cache could burn area? You do realize that potentially higher clocks could burn area? You do realize that Navi is pretty much incomparable to previous versions of AMD GPUs, because of vast departure from the previous designs?

That's a lot of could's..

It's a new uArc yes, so it gains a few things and looses a few things.. Kind of a rebalancing act for AMD as they need to make the most out of die space.. I am pretty sure this is a 256mm^2 chip and that fits perfectly into that 44-48 CU bracket. 40CU setup sounds good, but i think that's a bit less for a chip that's competing against a RTX2070, IMO.

There has been Alot of changes, 25% IPC and higher clocks, 2x shader engines? more cache?.. Is sure not going to come for free, my guess is new set up CU vs CU is less dense, at the benefit of efficiency, IPC and clocks.

If they skimp on resources and try to clock it past its efficiency spot like is AMD tradition... Then they it might come up short.

5-10% faster than 2070 at $399 with equal power consumption (better efficiency) would be fantastic, how likely though?

Acc. to Videocardz, the two chips are 180 and 225 Watts respectively. Also, all the Navi preview cards showcased so far have 2 8-Pin PCIe connectors Appears the chips are running at their maximum possible Clocks as a lot is riding on the first "Navi" GPU.. Can't be slower than the competition, right.. However, most can simply tweak them for power savings in Wattman, if required.

Even then, these are only sipping 20-40 Watts more than their counterparts, which is a LOT better than 100+ Watts more, as was with Polaris/Vega.. (having no compute / RT HW helps..)

Based on their track record, and that Sapphire Reps statement; I am pretty sure the lower SKU beats Rtx2060 convincingly. But, the faster part, that AMD demoed, most likely trades blows with Rtx2070 & leads only in games AMD has generally better performance. Would like nothing better than to be proven wrong though..

 

[DrMrLordX]

For $279, there is the GeForce GTX 1660 Ti

That's the product AMD needed to knock off its pedestal.

By contrast, Polaris is still profitable

Not if they can't move any units. Why would anyone buy a 590 now? Or really, any Polaris? It would have to be heavily discounted for people to want one vs. the 1660 or 1660 Ti.

 

[Stuka87]

Even then, these are only sipping 20-40 Watts more than their counterparts, which is a LOT better than 100+ Watts more, as was with Polaris/Vega.. (having no compute / RT HW helps..)

RX480 didn't use 100W more than the GTX 1060. It did use about 50W more, depending on which versions of each card were being used. And which BIOS the 480 was using. I agree that Vega was a huge power sucker.

We will have to wait and see what navi uses. Its very possible the boards displayed were development boards, which almost always have 2x8 pin connectors on them, nVidia included.

 

[Glo.]

That's a lot of could's..

It's a new uArc yes, so it gains a few things and looses a few things.. Kind of a rebalancing act for AMD as they need to make the most out of die space.. I am pretty sure this is a 256mm^2 chip and that fits perfectly into that 44-48 CU bracket. 40CU setup sounds good, but i think that's a bit less for a chip that's competing against a RTX2070, IMO.

Those "could's" are put there to give you hint, that all of that puts out the window what you were trying to calculate in your post.

Your calculations may be correct if it would be Vega based GPU. But it isn't and your calculations are pointless, considering vast redesign of the Architecture.

 

[french toast]

That's a lot of could's..

It's a new uArc yes, so it gains a few things and looses a few things.. Kind of a rebalancing act for AMD as they need to make the most out of die space.. I am pretty sure this is a 256mm^2 chip and that fits perfectly into that 44-48 CU bracket. 40CU setup sounds good, but i think that's a bit less for a chip that's competing against a RTX2070, IMO.



Acc. to Videocardz, the two chips are 180 and 225 Watts respectively. Also, all the Navi preview cards showcased so far have 2 8-Pin PCIe connectors Appears the chips are running at their maximum possible Clocks as a lot is riding on the first "Navi" GPU.. Can't be slower than the competition, right.. However, most can simply tweak them for power savings in Wattman, if required.

Even then, these are only sipping 20-40 Watts more than their counterparts, which is a LOT better than 100+ Watts more, as was with Polaris/Vega.. (having no compute / RT HW helps..)

Based on their track record, and that Sapphire Reps statement; I am pretty sure the lower SKU beats Rtx2060 convincingly. But, the faster part, that AMD demoed, most likely trades blows with Rtx2070 & leads only in games AMD has generally better performance. Would like nothing better than to be proven wrong though..

Unfortunately whilst my hope is 48CU, my thinking is we are looking at 40CU part that is clocked wayyy past its efficiency window to compete with Turing.
I think Turing caught AMD by surprise, just like pascal did in perf.
I think Vega was more of a gaming turkey than they envisaged, and I also suspect N7 didn't hit the clocks that they were expecting...indeed AMD revised the performance benefit of 7nm down from 35% to 25% if I remember correctly.

Navi was designed under raja Kudori, we know he didn't execute very well and we also know Navi was delayed, I suspect Lisa Su pulled it in for a redesign, pulling in some features from next year's uarch and thereby enabling AMD to market a clean break from GCN.
I think next year will be the true next gen uarch from top to bottom, whilst Navi will take some features from it.

If this was a 48CU on new uarch and 7nm I feel we would be getting better than 2070 perf for 200+w.
Seems like they are having to clock the transistors off of it.

 

[happy medium]

So far we haven't seen anything apart from Strange Brigade, we do not know the SKUs even .

And yet you already jumped to conclusion that it is OVERALL slower than RTX 2070 .

Why don't we wait, eh? .

While I agree it will be slower than a 2070, it will be faster than a gtx1080/2060.

Which is good news that means it's big brother the Navi 58xx should be about 10 percent faster than a 2070.

 

[railven]

haha nvm

 

[Glo.]

It could be slower than RX 580 and faster than RTX 2080 Ti. It's a clean sheet* design, you can't guess it's performance! We've never seen anything like this before! Trust me!

*pending confirmation, but I have it on good authority it's a fact!

Unless Lisa Su blatantly lied about it in the keynote, Navi really is Cleen Sheet design.