Nvidia to use Samsung 14NM to manufacture Pascal GPUs this year

[CAT-THE-FIFTH]

I wonder if this means the GP107 and GP108 are not TSMC:

http://www.tweaktown.com/news/53375/...ung/index.html

[Bagnaj97]

It'll be interesting to see how the clocks compare between these and the 16nm chips. It should at least hint whether the comparatively low clocks of Polaris are the fault of the architecture or process. Hopefully it's not the process or Zen may not perform as well as hoped.

[scaryjim]

... Hopefully it's not the process or Zen may not perform as well as hoped.

I think Polaris uses an efficiency optimised version of the 14nm process, while Zen will use a clock speed optimised one, but I can't remember where I read that now!

[badass]

I think Polaris uses an efficiency optimised version of the 14nm process, while Zen will use a clock speed optimised one, but I can't remember where I read that now!

Which ironically could be contributing to why Polaris is less efficient with 2016 Graphics workloads than Pascal.

A clockspeed optimised process can be more efficient than an efficiency optimised process at higher clockspeeds. It looks like AMD wound the clock speed right up on the RX480 which is easy if you don't mind burning power whilst changing the process tgo a clock speed optimised one would probably take 6 months-1 year and increase the die size as well.

[CAT-THE-FIFTH]

Which ironically could be contributing to why Polaris is less efficient with 2016 Graphics workloads than Pascal.

A clockspeed optimised process can be more efficient than an efficiency optimised process at higher clockspeeds. It looks like AMD wound the clock speed right up on the RX480 which is easy if you don't mind burning power whilst changing the process tgo a clock speed optimised one would probably take 6 months-1 year and increase the die size as well.

It depends if Polaris was designed more with laptops in mind than desktops,which is where AMD is suffering quite a bit. There are indications that Tonga was more a part for Apple and Polaris 10 seems to be the GPU used in the PS4 Neo and is clocked at under 1GHZ. It does seem AMD might be getting trying to re-use designs wherever possible.



Polaris 10 peak effiency seems more around the 900MHZ mark or thereabouts.

[scaryjim]

Which ironically could be contributing to why Polaris is less efficient with 2016 Graphics workloads than Pascal.

I suspect it's just that nvidia's shaders really are that efficient - they've been well ahead on efficiency at the same node since Kepler, when they took AMDs "many small shaders" model and really ran with it. Prior to Kepler AMD had efficiency sewn up. It's one of the main reasons that nvidia have done so well for themselves in the last 4 years - they simply have a better engineered solution than AMD.

A clockspeed optimised process can be more efficient than an efficiency optimised process at higher clockspeeds. It looks like AMD wound the clock speed right up on the RX480 which is easy if you don't mind burning power whilst changing the process tgo a clock speed optimised one would probably take 6 months-1 year and increase the die size as well.

True enough - my suspicion is that AMD thought they'd get another couple of hundred MHz out of Polaris before the voltage curve started rising, but in the end (perhaps down to yields, or just the lack of process maturity) they needed more voltage than they were hoping. We all know that Polaris 11 will be going into laptops, but I can't help wondering if Polaris 10 will be too - laptops can take chips with up to ~ 100W TDP, and if they can drop the voltage from 1.1v to 0.85v at around 950MHz they should be able to get a full Polaris 10 comfortably below 100W.

What I don't think there's any doubt about is that Polaris is power-optimised, rather than performance optimised - I was really interested by the graph that CAT posted a few days ago* that showed Polaris with basically a flat voltage set at around 0.8v all the way up to around 900MHz, followed by a really sharp rise over the next 300 MHz. Drop Polaris to 900MHz @ 0.8v and you should get the best out of it - unfortunately, that best wouldn't've hit the performance target AMD set themselves....

* and posted again whilst I was writing this (ninjaed by CAT again...)!

[CAT-THE-FIFTH]

Well even if the top mobile P10 SKU runs at 900MHZ to 1GHZ,that would still make it close to R9 290 level IMHO,and that would be a decent jump up from the R9 M380X:

http://www.notebookcheck.net/AMD-Rad....154332.0.html

Voltage and power consumption should be much less too and I suspect boost clocks would be maintained much longer. Another thing is if they are targeting the same sort of voltages,it looks like the boost could be closer to 1150MHZ.

[scaryjim]

... http://www.notebookcheck.net/AMD-Rad....154332.0.html ....

125W mobile card They'd barely have to drop the voltage or clocks on a 470 to hit that! We need someone with an RX 480 or RX 470 to start messing with the voltages and power targets and see just how fast these things can run when you optmise for different power levels....

[watercooled]

You know, I've had this thought at the back of my mind that, assuming the 14nm process might be more low-power targeted with 16nm more suited to high performance GPUs, there's a chance we might see Vega on TSMC? We've recently found out that the Xbox One S processor is produced on 16nm at TSMC too.

Assuming this news is true it adds some more credibility to that? It's not like it's the first time we've heard it though, but AMD finally made the rumoured move to GloFo/Samsung with Polaris I guess...

[watercooled]

Hmm, I had posted a reply but it seems it got eaten. It was something along these lines:

I've been thinking for a while that, if SS/GloFo 14nm is indeed more well-suited to low power applications and TSMC 16nm to high-performance ones, then it's conceivable we might see Vega on TSMC's 16nm? We've also seen AMD working on this node through the new Xbox One S SoC. This rumour about Nvidia adds some credence to that theory, assuming it's true.

[badass]

I did some calculations here http://forums.hexus.net/hexus-news/3...ml#post3694054

Looking at that Graph, Polaris 10 looks perfect for a laptop. Wind down the base frequency to 900 MHz, keep the GDDR5x but down to 7GHz or so (I think there's some bandwidth starvation on the Desktop Polaris 10) and you've got an MXM that probably uses 70W or less.

[scaryjim]

Hmm, I had posted a reply but it seems it got eaten. ....

I can see it - not sure if it got stuck in moderation (again!) or something odd happened.

Interesting thought, but I'm more inclined to assume that there are different power optimised and clock speed optimised versions of the 14nm process - most other processes seem to have various different optimisations. Polaris would have been built on a power optimised one that means you can keep the power curve flat for a wide performance range but won't clock as high. Vega could be fabbed on one that generally requires more voltage anyway but that then clocks considerably higher.

I'd guess (I'm not a hardware engineer though, so this is a guess, and I'd love to get some expert knowledge) that most silicon basically has one voltage that it requires to run aty all, but that remains relatively flat until you start pushing the silicon beyond a particular speed - that's certainly what CAT's chart seems to show. It looks like Polaris on 14ff can hold its lower voltage over a much wider range than previous chips - maybe there's a version of 14ff that typically needs 1v to run, but can hold that voltage at much higher clock speeds? So at lower clock speeds it needs a higher voltafe and is less efficient, but once you get past 1200Mhz it'll run those higher clocks at a lower voltage, and be more efficient...

[watercooled]

Yeah both of my posts are there now.

I think a lot of it comes down to layout libraries too, there being a trade-off between power, area (cost) and speed. You don't hear much about it from the GPU companies but the likes of ARM talk about it a fair bit given how many different implementations there end up being of the same IP core, e.g http://www.anandtech.com/show/10347/...mis-unveiled/3

WRT the 14nm process, it's possible but so far I've only heard of the LPE and LPP variants, with LPE being an early one and AFAIK supplanted by the current LPP process. Compare that to 28nm where you had lots of variants, using amongst other things different gate materials, SiON and HKMG depending on the market.

For instance, we saw a big improvement in clock speeds when the likes of the Snapdragon 800 adopted HPM over LP used in the 600 as I think the two were very similar architecturally.