280x vs Fury X scaling-- double shaders, double performance?

[soccerballtux]

And for that matter, what about the 7850 (half the shaders of the 280x?


I'm most interested in 280x for upgrade, but in general the question interests me.

 

[Enigmoid]

And for that matter, what about the 7850 (half the shaders of the 280x?

I'm most interested in 280x for upgrade, but in general the question interests me.

Fury most definitely has bad scaling problems

179% scaling over the 280X, which scaled poorly over the 270X (Pitcarin).

In fact, looking at the 270X, Fury X is 238% faster despite having 3.2x the shaders operating 5% faster or 3.36x more processing power.

4K is of course more favourable, but a niche market.

 

[monstercameron]

Fury most definitely has bad scaling problems

179% scaling over the 280X, which scaled poorly over the 270X (Pitcarin).

In fact, looking at the 270X, Fury X is 238% faster despite having 3.2x the shaders operating 5% faster or 3.36x more processing power.

4K is of course more favourable, but a niche market.

Bad scaling? Interesting choice of words. How do you know how well certain units scale?

 

[flopper]

And for that matter, what about the 7850 (half the shaders of the 280x?


I'm most interested in 280x for upgrade, but in general the question interests me.

Looking at the 380 2gb at 180 euro here you need a fury x to double the fps at 600euro.
that at 2560x1440. Its what I checked for my sis kid altough he plays 1920p so the card plays well for its price.
die shrink is needed to bring more value to midrange

 

[Enigmoid]

Bad scaling? Interesting choice of words. How do you know how well certain units scale?

The scaling is poor. Scaling for GPU's can easily approach the ideal 100% in many cases.

 

[monstercameron]

The scaling is poor. Scaling for GPU's can easily approach the ideal 100% in many cases.

Maybe, I don't know but I am interested in your sources or an explanation.

 

[bystander36]

I thought that shader/cuda core counts did not scale 1:1. Everytime you get a cut down version of the same card, the percentage of units cut is almost never lost in performance. The Titan X and 980ti are good examples, so are the 7970 and 7950.

 

[Red Hawk]

Fury most definitely has bad scaling problems

179% scaling over the 280X, which scaled poorly over the 270X (Pitcarin).

In fact, looking at the 270X, Fury X is 238% faster despite having 3.2x the shaders operating 5% faster or 3.36x more processing power.

4K is of course more favourable, but a niche market.

Well Tahiti didn't scale so well over Pitcairn because while the former has 60% more shaders (2048 vs 1280), it actually has the same amount of geometry shaders (2) and the same amount of ROPs (32). So when a game has geometry or raster related bottlenecks, Tahiti's advantage over Pitcairn deteriorates. In synthetic tessellation benchmarks, thanks to its higher clock speed a 270X will actually beat a stock 280, since theoretical geometry performance comes down to the amount of geometry engines and the clock speed.

Fiji should be a different story. Not only does it have twice the ROPs and geometry engines as Tahiti, but its ROPs are improved with delta color compression and its geometry engines have improved geometry handling. So they should scale very well with Tahiti. Who knows where the real bottleneck that's holding Fiji back lies.

 

[Enigmoid]

Maybe, I don't know but I am interested in your sources or an explanation.

Just look in many benchmarks and you will see the general trend.

I thought that shader/cuda core counts did not scale 1:1. Everytime you get a cut down version of the same card, the percentage of units cut is almost never lost in performance. The Titan X and 980ti are good examples, so are the 7970 and 7950.

This is seen a lot. However remember that the cutdown is never systematically constant. Lots of time shaders are cut but geometry, bandwidth, and the front end remains the same.

Well Tahiti didn't scale so well over Pitcairn because while the former has 60% more shaders (2048 vs 1280), it actually has the same amount of geometry shaders (2) and the same amount of ROPs (32). So when a game has geometry or raster related bottlenecks, Tahiti's advantage over Pitcairn deteriorates. In synthetic tessellation benchmarks, thanks to its higher clock speed a 270X will actually beat a stock 280, since theoretical geometry performance comes down to the amount of geometry engines and the clock speed.

Fiji should be a different story. Not only does it have twice the ROPs and geometry engines as Tahiti, but its ROPs are improved with delta color compression and its geometry engines have improved geometry handling. So they should scale very well with Tahiti. Who knows where the real bottleneck that's holding Fiji back lies.

Yep, and we seem to be seeing the same general trends with Fury X. The front end simply isn't big enough to feed all those shaders (pitcarin also defeated the 7950 in certain games - not just synthetic benchmarks).

 

[soccerballtux]

regarding scaling, architecturally, should they focus on wider, or clock speed?

 

[crisium]

960 is exactly half of a 980 and has the same scaling.

I guess Maxwell has poor scaling too.

Or actually this is typical scaling.

 

[Enigmoid]

960 is exactly half of a 980 and has the same scaling.

I guess Maxwell has poor scaling too.

Or actually this is typical scaling.

Yep, though a 960 has a typical TDP of 120W while a 980 has a TDP of 165W. The 960 does better comparatively because it is allowed more power (also why it is less efficient).

Set the 960 and 980 so that the 980 is using twice the power of the 960 and see what happens. 960 also seems to run at higher clocks than the 980 as well.

(Fury is more difficult to compare because it uses HBM).

 

[soccerballtux]

960 is exactly half of a 980 and has the same scaling.

I guess Maxwell has poor scaling too.

Or actually this is typical scaling.

it's probably typical scaling. You can't have perfect scaling unless your GPU is never waiting on you CPU-- which isn't really possible

I think if anything Fiji might direct AMD's attention to shoring up CPU scaling, if DX12 doesn't resolve that.

 

[raghu78]

Fiji GPU and the Fury X points to the limits of the GCN architecture. Nvidia's Maxwell architecture is an excellent example of a scalable architecture. With lesser transistors Nvidia can deliver more performance because the underlying architecture is very efficient, well designed and scalable. The main takeaway is AMD now has a huge architectural gap to overcome when the next gen GCN faces off against Pascal. Nvidia are going to be building off and improving an extremely efficient architecture while AMD's next gen will have to compete against not Maxwell but Pascal. Now thats like trying to close multiple GPU generations architectural and efficiency gap. AMD needs a radical redesign of GCN if they want to even be anywhere close to competitive with Nvidia Pascal in terms of perf/watt, perf/transistor and perf/sq mm when the next gen GPUs launch.

 

[3DVagabond]

Fiji GPU and the Fury X points to the limits of the GCN architecture. Nvidia's Maxwell architecture is an excellent example of a scalable architecture. With lesser transistors Nvidia can deliver more performance because the underlying architecture is very efficient, well designed and scalable. The main takeaway is AMD now has a huge architectural gap to overcome when the next gen GCN faces off against Pascal. Nvidia are going to be building off and improving an extremely efficient architecture while AMD's next gen will have to compete against not Maxwell but Pascal. Now thats like trying to close multiple GPU generations architectural and efficiency gap. AMD needs a radical redesign of GCN if they want to even be anywhere close to competitive with Nvidia Pascal in terms of perf/watt, perf/transistor and perf/sq mm when the next gen GPUs launch.

You know, I remember the same things being said vs. Maxwell. It turned out not to be true. We'll see.

 

[bystander36]

Of course there is a big elephant in the room; DX12. Maxwell is doing great with DX11, but DX12 may move the goal posts, and it could potentially prefer AMD architecture. We'll have to wait and see. A lot of this stuff is guess work, because they don't know what the dev's will do with their engines before they are created.

Look at Kepler, it started off strong, but did not keep up well with the game engines that later followed. Maxwell was a redesign that currently looks strong, but DX12 could change all that.

 

[soccerballtux]

AMD's engineering team aren't complete idiots. If GCN weren't a good design they wouldn't have kept it.

I'm surprised they haven't tweaked it a bit more (e.g., to clock past 1200mhz)

 

[omek]

The synthetics aren't showing a lack of scaling, and I know, it's a synthetic. The point of it however is to relieve any CPU limit or rid-of anomalies which could be caused by the CPU to produce a clean GPU benchmark - what seems to be happening is that in normal gaming the GPU is 'out-powering' the CPU and obviously causing a bottleneck skewing the results slightly.
The scaling doesn't look bad at all.

Almost 2x, 7300 to 14400

Pretty much 2x, 3330 to 6600

BFHL at about 190%-195% (without doing the actual math)
http://www.guru3d.com/index.php?ct=...dmin=0a8fcaad6b03da6a6895d1ada2e171002a287bc1
Not perfect but far better than the proposed 1.75x at the start of the thread.

Bioshock 48fps vs 107fps (280x/Fury X), far over 2x
http://www.guru3d.com/index.php?ct=...dmin=0a8fcaad6b03da6a6895d1ada2e171002a287bc1

 

[Mondozei]

There was an interesting discussion on this topic on a recent Tech Report podcast. It was devoted mainly to architecture discussions.
It's here for anyone who is interested,

The basic takeaway, if I remember correctly, was that GCN is in some cases ahead of its time. It has more focus on shading performance in general, often advanced types of shading techniques which will become more important in the future. So the "next" in GCN isn't for nothing.

Another key point is that GCN is a more general performance-oriented arch. Maxwell was designed from ground up to be much more focused on gaming and similar tasks since NV knew that was where the bulk of sales were going. Pascal does seem to be going in a different direction, especially with NVlink and similar technologies, which are important for researchers and those doing machine learning/convoluted neural networks etc.

The folks at Tech Report said that Maxwell is a better arch for the here and now, while GCN will age better. This isn't surprising, this was what happened in part between the Kepler arch and GCN 1.1. Yes, AMD has done better driver support, as well as the advantage of a unifying arch, but there is more to it than that.

Maxwell is better optimised to just be pushing more triangles and for the most part, that is still dominant in most games over shading. I'd recommend listening to the podcast. It's great.

 

[Mondozei]

179% scaling over the 280X, which scaled poorly over the 270X (Pitcarin).

In fact, looking at the 270X, Fury X is 238% faster despite having 3.2x the shaders operating 5% faster or 3.36x more processing power.

4K is of course more favourable, but a niche market.

Scaling has to do with a lot more than shaders.

Also, 100% scaling isn't the norm at all. And you still haven't provided a source for that. And no, most benchmarks don't support that view.