AMD: Speculation, Rumors, and Discussion (Archive)

[Kaotik]

I think there are some pretty unrealistic expectations regarding Polaris 10 in this thread.

Polaris 10 is a 232mm^2 chip that is set to replace the lower-end Hawaii and GM204 variants (R9 290 and GTX970) within the same performance levels but at substantially lower price and power consumption.
AMD's statements about Polaris 10 have been rather consistent, which is to make the "minimum-VR" performance level more affordable.

I doubt very much that even the highest-end Polaris 10 will be able to significantly surpass R9 390X levels of performance, much less reaching the GTX 1070 or even the 980 Ti.
Polaris 10 will be about taking down the GM204 which is a current sales champion and phasing out Hawaii which is too old, too big, too power-hungry and is probably giving them too little margins.
Reaching 980 performance might be a bonus but may not even be a requirement.

In fact, a very possible reason for AMD not rushing any further news until Computex is the fact that nvidia didn't (paper)launch anything that will go directly against their 2016 lineup.

You seem to think there's going to be just one Polaris 10 model, while with 2 chips, they pretty much have to release at least 3 of them, and since they've been consistent about Polaris 10 and VR min spec, it should be quite clear that slowest Polaris 10 hits the VR min spec. From that, add 2 models with 20 % performance between each and you're already at 980 Ti. Of course, it could end up a little under, but so might 1070.

 

[Deleted member 13524]

Yes, but if Fury is 300 watts and AMD said that Polaris is 2 times perf/watt (Koduri said 2,5) then, there should be a chip at 150 watts with that performance, if not they will have failed at fulfilling the expected performance target they themselves created with their marketing buzz.

First off, Fury has the HBM advantage and I'd bet Raja Koduri wasn't talking about HBM solutions when he was making those comparisons because no Polaris has HBM. Most probably, he was talking about current GDDR5 solutions such as Hawaii or Tonga. I mean we could even move the goalposts even further and suggest he was talking about the Nano, for which we'd have a miraculous chip with ~GTX980 performance at less than 75W.

Secondly, why must there be a 150W solution between Polaris 11 and 10?
Polaris 11 is bound to be a sub-75W part, probably with no PCI-E power connectors at all. If Polaris 10 is 2.5x more efficient than a R9 390X, then it'll be 275W/2.5 = 110W.
Maybe they could push the clocks up and hit Fury performance levels within 150W, but at the same time they would:
1 - Decrease chip yields
2 - Increase the cost for PCB and power regulators
3 - Completely cannibalize current Fiji solutions even if they have huge price cuts.
4 - Shrink the market at which Vega will be targeting in the future

You seem to think there's going to be just one Polaris 10 model

No, I fully expect to see Polaris 10 Pro and Polaris 10 XT graphics cards. I just don't expect the XT model to significantly exceed the performance of a R9 390X, if at all (which is already uncomfortably close to a Fury BTW).

 

[3dilettante]

If Polaris 10 is the big chip of the two, and is supposedly around 230mm2, where's a chip in the range between 230mm2 and the assumed huge Vega? Polaris 12 next year? A Vega 10 and 11?

 

[gamervivek]

According to TPU's latest review, at 4k, Fury X is only 26% faster than 390X, ref. 980Ti only 21%. If Polaris 10 has a Hawaii like configuration with the arch. improvements that AMD have been talking of they only need a ~30% clockspeed boost over a 390X, which is way overdue considering that Maxwell was doing better than that on 28nm, to easily get clear of Fury X, 980Ti and Titan X.

Unrealistic would be expecting a 256-bit bus to perform well at 4k and the clockspeed boost considering the leaks we've had were 800 and 1050Mhz.

If Polaris 10 is the big chip of the two, and is supposedly around 230mm2, where's a chip in the range between 230mm2 and the assumed huge Vega? Polaris 12 next year? A Vega 10 and 11?

There are two Vega chips just like Polaris, Anandtech confirmed it.

 

[Razor1]

http://www.techpowerup.com/reviews/Gigabyte/GTX_980_Ti_XtremeGaming/23.html

Talking about this one?

 

[Deleted member 13524]

If Polaris 10 is the big chip of the two, and is supposedly around 230mm2, where's a chip in the range between 230mm2 and the assumed huge Vega? Polaris 12 next year? A Vega 10 and 11?

Yup. gamervivek already mentioned it but here's the quote:

Meanwhile AMD has also confirmed the number of GPUs in the Vega stack and their names. We’ll be seeing a Vega 10 and a Vega 11.

There might be a Vega 11 that places itself between the GTX1070 and the reference GTX1080, and then a Vega 10 that will counter the rumored GTX 1080 special editions with >2GHz clocks (which may eventually become the GTX1080 Ti).

 

[Anarchist4000]

Two Vegas, at least one of which should be huge and the only advertised feature being HBM2. It's possible they're both huge but one features FP64 for pro setups. Other possibilities seem to be 490 and Fury tiers respectively.

I still can't help but think a pair of Polaris 10 on an interposer with even 4GB HBM1 (2GB each but actually shared via Onion) would be an interesting part. For Example a current FuryX replaced by 2 FINFET dies that still come in under 200W. I'd have to check the specifics on Onion, but it's possible it only connects two devices. So APU works, dual cards work, CPU + discrete over PCIE probably works. Then just step that up with the smaller Vega at around 300mm2 with HBM2. That configuration would definitely be pushing some limits.

 

[pMax]

I still can't help but think a pair of Polaris 10 on an interposer with even 4GB HBM1 (2GB each but actually shared via Onion)

...uh? GPUs has already a PCIE connection and an IOMMU i think, so they can interconnect without hassle. Why would you need Onion (coherent) bus... between GPUs other than a fast PCIe link?

Plus, the NUMA-like system you're building would suffer heavily if textures needed by one GPU stays on the other half of the memory...

 

[CSI PC]

According to TPU's latest review, at 4k, Fury X is only 26% faster than 390X, ref. 980Ti only 21%. If Polaris 10 has a Hawaii like configuration with the arch. improvements that AMD have been talking of they only need a ~30% clockspeed boost over a 390X, which is way overdue considering that Maxwell was doing better than that on 28nm, to easily get clear of Fury X, 980Ti and Titan X.

Unrealistic would be expecting a 256-bit bus to perform well at 4k and the clockspeed boost considering the leaks we've had were 800 and 1050Mhz.



There are two Vega chips just like Polaris, Anandtech confirmed it.

Lets be honest,
using 4k as a performance point is mostly irrelevant because the fps and frame times are just too weak.
Case in point, look at Witcher 3 at 4k or as a more equal playing field but does not push the boat out GTA V.
GTA V is still not acceptable in terms of playability at 4k.

From a performance mark perspective, it may be useful to compare an architectures performance trend as it goes up the resolutions including 4k.
I appreciate the argument changes slightly if playing in mGPU, but this adds a whole other set of variables that skew calculating a GPU performance.
Cheers

 

[SimBy]

It's reasonable to expect Nvidia will target 1070 minus 20% performance with its mainstream GPU. They have to match or beat that. And pricing does the rest. Simple as that.

So slightly faster than 980 for $250. That's you target AMD. But please, do aim higher.

 

[Anarchist4000]

...uh? GPUs has already a PCIE connection and an IOMMU i think, so they can interconnect without hassle. Why would you need Onion (coherent) bus... between GPUs other than a fast PCIe link?

Plus, the NUMA-like system you're building would suffer heavily if textures needed by one GPU stays on the other half of the memory...

To share the actual memory pool as opposed to duplicating everything. Goal being to texture some stuff from the other memory pool. Two chips sharing an interposer with a ridiculously wide link could probably pull it off. I thought they were replacing Garlic with Onion3, so it seemed practical. Also seems likely they'd have support built in for it already if Zen+Polaris MCMs were in play.

 

[Deleted member 13524]

So slightly faster than 980 for $250. That's you target AMD. But please, do aim higher.

If it's slightly faster than a 980 then it certainly won't be $250. If it's $250 then it's certainly not faster than a 980.
Having both of those at the same time would be just unnecessarily aggressive pricing for AMD.

 

[Razor1]

There doesn't seem to be a dedicated thread to Vega, so I'm guessing this belongs here

http://videocardz.com/59808/amd-vega-gpu-allegedly-pushed-forward-to-october

 

[Malo]

The Vega thread here

 

[Silent_Buddha]

I think there are some pretty unrealistic expectations regarding Polaris 10 in this thread.

Polaris 10 is a 232mm^2 chip that is set to replace the lower-end Hawaii and GM204 variants (R9 290 and GTX970) within the same performance levels but at substantially lower price and power consumption.
AMD's statements about Polaris 10 have been rather consistent, which is to make the "minimum-VR" performance level more affordable.

You might want to look again.

http://wccftech.com/amd-polaris-architecture-vr-minimum-spec/

In there is a quote from AMD

AMD has just completed the shrink to 14 nanometer [with Polaris Architecture]. What this means is, and this is where it comes home to everyone in this room, is that we can produce GPUs that will run the minimum spec of VR at a lower cost, in larger volume, consuming less power and running faster. That means in the second half of this year and going forward more people will be able to run those headsets which will make a larger for everybody in the room [VR Industry]. – AMD’s Roy Taylor

The Polaris architecture was designed to bring VR to the masses. That could mean either Polaris 10 or 11.

http://www.gamespot.com/forums/syst...polaris-10-to-replace-fury-x-33136353/?page=1

Their roadmap also implies that the low end Polaris 10 will replace the high end R9 300 series while the top Polaris 10 will replace the Fury series (X and non-X)

http://wccftech.com/amd-polaris-10-gpu-pictured/

Polaris 10 also demo'd running Hitman DX12 at greater than 60 FPS at 2560x1440 which is actually comparable or faster than Fury X. I can't find the link again since it was something I read months ago, but one of the prominent tech sites had a news blurb while they were at an AMD event that showcased Polaris 10 running some games at 2560x1440 faster than he'd ever seen Fury X run those games. Or maybe it was a twitter post.

Of course, there's always the possibility that final shipping hardware won't be as fast as the hardware they demo'd, so we'll have to wait until the actual unveiling and benchmarks before coming to any conclusions.

Regards,
SB

 

[Razor1]

http://www.gamespot.com/forums/syst...polaris-10-to-replace-fury-x-33136353/?page=1

Their roadmap also implies that the low end Polaris 10 will replace the high end R9 300 series while the top Polaris 10 will replace the Fury series (X and non-X)

It also implies Vega replaces Polaris too, HBM 2 cards in low end and midrange bracket sound right? I don't think those blocks mean much other than this is what is coming out at this time.

http://wccftech.com/amd-polaris-10-gpu-pictured/

Polaris 10 also demo'd running Hitman DX12 at greater than 60 FPS at 2560x1440 which is actually comparable or faster than Fury X. I can't find the link again since it was something I read months ago, but one of the prominent tech sites had a news blurb while they were at an AMD event that showcased Polaris 10 running some games at 2560x1440 faster than he'd ever seen Fury X run those games. Or maybe it was a twitter post.

Frame rate locked and at what settings? Settings weren't mentioned.......

 

[Ext3h]

It also implies Vega replaces Polaris too, HBM 2 cards in low end and midrange bracket sound right? I don't think those blocks mean much other than this is what is coming out at this time.

Yes and no.
Vega (respectively the new arch) will also replace Polaris at some point (even though probably not prior to Navi), but HBM2 on midrange and lowend probably won't happen any time soon. Interposers, as well as the extra thin slices in the HBM stack, are still far too expensive for that. We will probably have to wait until AMD finds a way to get sufficient memory into the package without resulting to stacking, or at least for further reducing the cost involved with that.

So, unless AMD has already an yet unknown GDDR5(X) relative of the known two Vega chips planed, no midrange Vega for the next 1 1/2 or 2 years.

 

[CSI PC]

There doesn't seem to be a dedicated thread to Vega, so I'm guessing this belongs here

http://videocardz.com/59808/amd-vega-gpu-allegedly-pushed-forward-to-october

Considering how tight HBM2 memory is at the moment and that is with Samsung being 1st into mass production surely means they are going to have to release this with either GDDR5X or more traditional (doubt they will go with HBM1).
I would expect NVIDIA has probably a high order contract with Samsung for HBM2, so not sure where AMD would find a large amount, and if they did (lets say SK Hynix can manage this) at what cost is this going to add to Vega.
The plan was to wait until HBM2 was more available to masses for cost (was words to that effect used either in a presentation or interview) and why it was pushed back to 2017.

On the plus side, fingers crossed they do release even if with GDDR5X for now, as that will also push NVIDIA to release their big card; great all round for fans of both companies.
I guess comes down whether they both have back-up plans to go with say 12Gb/s GDDR5X if they needed to release their large GPUS early Q4 to keep costs down.
Cheers

 

[Razor1]

Unless AMD has planned for Vega to use GDDR5x, I don't think they can swap in GDDR5x as they wish, cause that is a fairly large change lol not to mention a waste of silicon in for the bus. What I think if this rumor is real, Vega will be announced in Oct. and reviews will come out shortly after that with ES's and buy-able a couple months later which puts it at end of 2016 where HBM 2 should be well into mass production.

 

[LordEC911]

Unless AMD has planned for Vega to use GDDR5x, I don't think they can swap in GDDR5x as they wish, cause that is a fairly large change lol not to mention a waste of silicon in for the bus. What I think if this rumor is real, Vega will be announced in Oct. and reviews will come out shortly after that with ES's and buy-able a couple months later which puts it at end of 2016 where HBM 2 should be well into mass production.

The rumor definitely has something to it, there is something in the works that they are expecting in Q4.

Here is what I said about it last week-

So hearing some things... sounds like something is expected Q4.
I'm guessing it is the baby Vega which is the third chip Charlie was talking about? In my head baby Vega is Tahiti/Tonga sized, aka <400mm2.
Are we expecting both Vega's to have HBM2? 256bit with GDDR5x or 384bit with current 7Gbps GDDR5 would likely be enough with +300Gbps.

I honestly can't see them going with a 384bit bus though...

http://semiaccurate.com/forums/showpost.php?p=261209&postcount=771