• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

Realistic performace expectations from both NVidia and AMD 2016 GPU's

bru

bru

bru

Soldato
Joined
21 Oct 2002
Posts
7,361
Location
kent
OK we have had lots of guesses and wild ideas about just how fast the new cards are gong to be from both AMD and NVidia, but I thought it might be an idea to try to look at things a bit more sensibly.

So below I have laid out what I think is a sensible look at things how they stand, but as always there will be people that disagree, ad people that want to voice their own suggestions of how thing will pan out over this year. Sensible Debate is good. trolling and ridiculous comments are not.

Ok so AMD are showing us the first Benchmark from Polaris 10 at the Capsaicin Live Event tonight at around 11:00pm. Hopefully the numbers they show us, will give us something to work with to extrapolate the bigger parts performance.

For this post I will be using the charts from Techpowerup, as we know they give quite a broad range of benches with gives a reasonable indication of overall performance. Yes there are other sites and benches that can show differing results, but for this thread I will be using these.

What we know about the AMD side of things.

AMD flagship card as it stands today is the FuryX, performance wise it sits just behind the NVidia 980ti and moving to just in front of the 980ti as resolution increases.

Here is the latest TPU review showing the order of cards dated FEB 12th ( I'm not looking at the exact numbers here just an overall view of the order of cards.)

latest-TPU.jpg

Edit: TPU's test specs here

Now the part AMD are showing tonight, should show us that it is capable of being VR ready in the Steam VR benchmark test. (if it isn't then I really do not see the whole point of tonight's presentation) this should put it around the 290X/970 performance bracket. (seeing as they are the minimum specs for VR).

Now for Polaris 11, the bigger of the two Polaris chips that we know about, you would have to think that it will be AMD's fastest single ship card, as it just seems bonkers for AMD to launch a new card and it not be quicker than the FuryX. I mean if they can get 290X performance out of Polaris 10 then Polaris 11 being nearly four times bigger must be able to reach the FuryX. (Of course if the image isn't to scale than that blows that one out of the water, but if the image isn't to scale then what was the point of it in the first place.)

Polaris-excitment.jpg

Now we come to the crux of the matter on the AMD side, the FuryX2 or whatever they end up calling it.
We cannot get around the fact that it is late to the party, AMD showed us it back in June last year, saying it would be available in the fall, here we are in March and it still hasn't arrived.

But that lateness aside, what sort of performance are we going to be looking at. Well with reportedly with two Nano chips on board it will be giving performance something like Nano crossfire.

Nano-crossfire.jpg

Edit: TPU's test specs here.

Now is Polaris 11 going to make this card completely obsolete in a matter of a couple of months or is it not going to be as fast as this. Look at the 1080p graph there, that is a pretty small window to squeeze into (between FuryX and Nano crossfire). the gap is certainly bigger at 4k, but still not massive.

My self I am not sure how things will pan out on the AMD side of things.

Now on to the NVidia side of things

We know that NVidia will release new chips this year, there have been rumours, but very little solid info from the green camp.

The presentation slides that NVidia have shown show things like 10 times faster and high speed NVlink, but we know that those sort of performance claims are for very specific circumstances and for a very specific aspects of the chip, (Even Jen-Hsun called it CEO math and rough estimates) Bottom line we certainly do not expect a general 10 times performance increase from Pascal.
This is where we run into problems, because other than showing us a mock up drive px2 unit, not with actual Pascal chips on it and telling us some numbers that don't really equate to gaming at all, we haven't seen anything about NVidia's new Pascal chips at all.

Hopefully NVidia will give us medium Pascal first, unless they go for a smaller design first like they did with the Maxwell 750's, but the most recent Nvidia cards have been of the smallest Maxwell chips GM206, so replacing them so soon. It all gets into speculation and I wont go there.
Bottom line we know next to nothing about NVidia's Pascal chips, with any luck we should find out more at GTC.


It seems quite odd, that even though AMD have shown us more leaks and titbits than NVidia have rumour would have it that NVidia might announce first.

This entire post is just my thoughts and opinion's on the subject, things may be incorrect and mistakes happen but hopefully you will get the point.. I certainly look forward to reading other peoples thoughts on realistic performance expectations, because lets face it most of us would love the new cards to be 10 times more powerful and a quarter of the price, but we know that isn't going to happen. ( I know that is the first time I have mentioned price, this is deliberate, as the price of the new cards and indeed even the current cards is another discussion altogether.)
 
Last edited:
Which version of AMD Drivers are used on the charts you posted?
Because Cat 15.7.1 or 15.15beta (from July 2015) when the cards came out and most reviews were done, to Crimson 15.12 (let alone 16.3) the performance gains on all AMD cards is pretty huge (even old 290X ones). Let alone the Fury/FuryX series who got the biggest boosts.
 
Last edited:
Panos said:
Which version of AMD Drivers are used on the charts you posted?
Because Cat 15.7.1 or 15.15beta (from July 2015) when the cards came out and most reviews were done, to Crimson 15.12 (let alone 16.3) the performance gains on all AMD cards is pretty huge (even old 290X ones). Let alone the Fury/FuryX series who got the biggest boosts.
Click to expand...

Edited in the test specs that TPU published at the time of the reviews.
Of course these might not be with the very latest drivers but unless someone cares to find more up to date reviews with all the cards and games benched these will have to do. (I did also state "I'm not looking at the exact numbers here just an overall view of the order of cards")
 
Being realistic I expect as follows considering Polaris is rumoured to release first. The is based on the architectural improvements to GCN4 as well as hardware improvements coupled with the new node performance and power saving increases.

Top tier Polaris ~30% faster than Fury X/980Ti selling at ~£520-£550
2nd tier Polaris ~= to 10% faster than Fury X/980Ti selling at ~£420-£450
3rd tier Polaris ~ Fury pro performance selling for ~£330

My prices and performance speculations are within 5% best guesses considering how new higher performance GPUs released in the past. IMHO 980 was only marginally faster overall (~10%) than the existing high end so the performance tax premium could not be applied. The same applied to the GTX680 which because it had similar performance to 7970 was released to undercut the 7970. The 7970 and GTX780 initial prices when no competitive GPUs were available shows both AMD and Nvidia are happy to price at ~£550-£600 when they have the top tier performance market to themselves. In fact Nvidia are happy to price at £900+ if the option presents itself.

Only when Pascal is released to bring competition to the market will the prices stabilise at a lower level. Of course if Pascal releases at the same time as Polaris then a price war will benefit us all.
 
Last edited:
The Fury X has considerable more theoretical performance than the 980ti when both are at stock. And if the polaris 11 has similar numbers of shaders then it will have greater performance than the 980TI by a margin.

But the above assumes that the changes in GCN4 drastically improve GCN's ability to make use of its shaders. And considering the additions to GCN that they mention then this should be close. All before any clock improvements come into play from the node change.
 
Tpu slides.:D

My guess on tonights show, damp squib of a polaris 'preview' and more concentrated on the late late late show that is dual Fiji and AMD's committment to VR.
 
Mauller said:
The Fury X has considerable more theoretical performance than the 980ti when both are at stock. And if the polaris 11 has similar numbers of shaders then it will have greater performance than the 980TI by a margin.

But the above assumes that the changes in GCN4 drastically improve GCN's ability to make use of its shaders. And considering the additions to GCN that they mention then this should be close. All before any clock improvements come into play from the node change.
Click to expand...

Honestly we only need it to be on par with the GTX 980 ti (overclocked) in DX11 with AMD's ****** DX11 performance, and be much faster in DX12 (this will get people to upgrade and switch). Nvidia's DX12 performance right now is not what it should be.
 
I don't think its realistic at all to think that a top tier chip which basically contains twice of everything + a newer architecture + presumably better overclocker will only be faster by 30%.

I'd say it should be faster by minimum 50%.
 
I agree with Mtom. It's been so long since we've seen a significant die shrink that maybe we've forgotten what one looks like. I'd say at least 50% performance improvement for a chip of similar size. Minimum. Keep in mind that a move from 28nm -> 14nm node represents a potential 4x increase in transistor density for the same area.

Taking AMD as an example; Take the fastest 55nm single-chip card, the Radeon 4890 which chucked out 1360 GFLOPs - not bad! The die size was 282mm^2.

Now let's take the fastest 28nm single-chip which represents roughly a 4x doubling in transistor density; the R9 Fury X which chucks out 8601 GFLOPs - yeesh! The die size is more than double the 4890 with 596mm^2.

Now for the sake of argument let's chop the Fury X in half in terms of die-size and compare rough performance; 1360GFLOPs vs 4300. That's roughly 3x faster for the same die size.

We all know what kind of scaling issues Intel has had down at those teeny, tiny node sizes so let's try to keep apples to apples. A chip on the 14nm node with 4x transistor density and a similar ~290mm^2 die size, all else being equal, should represent a 2x - 3x performance improvement. In other words we'll see Fury X performance from mid-range cards. This is ignoring architectural changes and issues with power consumption. I think there is a potential to see some interesting bottlenecks appear within both AMD and Nvidia's architectures. At the same time performance is going to be a riot on both red and green teams.

This generation is going to be pretty awesome.
 
Last edited:
Worth noting the 14nmLPP from Samsung/Glofo and the 16nmFF+ from TSMC are actually 20nm with finfets.

The 14nmLPP process is ~15% denser than TSMC's though.
 
D.P. said:
The TMSC process apparently yields chips with better power characteristics.
Click to expand...

I could be wrong, but I believe that's when they were comparing it to 14nmLPE (referencing the comparisons of the iPhone having better battery life with 16nm than 14nm).

According to Samsung, 14nmLPP consumes 15% less power and produces 15% more performance. Making ~35% increase in performance per watt over LPE.
 
Next 5 months going to be interesting.
On the green side, Pascal GP104 which is just a die shrink of the Maxwell with FP16/32/64 added and 256bit bus GDDR5X

On the red side Polaris, is a brand new GCN core with HBM, HDR, DP 1.3, HDMI 2.0a, and superior version of delta compression and rasterization for complete DX 12.1 support. But the most important is compute pre-emption, which is usually seen in CPU's.
 
Last edited:
If their claims are anything like Intel's, then don't expect anything massive :P

Anyway, I feel this will be the year I switch back to red.
 
Panos said:
Next 5 months going to be interesting.
On the green side, Pascal GP104 which is just a die shrink of the Maxwell with FP16/32/64 added and 256bit bus GDDR5X

On the red side Polaris, is a brand new GCN core with HBM, HDR, DP 1.3, HDMI 2.0a, and superior version of delta compression and rasterization for complete DX 12.1 support. But the most important is compute pre-emption, which is usually seen in CPU's.
Click to expand...

So all the AMD features are new shiny and bright, where as NVidia will just sit on their haunches and not bring anything new to the table at all.
Well that is certainly a possibility but very unlikely.

I seriously doubt that any of the new cards coming middle of this year will be using GDDR5X.
 
Perhaps there won't be any ever. Anyone else think this is Micron trying to grub some quick cash and trying to fool AMD/NV with something that won't be ready, but they saw through it?
 
You must log in or register to reply here.
Back
Top Bottom