AMD 7nm GPU News and Rumours 2018/2019

[Beren]

Hey Folks,

Instead of polluting the Nvidia 2019 thread with AMD stuff how about we actually have one thread for AMD news and rumour and stop the bickering in that thread.

This is a thread for news/rumour and leaks of AMD cards. Can we please keep it to topic - unless there is a direct review mentioning an Nvidia card and a new AMD GPU keep opinions about other cards out of the thread please.

Mod's - Could you keep an eye open and delete anything not relevant so we actually have a thread of AMD related information.

It looks pretty certain that we are going to see Vega20 on 7nm in at least a limited release in the next three months for Pro Cards. It looks like wccf is trying to respin the same news as a rumour of a Consumer 7nm graphics card this year - at this point that seems unlikely, but I do think there is enough stacking up to suggest a 7nm card in Q1 2019, probably a mid sized Polaris replacement.

Do we think that the business that released a skunkworks threadripper on the down low is capable of releasing a gaming GPU based on Vega20? Is there anything out there at all that points to this?

I believe that both teams have a 1060/580 replacement in the works for a release Q4 2018 or Q1 2019. We are starting to see a few leaks, but I believe the quiet is because both are trying to not let anything slip to the other side on just how much power the mid range is going to get.

 

[Beren]

What's happening is, all players are overbidding themselfes to get capacity from TSMC. Smartphone socs are small and margins are ok, so they come first. Next will be serverstuff with very good margins but lower yield, like V20 this year, Zen2 probably Q1, nvidias volta HPC successor Q2. it will go on from high margin to lower margin when there's enough capacity. As for AMD, after V20 and Zen2 Server, Zen2 Desktop will come mid year, as CPUs have better margins than GPUs and they can earn more money in that space. Navi might then fit in in Q3.

You might be right for volume which would be a shame - I would expect a slight difference though, if I was AMD and I could make the cards I would sweet talk just enough wafers out of TSMC early to do a limited run, get some product out to the market and show what they were going to be capable of.

I think there is going to be something coming out of GlobalFoundries on 12nm. They have the process and the ability to produce Vega at all CU variants there.

 

[Beren]

I love how the press grabs a rumour and circulates it so much it sounds legitimate...

Other than a random ChipHell post, which doesn't seem to actually say a lot - Has anyone heard anything about the 12nm Polaris refresh? If I were TSMC I would have been pitching hard for any new GPU release so it makes a reasonable amount of sense to me - although I thought this chip was going to come about 9 months ago.

There appears to be two guesses - a straight re-spin of the 480/580 Polaris to a 680? with 10% higher performance but the same ram. There is also a rumour that they might be using the Xbox1X layout which would give a few more CU's.

There was Rumour in June that a GDDR 6 card was coming, but I doubt it is this one - more likely that will be Navi in Q2 2019.

So, has anyone seen anything around that is not directly coming back to a single uncorroborated chiphell source about a new AMD card?

 

[Beren]

It would, but there is absolutely nothing to indicate foundry either - Actually it would make much more sense for this to be on the GF 12nm process. I was being dumb earlier.

 

[Beren]

It is probably a completely pointless exercise but has anyone graphed transistor count to performance to calculate the most efficient current chip design? The great disappointment with Vega 64 from a pure gaming perspective was that at 12.5B transistors it could have reasonably have been expected to perform at 1080ti levels - but because it was trying to be all things to all men it could not use all it's power in gaming and lagged massively behind.

At 5.7B Transistors the RX580 is actually a far more effective pure gaming card for AMD in terms of perf/transistor. A larger chip on a better process and faster Ram would potentially perform significantly better. It wont happen - but I think it was possible.

 

[Beren]

To be fair the 580 is a tiny die, they could just double it to 400mm but they also wanted that $230 price point.

If they do a 250mm die on 12nm they would get 30% more compute and also 10% higher clocks would net a significant upgrade, certainly within reach of 1070 performance. But it would be a redesign.

 

[Beren]

Yes it is a keynote - but I am definitely expecting her speech to be punctuated by either the launch or announcement of at least one product. CES is too good an opportunity to miss.

This is CES so I am not expecting it to be about enterprise kit.

There are two possibilities to me. The way the announcement is worded focusing on VR and Gaming... it might be possible that this will be about Console and not PC.

I think it is far more likely to be a glowing conversation about 7nm the continuing journey about Zen, the confirmed progress to 7nm GPU's but I think there is likely to be a 'surprise' product.

 

[Beren]

So this is Polaris++... Didn't AMD finish Intel's book on how to do it? First establish market dominance, THEN implement incremental gains.

The annoying thing is that for 1080p gaming the 680 is actually likely to be a really great card, I just wanted them to do better.

 

[Beren]

We are seeing some really aggressive pricing on Vega56 right now.

I am only expecting a 12% jump for the RX590/680/690 but the pricing on Vega is making me wonder if there is enough juice in the upgrade to get it close to Vega56 territory... but that would mean nearly 40% improvement... so I am probably dreaming.

 

[Beren]

480 to 590 should be something like a 20-25% improvement. A Vega 56 is at least 50% faster than the 480.

Navi... we are all in the dark, in terms of roadmap it makes sense for Navi to be a Polaris replacement - and as a 7nm shrink and with architectural improvements you can probably expect a 30-50% increase in performance over the RX590. Possibly mid next year. I would guess we are going to get Vega64 performance at a much lower price point, but that is pure conjecture.

If you have the cash I would seriously consider the Vega56 it is a genuine and serious upgrade to your 480.

 

[Beren]

Probably a dumb question... But based on the pictures of the Vega 7nm chip it seems it is going to have a whole lot more transistors than a Vega64. Since GCN is limited to 4096 SPs and Vega already has that, what is the rest of that silicon real estate going to be doing?

 

[Beren]

Why do we get ourselves so in a twist over this stuff? We absolutely don't know, but I have already started to guess so waffle for waffles sake...

Instead of concentrating on a couple of mentions on a slide released 6 months ago about a card that will probably be released in almost a year - what is the design brief going to be on this architecture?

AMD have publicly declared Navi is a Mid Tier card - exactly how they define Polaris. They already have a price bracket in mind for that probably at the $200-$300 price point.

There has been significant noise about Navi going into the PS5. I believe that one of Sony's must haves will be the ability to game at 4k 60Hz at medium to high settings. I think that will mean hitting close to 1080 levels of performance.

They are also going to have to be able to create the chip to hit a console defined price point which I think also backs up the $2-300 number.

Therefore based purely on what I think the Engineering design brief has to be I think we are going to be looking at performance sitting somewhere around 1070-1080 levels for something like $300.

Just to chuck a guess on record so that I can come back and gloat if I get close... Anyone else fancy a guess?

The top Navi card will be 140mm2 die on a 7nm process with 52 CU's clocked at 2Ghz matched with 8 gigs of GDDR6 a launch price of $289 and performance within 5% of a 1080.

 

[Beren]

https://wccftech.com/amds-vega-20-gpu-early-engineering-sample-benchmarked-in-final-fantasy-xv/

AMD’s upcoming Vega 20 GPU was recently spotted (courtesy of Videocardz) in Final Fantasy XV benchmark db and represents one of the first data points we have in how it will fare in gaming. It goes without saying that this is a very very early preview and will almost certainly not be indicative of the final product. This is in all likelihood an early engineering sample so clocks and drivers are expected to improve further as well.

AMD’s Vega 20 GPU benchmarked in Final Fantasy XV – trades blows with a GTX 1080
Before we begin, it’s worth pointing out that we have no information on whether this is the RX version or not. In fact, as we have reported earlier, we believe AMD will not be rolling out an RX version of their 7nm Vega 20 at all and will instead move directly to the next generation architecture. This is most likely the Instinct variant – which isn’t really designed for gaming of course. The PCI-E ID is 66AF:C1 and performs roughly equivalent to a Pascal GeForce GTX 1080.

Not great news if this is actually indicative of node improvements for 7nm... although obviously massive bags of salt to be used here.

 

[Beren]

Hence bags of salt. However, just because a compute card can compute at a vast tflop level does not mean it can game. I suspect the first instinct card will have the same 64 Vega CUs it will just have a lot of extra stuff on top as well.

 

[Beren]

Holy Context Batman. Good spot, Competetive in high end depends entirely on definition and says nothing about beating Nvidia just competetive with.
The 590 is looking good though.

 

[Beren]

Nothing is certain. The signs point to a Polaris replacement still though. Mid range for 7nm is still likely to be quick mind.

 

[Beren]

The specs make perfect sense. They want a capable 4k console with strong VR capabilities - potentially even the ability to run two VR sets off one console which would be absolutely epic for a whole heap of reasons.

We are looking at the final iteration of GCN on 7nm. They have always had the ability to produce a chip at this level.

 

[Beren]

The 7nm Vega shrink is very specifically an enterprise card. It is vastly different to a vega64.

I would love to think there's going to be a vega 64 refresh with 30% higher clocks, but I seriously doubt it at this stage.

 

[Beren]

So to all those who have commented and seem to be saying its ok for this sort of thing to happen in presentations, let me ask you this and try to answer honestly.

If it was NVidia pulling a stunt like this and AMD were the ones losing out, wouldn't you be a little put out by it and maybe comment on this forum that we all like to frequent.

Definitely - I will add the slight caveat, this is stuff technical enough that I don't know if they have definitely fiddled with something. I would love to see proper commentary from someone using this tech in the wild who has seen both new technologies.