AMD Polaris slides leaked and first cards will be released in mid-2016!

[CAT-THE-FIFTH]

The whole slide deck has been released:

http://videocardz.com/58031/amd-anno...ecture-gcn-4-0

The early sillicon is showing a 60% improvement in performance/watt over Maxwell,and thats six months before release. That would equate to a GTX980TI consuming between 125W to 150W,which would be one PCI-E power connector!

[kalniel]

If the improvement scales to the upper chips.. their slides are showing geforce 950 class card performance ie quite near the low end.

[CAT-THE-FIFTH]

If the improvement scales to the upper chips.. their slides are showing geforce 950 class card performance ie quite near the low end.

I have a GTX960 which is quite close to a GTX950. Even if you to keep to that same power consumption of around 100W(as measure by TPU),a 60% improvement(as indicated in the slide) would easily mean around GTX970 level performance on a card like that,which is well under the 150W limit for a single PCI-E power connector. Even a lower 50% improvement would probably be around the same effect.

[DanceswithUnix]

Hmm, yet more laptop directed power reduction statistics. I was hoping for a doubling in performance for a given price but left none the wiser here. Still, nice to see they have working silicon to benchmark with.

Interesting choice to do power measurements vs a GTX 950, is that a good card for efficiency vs the other Nvidia cards using that chip?

[kalniel]

I have a GTX960 which is quite close to a GTX950. Even if you to keep to that same power consumption of around 100W(as measure by TPU),a 60% improvement would easily mean around GTX970 level performance on a card like that,which is well under the 150W limit for a single PCI-E power connector.

I mean they don't say anything about the perf/watt improvement scaling linearly or applying in both directions. You could take the same performance for a 40% reduction in power as shown on the slides, but it's harder to take the same power and hope for a 40% improvement in performance.

[DanceswithUnix]

I mean they don't say anything about the perf/watt improvement scaling linearly or applying in both directions. You could take the same performance for a 40% reduction in power as shown on the slides, but it's harder to take the same power and hope for a 40% improvement in performance.

Depends if you can afford to increase the size of the die/transistor count by 40% which should scale you nicely back up. You are right that just turning the clocks & voltage up is unlikely to get you very far.

Edit to add: My hope with this generation is that the transistor count will roughly double. From 28nm to 14nm is two full process nodes which in days past would mean a quadrupling in density, but things aren't so good these days but double would be nice *before* you start using HBM2 to bump the performance further.

[kalniel]

Depends if you can afford to increase the size of the die/transistor count by 40% which should scale you nicely back up. You are right that just turning the clocks & voltage up is unlikely to get you very far.

I guess transistor count does scale fairly linearly, but still some overheads.

[CAT-THE-FIFTH]

I mean they don't say anything about the perf/watt improvement scaling linearly or applying in both directions. You could take the same performance for a 39% reduction in power as shown on the slides, but it's harder to take the same power and hope for a 39% improvement in performance.

Its actually a 62.7% improvement in performance/watt over the GTX950. Also you are over-thinking it. If you looked at the GCN cards from the HD7750 to the HD7870 there was almost the same performance/watt,so for the lower end parts to the midrange parts so I expect it will be the same. The same goes for Maxwell - the GTX950,GTX960 and GTX970 are very similar.

So even going as over-pessimistic as you can,at even a 40% improvement in performance/watt over Maxwell midrange parts,you are still getting GTX970 level performance in a GTX960 class card.

Edit!!

Nvidia managed to get like a 40% improvement with the GTX980 when compared to the GTX780 on the same node.

So,I think 40% to 60% improvement using a 14NM node(Hybrid 20NM with Finfets) and a modified uarch is not at all unreasonable.

[kalniel]

Its actually a 62.7% improvement in performance/watt over the GTX950.

How do you get 62.7% improvement over the 950?

[DanceswithUnix]

How do you get 62.7% improvement over the 950?

I think he went this way...

(140-86)/86 = 0.627

I read that as "950 is 62.7% worse power consumption than the new card."

[kalniel]

I think he went this way...

(140-86)/86 = 0.627

I read that as "950 is 62.7% worse power consumption than the new card."

Yes, awkward! You could say the new card uses 0.614 of the power of the 950 I guess.

But simpler to say it's a 40% improvement (reduction in power for the same performance) compared to the 950. (140-86)/140 = 0.385 or roughly 40.

[EndlessWaves]

The anandtech article is well worth a read as it goes into more detail about various bits and pieces:
http://www.anandtech.com/show/9886/a...u-architecture

Interesting choice to do power measurements vs a GTX 950, is that a good card for efficiency vs the other Nvidia cards using that chip?

It's slightly worse than most of the others, but only slightly.

I think he went this way...

(140-86)/86 = 0.627

I read that as "950 is 62.7% worse power consumption than the new card."

Those figures are system power consumption according to the Anandtech article (which also mentioned they a live demonstration produced 88.1W and 150W figures from an electricity meter measuring at the wall).

The slide claims an i7-4790K and 4x4GB DDR4-2600, although I didn't realise the 4790K had a DDR4 controller.

I'd suspect something like a 90W/50W split for the 950/rest of system. Assuming similar system power draw from the system on the AMD setup that suggests the baby Polaris is pulling 36W. If we err on the side of caution and assume 80W/60W that still gives us 46W vs. 80W.

If the 120mm² estimate is accurate then this isn't just a low clocked but large chip (efficient but expensive to produce) - that's a smaller die than the GTX 750ti.

[DanceswithUnix]

The slide claims an i7-4790K and 4x4GB DDR4-2600, although I didn't realise the 4790K had a DDR4 controller.

The video said DDR3 which sounds like the one to believe. It lists motherboard used etc in some detail.

Edit to add: It is intriguingly exact though for a typo, so I guess they tested *something* with DDR4 2600 ram. Wonder if that is a boring old i7 or something else they had in the lab

[kalniel]

The anandtech article is well worth a read as it goes into more detail about various bits and pieces:
http://www.anandtech.com/show/9886/a...u-architecture

And the full slide set shows AMD don't know the different between 'less' and 'fewer'. Poor form.

[CAT-THE-FIFTH]

The Polaris sample is running at 850MHZ at 0.8V it appears. Interesting both TSMC and GF are being used. The GF process is meant to be denser but has worse characteristics with regards to power consumption.

Edit!!

The chip is also on a GF process too,plus there is no confirmation if it uses HBM2 or GDDR5/GDDR5X too.

[EndlessWaves]

And the full slide set shows AMD don't know the different between 'less' and 'fewer'. Poor form.

You're talking about "Discrete cards with less power connectors"?

It's an abbreviated sentence in a bullet point list. Unless you know what point was being made you can't say whether less or fewer would be the more common usage for the full point being made. There are many other examples of grammar you wouldn't use in other contexts such as "First in production 2012 at Intel" and "But doesn't help active states and would hurt performance".

From the same slide "FinFET enables performance per/watt innovation" is more unusual in it's use of "performance per/watt". Although I have the feeling I've seen that usage before so it may be intentional instead of incorrect, or maybe just a common mis-writing.