AMD - Piledriver chitchat

[watercooled]

I think I might have posted something along these lines before but I wonder where the new console APUs will be made?

ISTR it will be GloFo/SS as that's where Polaris is made and using TSMC would mean porting it to their process. But if that happens, I wonder what AMD will use TSMC for, if anything? It's a potentially risky move on the GPU side as they're counting on GloFo/Samsung competing favourably with TSMC for their higher power process, and already it doesn't look like AMD gained much from the 28nm TSMC > 14nm GloFo transition (I'm trusting they didn't regress in architectural efficiency). But let's just hope that's early yields...

[kalniel]

and already it doesn't look like AMD gained much from the 28nm TSMC > 14nm GloFo transition (I'm trusting they didn't regress in architectural efficiency). But let's just hope that's early yields...

Is that true? I had the impression Polaris was delivering very impressive perf/watt gains over hawaii.

[scaryjim]

Is that true? I had the impression Polaris was delivering very impressive perf/watt gains over hawaii.

Indeed, higher clocks, lower power draw ( ~ 160W vs ~ 250W), very similar performance.

Don't forget that Fiji was already a big perf/watt gain over Hawaii, and AMD have never compared Polaris to Fiji in terms of perf/watt. Also remember Nano was ~85% the performance of Fury X for ~70% of the power budget. What we're seeing released with Polaris 10 is essentially the Fury X/Fury iterations of that arch - the ones pushed up to the limit of the silicon to get the best performance, rather than the best efficiency. It'd be interesting to see what Polaris 10 could do - efficiency-wise - if you dropped the voltage & power target to maintain ~ 1100MHz...

[CAT-THE-FIFTH]

AMD launches new Polaris based Radeon Pro WX cards.

The top one,ie,the WX 7100, seems to have similar specs to the RX470:

http://wccftech.com/amd-radeon-pro-w...kstation-card/

It is single slot and has over 5TFLOPS - if the RX470 is close to that,it should not be far off a RX480. Plus TDP must be reasonably lower if the card is single slot.

Interestingly there is a 1792 shader WX 5100 which is single slot and bus powered,with a 256 bit memory controller and looks more like a single slot HD7750.

Then there is the WX 4100 which has 1024 shaders and is low profile.

The WX 5100 looks very impressive indeed.

[watercooled]

Is that true? I had the impression Polaris was delivering very impressive perf/watt gains over hawaii.

Indeed, higher clocks, lower power draw ( ~ 160W vs ~ 250W), very similar performance.

Don't forget that Fiji was already a big perf/watt gain over Hawaii, and AMD have never compared Polaris to Fiji in terms of perf/watt. Also remember Nano was ~85% the performance of Fury X for ~70% of the power budget. What we're seeing released with Polaris 10 is essentially the Fury X/Fury iterations of that arch - the ones pushed up to the limit of the silicon to get the best performance, rather than the best efficiency. It'd be interesting to see what Polaris 10 could do - efficiency-wise - if you dropped the voltage & power target to maintain ~ 1100MHz...

I said not much, not nothing. Comparing against even the Fury version of Fiji it's a fairly modest improvement - yeah it's a big improvement over Hawaii but bear in mind a lot of that is from architectural improvements already seen in Fiji (like I say I very much doubt they've regressed in architectural efficiency since Fiji). https://www.techpowerup.com/reviews/AMD/RX_480/25.html

They don't include the Nano in those charts but looking at the Nano review it's not hugely different, but as you say that's basically best case on 28nm vs worst-case on 14nm so not a life-for-like comparison.

On the other hand Nvidia seem to have gained a large improvement moving to 16nm despite a significant increase in clock speed.

I agree though it will be interesting to see how Polaris/Vega do with future lower-voltage parts.

[CAT-THE-FIFTH]

I said not much, not nothing. Comparing against even the Fury version of Fiji it's a fairly modest improvement - yeah it's a big improvement over Hawaii but bear in mind a lot of that is from architectural improvements already seen in Fiji (like I say I very much doubt they've regressed in architectural efficiency since Fiji). https://www.techpowerup.com/reviews/AMD/RX_480/25.html

They don't include the Nano in those charts but looking at the Nano review it's not hugely different, but as you say that's basically best case on 28nm vs worst-case on 14nm so not a life-for-like comparison.

On the other hand Nvidia seem to have gained a large improvement moving to 16nm despite a significant increase in clock speed.

I agree though it will be interesting to see how Polaris/Vega do with future lower-voltage parts.

Look at the Radeon Pro parts! A bus powered part based on Polaris 10.

[DanceswithUnix]

Interesting article on Anandtech about the Power 8 CPU.

A single core has pretty much the hardware to run two threads before it starts dipping properly into SMT, making it look like Piledriver turned up way past 11!

http://www.anandtech.com/show/10435/...-power8-part-1

[watercooled]

Yeah that article made for some really interesting reading.

TBH I don't at all think the CMT concept behind the Bulldozer family was a bad idea; of course it has performance trade-offs with more conventional designs but as far as I can tell it's not the main problem the family had. Other things like excessive branch misprediction pentalty, lower IPC, high L3 cache latency, and of course power management were more major contributing factors. In terms of per-core scaling though it really did impressively well in many workloads, the decode width being effectively limited to 2 per core didn't seem to matter a lot of the time (but didn't Steamroller change that to 4-wide by giving each integer core its own decoder anyway?).

[Socras]

is that the "next gen memory"? ssg?

[CAT-THE-FIFTH]

Someone mentioned this on SA forums:

http://dl.acm.org/citation.cfm?id=2934586

Does that mean Bristol Ridge and Stoney Ridge use FIVR??

[watercooled]

Interesting. Maybe they're trialling it on a mature core and node before rolling it out to Zen? Having said that, the APU platform would stand to benefit more given its application in mobile devices anyway.

[CAT-THE-FIFTH]

LMAO:

http://www.hardocp.com/article/2016/..._in_raw_data/1

Seems to conveniently forget this:

http://www.geforce.com/whats-new/art...nvidia-vrworks

Oh its early access too:
http://www.pcgamesn.com/nvidia-showc...ashed-on-steam

[watercooled]

Something I didn't expect - the Xbox One S uses a 16nm processor: http://www.eurogamer.net/articles/di...tech-interview

So AMD is indeed using TSMC 16nm for something!

Given how cost sensitive this segment is, this is quite interesting - I thought 16nm was going to be financially unattractive for things like this for some time? Or maybe the small performance bump and lower power consumption was worth it for MS?

[CAT-THE-FIFTH]

Posted on SA forums today.

The shader count for the Zen APU looks to be 768 shaders.

[watercooled]

I thought the latest from AMD was still claiming late 2016 availability for Summit Ridge?

[CAT-THE-FIFTH]

I thought the latest from AMD was still claiming late 2016 availability for Summit Ridge?

The slides are not for server availability.