Exclusive: The AMD Inside Story, Navi GPU Roadmap And The Cost Of Zen To Gamers

[Panos]

There isn't a hope in hell that AMD will topple Nvidia any time soon, but the prospect of a refined Vega consistently trampling over Nvidia's midrange segment for reasonable money is probably enough for Nvidia to push the new RTX generation and reinforce the image that AMD are still at least a generation behind and are the inferior product.

And before the usual suspects here start bleating "but AMD ARE a generation behind and inferior", who really cares if you can get your 1440p gameplay at unnecessary Hz for a respectable price? Suddenly competition comes back.

Define a generation behind. Where is hardware async compute on NVidia cards? That's 3 generations behind.
Why NV cards in HDR tank so much, having the Vega beating the Titan Xp at 4K HDR gaming?
Why the above is happening, to lesser extend in DX12?

 

[KompuKare]

Actually, if they had a larger GPU budget...

...the cheapest thing would be to make a Vega type thing with most of the compute stuff taken out. And using a GDDR (5/5X/6) controller. And fab it at TSMC because they must have enough stuff now to no longer have to worry about the WSA.

Stubborn pride or wishful thinking but all the heavy compute stuff is probably why their gaming cards have run hot for ages now (that and not binning as heavily as they could - the 7870LE Tahiti thing being their last heavily binned card AFAIK), and they hardly have a big enough compute marketshare to be able to afford to make their gaming cards suffer.

Despite the high DP performance Hawaii was actually quite competitive with GTX780/780Ti even in terms of perf/area but even then if they'd cut out most of the computing stuff and binned a lot they'd have a far more competitive product. It was only when Maxwell came out with much of the compute stuff cut that AMD really fell behind in perf/watt.

 

[4K8KW10]

Yes, they need to bring VLIW back.

 

[subbytna]

Nvidia's mindshare will always give it the edge over AMD. It's like Apple and their cult following. When Nvidia did some shady stuff over the years, their fanboys couldn't see any wrong and still bought their cards. People will say, "ah but Nvidia make the best cards so I'll buy them no matter what"

That's like saying Jimmy Saville was a great entertainer, so you'll forgive him for what he did. I'm old enough to actually remember him actually being a good Sat night entertainer...

I've had many Nvidia cards from the venerable 8800 to a 1070 but I'm going to give my money to AMD now when I can as I'm fed up with Nvidias antics and also Intel's antics too.

 

[Rroff]

And fab it at TSMC because they must have enough stuff now to no longer have to worry about the WSA.

AFAIK even with the new agreement they still have to pay a certain percentage for all wafers they have made elsewhere to GF. Which means they are somewhat limited in what is cost effective for them to produce :s

 

[Panos]

That's like saying Jimmy Saville was a great entertainer, so you'll forgive him for what he did. I'm old enough to actually remember him actually being a good Sat night entertainer...

Yeah, trying to persuade my mother in law about this, was difficult to comprehend, when she also replied "different times back then".

 

[Gerard]

Polaris 10/20 aka Ellesmere, part of the Arctic Islands family with GCN 4 is a rather small chip.
We haven't seen a upwards scaled to 4096 shaders Polaris.

Vega is Vega, and belongs to different iteration of the GCN generations, and quite possibly is weaker than a potential big Polaris.

Polaris is in the past now, it was meant to be something entirely different and turned out to be a hot running power sucking mid range performer. With the amount of overhauling they would have to do its probably not worth their time.

 

[4K8KW10]

Polaris is in the past now, it was meant to be something entirely different and turned out to be a hot running power sucking mid range performer. With the amount of overhauling they would have to do its probably not worth their time.

What was Polaris meant to be?
And what are they doing now? Another fail in the form of Navi, just like Fiji, Polaris and Vega?!
Fiji - only 4GB of unsuitable memory;
Polaris - you described it;
Vega - underperforming at ultra-high gaming, but still satisfactorily good in APUs and for prosumers...

 

[Gerard]

What was Polaris meant to be?

A much better performer than what it turned out to be, read this.

https://www.hardocp.com/article/2016/05/27/from_ati_to_amd_back_journey_in_futility

 

[Mauller]

Always more than one side to a story and generally people don't accept the problem is themselves heh.

Oh no, the problems were not with raja. They are just not what people think they are or what they have been speculating.

 

[LePhuronn]

Define a generation behind. Where is hardware async compute on NVidia cards? That's 3 generations behind.
Why NV cards in HDR tank so much, having the Vega beating the Titan Xp at 4K HDR gaming?
Why the above is happening, to lesser extend in DX12?

Well, that's my point. All of this will be ignored just by Nvidia releasing a new series of cards when AMD have nothing officially scheduled for 18 months. That makes them a "generation" behind, and given choice performance numbers will show GTX 1000 was already the better of Vega and now RTX 2000 is landing uncontested, that shows AMD's "inferiority".

Vega was late to the party and had its issues against GTX 1000 and in its own right, if nothing happens on the gaming front until Navi then AMD might as well not bother with the gaming segment any more because Nvidia's mind and market share will be near total. And this is why I'm expecting a 7nm RX Vega with Hynix HBM2 at the end of this year.

Plus it'll nicely stretch the 7nm node's muscles ready for Navi end of next year. Get the bugs out, get them yields up.

 

[KentMan]

Oh no, the problems were not with raja. They are just not what people think they are or what they have been speculating.

So you met the guy who everyone blames and he gave you the "Real" side of the story that totally exonerates him from any blame... right sounds legit from Raja.. of course he has no blame from any of his time at AMD.

 

[D.P.]

Define a generation behind. Where is hardware async compute on NVidia cards? That's 3 generations behind.

Nvidia have had Hardware async compute since Kepler, 3 generations ago.

Why NV cards in HDR tank so much, having the Vega beating the Titan Xp at 4K HDR gaming?

Potentially drivers. TBH I'm not and it is also not consitent between games.

Why the above is happening, to lesser extend in DX12?

Nothign is happenign in DX12, Nvidia dominates, Vega64 is still a long way behind the 1080ti.

What happens is Nvidia has a very advanced front-end that can handle DX11 workloads. AMD doesn't. AMD's performance also tanks in many DX12 games - what happens there? Well the answer is DX12 is incredibly complex to extract performance form and require much more intricate knowledge of the underlying hardware. Something game developers don;t have. This is why DX12 is dead.

 

[D.P.]

Actually, if they had a larger GPU budget...

...the cheapest thing would be to make a Vega type thing with most of the compute stuff taken out. And using a GDDR (5/5X/6) controller. And fab it at TSMC because they must have enough stuff now to no longer have to worry about the WSA.

Stubborn pride or wishful thinking but all the heavy compute stuff is probably why their gaming cards have run hot for ages now (that and not binning as heavily as they could - the 7870LE Tahiti thing being their last heavily binned card AFAIK), and they hardly have a big enough compute marketshare to be able to afford to make their gaming cards suffer.

Despite the high DP performance Hawaii was actually quite competitive with GTX780/780Ti even in terms of perf/area but even then if they'd cut out most of the computing stuff and binned a lot they'd have a far more competitive product. It was only when Maxwell came out with much of the compute stuff cut that AMD really fell behind in perf/watt.

AMD strippoed out most of the FP64 support in Fiji. Fiji, Polaris and Vega (to date), all lack the DP support of Hawaii. Vega20 will be putting that back in.

 

[jigger]

Nvidia have had Hardware async compute since Kepler, 3 generations ago.


Potentially drivers. TBH I'm not and it is also not consitent between games.



Nothign is happenign in DX12, Nvidia dominates, Vega64 is still a long way behind the 1080ti.

What happens is Nvidia has a very advanced front-end that can handle DX11 workloads. AMD doesn't. AMD's performance also tanks in many DX12 games - what happens there? Well the answer is DX12 is incredibly complex to extract performance form and require much more intricate knowledge of the underlying hardware. Something game developers don;t have. This is why DX12 is dead.

Nvidia still haven't implemented async compute and I don't believe the chips have the functionality.

If DX12 is dead then so is Windows as a gaming platform. Maybe it's time to revert back to something like Mantle. Mantle performed batter than DX12 anyway.

 

[Mauller]

So you met the guy who everyone blames and he gave you the "Real" side of the story that totally exonerates him from any blame... right sounds legit from Raja.. of course he has no blame from any of his time at AMD.

I already said, the problem is not what people think, every bit of speculation is miles from the truth. The real issue is much more contrived and crappy than what people think.

 

[FoxEye]

I already said, the problem is not what people think, every bit of speculation is miles from the truth. The real issue is much more contrived and crappy than what people think.

Sorry but saying "You're all wrong; I can't tell you the truth because it's super secret," is the worst kind of trolling

 

[Mauller]

Sorry but saying "You're all wrong; I can't tell you the truth because it's super secret," is the worst kind of trolling

lol, I would love to outright say, raja gets far too much flak for something that was not his problem. But he also told me and my friend in trust and even if people don't want to outright believe, I would rather keep his trust.

His part of Vega that was his baby was the HBCC, which me and my friend too massive advantage of in our demo for siggraph. No Nvidia card can come close to what we did, software caching abd texture streaming just don't cut it in comparison.

 

[CAT-THE-FIFTH]

Well all I get is AMD basically said most of their R and D was diverted to development of Zen,and that probably means both hardware and software people,and that even Raja said AMD had thought GPUs were not worth pushing and that "drivers were not easy",which possibly is partly down to the fact I suspect ATI/AMD didn't make as much from the HD4000~HD7000 series as they expected.

I personally think that lead to a situation where RTG and the preceding parts of AMD which were tasked with graphics development had not enough funds and personal to have any contingency plans in place,so Vega not only had to use more expensive and slower HBM2 as a result,they probably shipped it with a number of features not working,since getting devs to use features needs resources too.

 

[JessieD]

AMD peoples when is Navi due please?