Predict: The Next Generation Console Tech

Status
Not open for further replies.
Are you guys talking about Single Precision FLOPS oder Double Precision FLOPS? The latter would be quite interesting in a HSA system that relies heavily on GPGPU algorithms, right?
 
Are you guys talking about Single Precision FLOPS oder Double Precision FLOPS? The latter would be quite interesting in a HSA system that relies heavily on GPGPU algorithms, right?

If Durango has double precision flops greater than 1.2 teraflops, it'll sport the most powerful single GPU yet, desktop or mobile. The 7970 is 950 gigaflops double precision.
 
But if both systems are measured in SP-FLOPS, how can Durango have "better" FLOPS than Orbis? Or what did you say exactly? "the best FLOPS you'll seen yet"

Sounds pretty contradictory, if you ask me...

There is no fundamental difference between the quality of a theoretical flop in ideal conditions.

However in a GPU on a real workload not all the flops are used all of the time, for many reasons. Ignoring for a minute things like rendering shadows, Z pre-passes or the geometry pass in a deferred renderer where for the most part the GPU's ALU's sit entirely idle. Even when a GPU is working on Shader heavy code ALU's are often under utilized, to get the massive peak computational performance they batch data into wavefronts often these wavefronts run with portions of the ALU's disabled, or because of register pressure not enough wavefronts can run to hide memory reads and the ALU's sit idle while the data is fetched.

It is quite possible for a chip with a lower number of total flops to have better performance that a machine with a higher number. It would require some probably significant architectural difference. For example a much larger register pool, narrower wavefronts, not batching pixels as quads, but all of those things would eat additional silicon, and potentially introduce other issues. So it's all a trade off.

Having said that I know nothing about the actual Durango GPU and, it's unlikely this level of detail will ever be made public by MS, so...
 
So you're talking about Paper-FLOPS on a VLIW5 chip like the Cypress XT (I think this is a good example) versus a Pitcairn-class GPU.

Fair enough, but I wouldn't call it "better" or "best" FLOPS, since architectual tweakings are pretty common for core-gaming consoles.
 
If Durango has double precision flops greater than 1.2 teraflops, it'll sport the most powerful single GPU yet, desktop or mobile. The 7970 is 950 gigaflops double precision.
Not necessarily. The GCN architecture can be configured with 1/2 rate double precision performance, as well as the 1/4 and 1/16 rates seen in current products.

If Microsoft were using a 1.2TF DP/2.4TF SP graphics part that would make the large APU and 384bit memory bus rumours more believable, and explain why there haven't been any leaks from developers about its disappointing graphical performance.

The 1.2TF rumours could be purposeful misdirection from Microsoft with an element of truth.

BKilian say FLOPS great for comparing protein folding performance. Crap for comparing Game Console performance.

It seems likely both next generation consoles are using architecturally similar GPUs from AMD, in which case FLOPS can be a good guideline.

GCN is as far as I know an architecture without any great flaws, so I have my doubts even access to a large SRAM cache can bridge the rumoured 50% FLOPS deficiency in Durango.
 
Last edited by a moderator:
The speculations in this thread imply that a 1.2 SP TFLOPS GPU can be a 1.8 SP TFLOPS GPU (or beyond) but that is simply wrong.

You can argue that a 2.7 SP TFLOPS GPU like the Cypress XT is not able to deliver this peak floating point performance all the time, of course, but a 1.2 SP TFLOPS GPU can not deliver more peak performance than this exactly 1.2 SP TFLOPS.

I just wanted to remind you that some readers might get a completely wrong impression of what Durango can and what not (*cough* NeoGAF *cough*). If Durango has a 1.2 SP TFLOP GPU it may can deliver this peak performance more steadily than a desktop GPU, but it will never deliver more than this 1.2 SP TFLOP.
 
So you're talking about Paper-FLOPS on a VLIW5 chip like the Cypress XT (I think this is a good example) versus a Pitcairn-class GPU.

Fair enough, but I wouldn't call it "better" or "best" FLOPS, since architectual tweakings are pretty common for core-gaming consoles.

I wouldn't call it a good or better FLOP either. In practice, it's up to the developers to optimize for the architecture. For a closed system, if there is more headroom for peak performance, it may be possible for creative and hardworking devs to hit greater height.

I suspect, like this gen, the developers' effort should count the most.
 
I guess theres some sort of suggestion from thuway on GAF that Orbis uses GCN and Durango GCN2? (the latter presumably better, but then some gaffers are saying theres little or no performance difference anyway?)

I have no clue about this though...

It wouldn't seem to make sense for Sony to not use AMD's latest tech at time of shipping, so I'm dubious.
 
There is no fundamental difference between the quality of a theoretical flop in ideal conditions.

However in a GPU on a real workload not all the flops are used all of the time, for many reasons. Ignoring for a minute things like rendering shadows, Z pre-passes or the geometry pass in a deferred renderer where for the most part the GPU's ALU's sit entirely idle. Even when a GPU is working on Shader heavy code ALU's are often under utilized, to get the massive peak computational performance they batch data into wavefronts often these wavefronts run with portions of the ALU's disabled, or because of register pressure not enough wavefronts can run to hide memory reads and the ALU's sit idle while the data is fetched.

It is quite possible for a chip with a lower number of total flops to have better performance that a machine with a higher number. It would require some probably significant architectural difference. For example a much larger register pool, narrower wavefronts, not batching pixels as quads, but all of those things would eat additional silicon, and potentially introduce other issues. So it's all a trade off.

Having said that I know nothing about the actual Durango GPU and, it's unlikely this level of detail will ever be made public by MS, so...

Is what happens with Nvidia flops vesus AMD flops. Nvidia always has less teorical flops but bigger dies due to bigger registers, more complex scheduler, more shader arrays...that become them 'super flops'.
 
I guess theres some sort of suggestion from thuway on GAF that Orbis uses GCN and Durango GCN2? (the latter presumably better, but then some gaffers are saying theres little or no performance difference anyway?)

I have no clue about this though...

It wouldn't seem to make sense for Sony to not use AMD's latest tech at time of shipping, so I'm dubious.

So, GCN2 would be like going Kepler architecture ( well, more like Fermi in which special sauces like hardware sheduler were greater and theorical flops less ) way with less flops but 'super flops' instead?. jejeje. I don´t think GCN2 will be such a departure from the already great GCN architecture.
 
The
The speculations in this thread imply that a 1.2 SP TFLOPS GPU can be a 1.8 SP TFLOPS GPU (or beyond) but that is simply wrong.

You can argue that a 2.7 SP TFLOPS GPU like the Cypress XT is not able to deliver this peak floating point performance all the time, of course, but a 1.2 SP TFLOPS GPU can not deliver more peak performance than this exactly 1.2 SP TFLOPS.

I just wanted to remind you that some readers might get a completely wrong impression of what Durango can and what not (*cough* NeoGAF *cough*). If Durango has a 1.2 SP TFLOP GPU it may can deliver this peak performance more steadily than a desktop GPU, but it will never deliver more than this 1.2 SP TFLOP.

The point is that which GPU achieves averagely the best performance.
And that's depends entirely on the architecture. The guy who got banned said also 1D SIMD. What does it mean? Aren't GNC SIMD 16-way? Of course it could not be true but just for the fun of discussion.

Although I remember bkllian stating that next generation MS system will do things none PC can do: I think we are still missing a big element about Durango.
 
I guess theres some sort of suggestion from thuway on GAF that Orbis uses GCN and Durango GCN2?

This one sounds pretty contradictory as well. Microsoft is rumoured to use a heavily cusomised GPU which implies that they had to spent a long time on R&D. In my eyes GCN2 makes more sense for a more "off-the-shelf" approach wich would point to Sony.
 
I think this quote from a interview to a Dragon Age 3 developer could hide a interesting clue:

"For the next generation there will be quite a big leap, but it won't be as obvious," Thompson revealed. "People will do things in a cleverer fashion - and I have to be careful here as there are non-disclosure agreements involved!

http://www.neogaf.com/forum/showthread.php?t=508577&highlight=

What could make the life of an artist easier due to cleverer ways of making things -remember, things under NDA-?. Real time lighting for example...
 
my news posted here about digitalfoundry having informations that xboxnext and ps4 are technically very similar, passed without attention here...I think people on this thread should seriously start questioning their BLIND beliefs that xboxnext has double the RAM amount than PS, with the latter having triple bandwidth....I wouldnt call that myself as technically very similar...

We could end up this gen with practically the same hardware for both sony and microsoft...not only the same CPU, same GPU but also same Memory architecture...
 
Status
Not open for further replies.
Back
Top