Xbox One (Durango) Technical hardware investigation

[french toast]

Love this...I cant keep up with you guys when you get down and dirty this technically..so im just gonna sit back and read

Edit I do understand what your on about...just cant add to it lol.

 

[Proelite]

So that's why Microsoft put an outdated GPU inside the 720 , because of the ESRAM ... maybe it will more than make up for any limitations ... that's how you get the"680gtx performance" .
That's amazing .

What makes you thinks it's outdated.

I remember reading an article that seemed to show the the Durango caches are larger per CU than GCN caches.

The GPU IS achitectured for 100% peak throughput on average.

 

[anexanhume]

What makes you thinks it's outdated.

I remember reading an article that seemed to show the the Durango caches are larger per CU than GCN caches.

The GPU IS achitectured for 100% peak throughput on average.

I think that was debunked here. Maybe an AIStrong post?

 

[kots]

What makes you thinks it's outdated.

Come on now , it's a 7770... already outdated , late 2013 it'll be a joke .

The GPU IS achitectured for 100% peak throughput on average.

Too bad the ceiling is very low ...
Anyway , seems that the buzzword lately is efficiency , i wonder why

 

[Proelite]

Come on now , it's a 7770... already outdated , late 2013 it'll be a joke .

It has 2 more CUs than 7770, esram + DMEs, and, possibly, different caches. I don't know what the final performance envelope will be, but I am willing to bet that it _curbstomps_ the 7770 GHZ edition.

 

[anexanhume]

Come on now , it's a 7770... already outdated , late 2013 it'll be a joke .

Yeah, with all the new GPUs coming out this year...

 

[liolio]

Come on now , it's a 7770... already outdated , late 2013 it'll be a joke .

Luckily AMD, it seems, is not to release any new GPU throughout 2013, :undefined smiley blending all sort of mixed feeling on the matter:

Too bad the ceiling is very low ...
Anyway , seems that the buzzword lately is efficiency , i wonder why

Well it is indeed not a beast but a few people were expecting that much (though it is on the low side of of my reasonable guesstimates based on power consumption foremost).

I don't not believe in that 100% efficiency anyway, at some point while rendering as sebbbi explained in another thread, most of the GPU (ALUs) are idle with the "fixed" function part of the chips doing their thing. I don't see how that can be change without being able to do what the "big kepler" does ie support multiple command inputs/lists (/ insert proper vocable...) (if I understand the issue properly).

 

[Ninjaprime]

What are the average latency differences between:
-6T - SRAM
-1T - SRAM
-DRAM

6T SRAM, aka real SRAM, sub nanosecond. Seems like someone had some 500ps(half nanosecond) cells they were bragging up not too long ago.

1T SRAM, aka eDRAM with buffers, 2-3 nanosecond was the fastest I've heard them brag up.

eDRAM, depends, IBM is supposedly at only 1.7 nanoseconds with their caches, but that is not typical, and its expensive high performance stuff. More typical would be in the 4-6 nanosecond range.

Normal DRAM varies, but is typically 30-60 nanoseconds.

 

[Kaotik]

Since you obviously missed it, after the roadmap leak thingy saying "HD7 stable through 2013" AMD has already stated that they are releasing new GPU's this year and will clear up the situation this week (or was it next)

 

[ERP]

The low latency won't help rendering much, but it might very well boost GP compute.

FWIW I'm not sure this is true, the effect is certainly less dramatic in rendering than in GP compute where IME most none trivial compute shaders are memory bound.
But I think it will have an impact on rendering IF you can source data from the low latency RAM, and IF something else isn't the bottleneck. Exactly how much I'm not sure.

 

[Proelite]

FWIW I'm not sure this is true, the effect is certainly less dramatic in rendering than in GP compute where IME most none trivial compute shaders are memory bound.
But I think it will have an impact on rendering IF you can source data from the low latency RAM, and IF something else isn't the bottleneck. Exactly how much I'm not sure.

What kind of impact are we talking about here on a "good" not best case scenario? 5-10% increase?

 

[pjbliverpool]

In that case the example of 3x800 mhz would be only for GPGPU ops, and in this way we already would have matched the 7970 performance for GPGPU ops.

This is a ridiculous assertion. Do you not think that if it were as straight forward as providing a small pool of low latency esram on die to allow 3x greaterl GPGPU performance across the board they would not have done so already in the PC parts which are by contrast virtually unlimited in die size, cost or power draw?

I really wish people would stop grasping at the tiniest straws to justify their hopes that Durango will outperform Orbis or even more ridiculous compete with high end PC's. Can't we just accept it for what it seems to be? i.e a slick family gaming/media device that is also "competitive" when it comes to core gaming and is still nice and cheap.

 

[ERP]

What kind of impact are we talking about here on a "good" not best case scenario? 5-10% increase?

If it were easy to quantify I would have quantified it.
On compute it's really dramatic.
On Graphics it's much easier to hide the latency, and that's what GPU's are designed to do, but anytime memory accesses don't cache well ALU's will likely sit and twiddle their fingers.

I actually think on modern GPU's ALU's sit idle a lot, partly because they are underutilized for a lot of the frame, and partly because the complex shaders can be memory bound. It's one of the reasons a well set up Compute model is interesting, because it may not have a dramatic effect on graphics at all even if it's sharing resources. But this is speculation, I've never sat down and measured it.

 

[Shifty Geezer]

This is a ridiculous assertion. Do you not think that if it were as straight forward as providing a small pool of low latency esram on die to allow 3x greaterl GPGPU performance across the board they would not have done so already in the PC parts which are by contrast virtually unlimited in die size, cost or power draw?

Yeah. Any uber solution would have been done before, unless it's something completely new. If caches could get that much more performance, nVidia's highly lucrative GPGPU range would be going that route and selling gangbusters to supercomputing cluster builders. Technology doesn't exist in a vaccuum and must be considered in relation to other developments. It's extremely rare to get something completely new that pushes performance beyond whatever anyone else can do. Any gains are invariably accompanied by equivalent drawbacks, such as complexity in use or fringe cases where the performance advantage can be leveraged.

 

[Proelite]

Yeah. Any uber solution would have been done before, unless it's something completely new. If caches could get that much more performance, nVidia's highly lucrative GPGPU range would be going that route and selling gangbusters to supercomputing cluster builders. Technology doesn't exist in a vaccuum and must be considered in relation to other developments. It's extremely rare to get something completely new that pushes performance beyond whatever anyone else can do. Any gains are invariably accompanied by equivalent drawbacks, such as complexity in use or fringe cases where the performance advantage can be leveraged.

On PC you can have the case where über solution causes decreased performance in older titles. You can't have that.

On consoles this is not the case.

 

[Gubbi]

I actually think on modern GPU's ALU's sit idle a lot, partly because they are underutilized for a lot of the frame, and partly because the complex shaders can be memory bound. .

GCN @800MHZ can tolerate 50 ns memory latency ( 10 wavefronts* 4cycles/wavefront * 1.25 ns/cycle). If you have lots of texture/data fetches missing caches you'll stall your ALUs.

You'll also waste ALUs if you can't fill your wavefronts (small tris), here lower latency wont make a difference.

Cheers

 

[Love_In_Rio]

This is a ridiculous assertion. Do you not think that if it were as straight forward as providing a small pool of low latency esram on die to allow 3x greaterl GPGPU performance across the board they would not have done so already in the PC parts which are by contrast virtually unlimited in die size, cost or power draw?

I really wish people would stop grasping at the tiniest straws to justify their hopes that Durango will outperform Orbis or even more ridiculous compete with high end PC's. Can't we just accept it for what it seems to be? i.e a slick family gaming/media device that is also "competitive" when it comes to core gaming and is still nice and cheap.

Nothing ridiculous.Desktop products have to deal with ancient engines not latency sensitives and beat the competitor in benchmarks,so better more alus

 

[anexanhume]

On PC you can have the case where über solution causes decreased performance in older titles. You can't have that.

On consoles this is not the case.

Examples?

 

[Love_In_Rio]

Examples?

Any directx 9 game with light shaders will not suffer from cache misses, for example COD.

 

[Shifty Geezer]

On PC you can have the case where über solution causes decreased performance in older titles. You can't have that.

What about supercomputers? If SRAM can make a 7770 run like a 680gtx (the original point Love_In_Rio is defending), why isn't nVidia adding it to its Tesla range? Or are only MS able to see the benefits and nVidia's going to be kicking itself when they see the incredible performance Durango gets with such cheap silicon?