Predict: The Next Generation Console Tech

[Carl B]

Assuming the end goals are the same, the FLops (in my mind) could only be substantially differentiated from the alternative architecture were they a) directed towards different use-cases, or b) more consistently able to achieve peak performance. The whole thing might play more to available bandwidth than anything else, though I am interested to see the ways in which the GPUs materially deviate from the world of desktop offerings.

I am very much wait-and-see when it comes to those specs, but this most recent set of anticipatory leaks has me getting excited for that pre-launch timeframe.

 

[Ruskie]

Only 32MB? Son, I am disappoint.

But, I guess its enough...

This is whats wrong with B3D these days...Next gen migration of alot of posters from other forums.

Why do you assume the poster with 1 post and easy cryptic word is someone legit? Why do you people believe everything that gets posted?

Seriously, today I learned that PS4 is 4GB of GDDR5 with 50% more theoretical performances on GPU in comparison to Durango. I also learned Durango has blitter and 12 CUs even though nobody really said anything about it, people just picked it up and spread it like its a fact. We came from "1TF+" estimate from BGs sources source, to 1-1.5TF, to 1.2TF and in the end we finally "settled" for 1.23TF. I mean, can we at least limit that fud in B3D? This seems to be a emotional roller-coaster ride for some members of this board.

 

[DuckThor Evil]

A dude calling himself Jehova throwing riddles at us... I can't wait for all this BS to end.

 

[Carl B]

Seriously, today I learned that PS4 is 4GB of GDDR5 with 50% more theoretical performances on GPU in comparison to Durango. I also learned Durango has blitter and 12 CUs even though nobody really said anything about it, people just picked it up and spread it like its a fact. Can we at least limit that fud in B3D? This seems to be a emotional roller-coaster ride for some members of this board.

That's always the case though at this stage in console development - look back pre-360 and PS3 announcements and the rumors were flying off the walls. Add to that the fact that this is a thread dedicated to "predictions," and really... it's honestly better to have tons of off-the-wall posts in one thread than tons of off-the-wall threads.

 

[3dilettante]

I like how the blitter thing was probably started in B3D in idle discussion.
I wish I could start an internet rumor (Orbis FPGA-driven programmable I/O bus).

 

[Bagel seed]

Fabric controller? Data center bridging?

This thing will be cloud enabled day 1?

 

[bbot]

What's this talk about an "assist"? Is it going to to be like the PhysX co-processor?

 

[Ruskie]

I really think people should move away from flops as a measure of performance, especially for the Durango. I think everyone is privy to the fact that not all flops are equal, and Durango flops are the "best" flops you'll seen yet. If my understanding of some technical details are correct, they're substantially better.

Bkilian? What say you?

Coming from a prominent "FLOP war" poster this post surely sounds interesting.

That's always the case though at this stage in console development - look back pre-360 and PS3 announcements and the rumors were flying off the walls. Add to that the fact that this is a thread dedicated to "predictions," and really... it's honestly better to have tons of off-the-wall posts in one thread than tons of off-the-wall threads.

I understand that and I agree, its just that I have never really seen so many people taking every post like a gospel and changing their tune every few hours, depending on their mood.

 

[Bagel seed]

Seriously, today I learned that PS4 is 4GB of GDDR5 with 50% more theoretical performances on GPU in comparison to Durango. I also learned Durango has blitter and 12 CUs even though nobody really said anything about it, people just picked it up and spread it like its a fact. We came from "1TF+" estimate from BGs sources source, to 1-1.5TF, to 1.2TF and in the end we finally "settled" for 1.23TF. I mean, can we at least limit that fud in B3D? This seems to be a emotional roller-coaster ride for some members of this board.

I think the trouble is people forget to use the words "I think" in their proclamations

 

[patsu]

That's always the case though at this stage in console development - look back pre-360 and PS3 announcements and the rumors were flying off the walls. Add to that the fact that this is a thread dedicated to "predictions," and really... it's honestly better to have tons of off-the-wall posts in one thread than tons of off-the-wall threads.

Are you going to flush us down the toilet once all of us are in this thread ?

Assuming the end goals are the same, the FLops (in my mind) could only be substantially differentiated from the alternative architecture were they a) directed towards different use-cases, or b) more consistently able to achieve peak performance. The whole thing might play more to available bandwidth than anything else, though I am interested to see the ways in which the GPUs materially deviate from the world of desktop offerings.

I am very much wait-and-see when it comes to those specs, but this most recent set of anticipatory leaks has me getting excited for that pre-launch timeframe.

I suspect he meant to say the fixed function calls are more high level, but traditionally we only count programmable FLOPs. And the fast memory helps improve the efficiency.

 

[Hornet]

I really think people should move away from flops as a measure of performance, especially for the Durango. I think everyone is privy to the fact that not all flops are equal, and Durango flops are the "best" flops you'll seen yet. If my understanding of some technical details are correct, they're substantially better.

Bkilian? What say you?

I guess a sufficiently large, high-bandwidth and low-latency pool of 1T-SRAM will certainly make many algorithm much more efficient on Durango than they are on current PC GPUs*. Are you referring to this or are you hinting something more than the memory layout?

* For instance, a Radeon HD 7970 is a 3.8 TFLOPS GPU with 264 GB/s of bandwidth and 768 KB of L2 cache with 355 GB/s of bandwidth (not sure about the latency). If Durango has a 1.2 TFLOPS GPU with a relatively low main memory bandwidth but something like 64 MB of 1T-SRAM at 1 TB/s with much lower latency than the main memory, I could see it outperforming a Radeon HD 7970 in certain workloads.

 

[MrFox]

This is whats wrong with B3D these days...Next gen migration of alot of posters from other forums.

Why do you assume the poster with 1 post and easy cryptic word is someone legit? Why do you people believe everything that gets posted?

Seriously, today I learned that PS4 is 4GB of GDDR5 with 50% more theoretical performances on GPU in comparison to Durango. I also learned Durango has blitter and 12 CUs even though nobody really said anything about it, people just picked it up and spread it like its a fact. We came from "1TF+" estimate from BGs sources source, to 1-1.5TF, to 1.2TF and in the end we finally "settled" for 1.23TF. I mean, can we at least limit that fud in B3D? This seems to be a emotional roller-coaster ride for some members of this board.

When has the blitter idea been anything more than Amiga nostalgia, and random speculation about new stuff? The other thread was started by listing 1985 techs (including the blitter) then I and others picked up on it for fun. Nobody claimed the blitter was real, or did I miss it?

(btw, many of my recent posts have been removed, if I pissed off the mods I'm really sorry)

 

[Ruskie]

When has the blitter idea been anything more than Amiga nostalgia, and random speculation about new stuff? The other thread was started by listing 1985 techs (including the blitter) then I and others picked up on it for fun. Nobody claimed the blitter was real, or did I miss it?

Well its considered a fact on some other forums since it "originated" from B3D. You shouldn't doubt human stupidity to take something they feel good with and run around saying its a fact, they will even start to believe it themselves.

 

[-Sweeper_]

While I dont expect any miracles from GNC 2.0 it will probably bring 10-15% more performance at same TDP (similar to HD5k to HD6k transition) + more efficient use of die space, on a more mature 28nm process.

A custom GCN2.0 GPU with 16 CUs GPU would probably take less than 150mm² while oferring a very respectable >1.6Tflops at moderate clocks (~800MHz). This would easily fit in a 00mm² APU with a bunch small Jaguar cores. Couple this with ESRAM and the right amount of RAM and M$ would have a very capable yet efficient next gen system.

I dont think 12 CUs would bring the kind of next gen bump that most people expect from a system launching 8 years after the 360, even though its Microsoft and they could still surprise us (maybe current rumours dont match the final specs... not to mention the special sauce talk). Both NV2A and Xenos were very capable GPUs at their times, and dont think it will be radically different this round.

 

[Shifty Geezer]

38 deleted posts and two bans later...

 

[DoctorFouad]

some new info of the ps4/xbox next coming soon from digitalfoundry :
http://www.eurogamer.net/articles/digitalfoundry-vs-4k-gaming

"Our information from developer sources suggests that even Sony - with much to gain in promoting 4K gameplay bearing in mind its upcoming range of screens - is making no attempt to evangelise the new display format to third-party developers, with 1080p the target resolution for Orbis titles. As both the next-gen PlayStation and Microsoft's Durango have a hell of a lot in common from a technological standpoint (more on that soon), we'd also venture to suggest that 1080p is the target for this year's new Xbox too."

having a lot of common from a technological standpoint ? to have a lot of common, even the amount of RAM and their respective bandwidth should be quite similar...I expect 4 Gb of RAM for both consoles, we could end up this gen with almost the same hardware but different brands with only software and marketing a differentiating point...the end of harwdare console competition atnd the days of "same hardware licensing, different vendors" are maybe closer than ever, only this time its AMD the licensing company...very strange and unexpected turn of events if true.

 

[Carl B]

Are you going to flush us down the toilet once all of us are in this thread ?

It looks like Shifty already had his hand on the lever.

I suspect he meant to say the fixed function calls are more high level, but traditionally we only count programmable FLOPs. And the fast memory helps improve the efficiency.

Yup, that could certainly be it (Proelite?), though in that case given that we are expecting "something" on the Sony side as well, it would be a matter of transistor count devoted to each architectural enhancement, respectively, and what their net yields are from a performance increase perspective relative to one another. Insofar as the "quality" of the Flops might be premised on some fixed function assistance, it seems we should be left to expect something above and beyond what the mainline PC architectures would imply in both cases.

 

[liquidboy]

Yep, quite a compelling vision. It does look like a son of Cell; an application level Cell so to speak. ^_^

If they do pursue that route, it may take years to realize. It looks hot and expensive too. Perhaps we will see a simplified first cut first ?

It all started in 2005/6 , they've been working on this already for many many years! Only in the last year have they needed to go into mass production..

Traditionally a SoC has been made up of several costly components, CpU/Gpu/Accelerators/encoders/decoders/dedicated single purpose components etc. etc. ...

In this new world most of those will be replaced with a specialized BTE. So instead of having to worry about testing several components integrated with each other ... they only need to test a few, as the rest are just specialized versions of a BTE.

Also because there are huge fat pipes (fast bus) between these BTE based processors and because they are laid out so close to each other in design, I wouldn't be surprised if the work needed to render a frame would be much smaller than what traditional SoC designs require.

Remember an entire scene, with all its objects, can be rendered on the HW thru these BTE's.

Completely new way of thinking about the structure of your game engine, when most of the software parts of your traditional game engine can now exist and executed ALL in hw ...

And im not even getting into the cool bits of the Raytracing advantages...

 

[pjbliverpool]

* For instance, a Radeon HD 7970 is a 3.8 TFLOPS GPU with 264 GB/s of bandwidth and 768 KB of L2 cache with 355 GB/s of bandwidth (not sure about the latency). If Durango has a 1.2 TFLOPS GPU with a relatively low main memory bandwidth but something like 64 MB of 1T-SRAM at 1 TB/s with much lower latency than the main memory, I could see it outperforming a Radeon HD 7970 in certain workloads.

If such a setup were to produce a noticeable performance increase over the wide memory bus / lots of CU's model of the 7970 then why wouldn't ATI already be following it in the PC market? i.e. much larger / faster cache at the expense of CU's and memory bus width?

Are these performance increases dependant on using this SRAM for the framebuffer and thus it has to be sized to a known maximum framebuffer size (1080p)? But then it's usage would obviously be completely different to the usage of the L3 in the 7970.

I must say all this talk of "super flops" does sound a bit dubious. I totally understand the futility of using flops as a single measure of power. Clearly there are plenty of other factors to consider (fill rate, texturing performance, geometry/tessellation performance, memory setup etc..) but all this talk of a 1.2 TFLOP GPU performing like a 3-4 TFLOP PC GPU because of some special sauce begs the question; how could AMD (and NV) be so incompetent to build these massive expensive GPU's with all this shader/texture/ROP/geometry power connected to big wide expensive buses and the fastest GDDR memory available when it's possible to build a GPU with 1/3rd the theoretical power which is cheaper and draws far less energy but performs the same.

I just can't see AMD/NV designing something so horribly unbalanced that results in a 4TFLOP GPU performing like a 1.2 TFLOP GPU using a different and cheaper memory configuration. From a common sense point of view it doesn't seem to add up.

 

[Hornet]

If such a setup were to produce a noticeable performance increase over the wide memory bus / lots of CU's model of the 7970 then why wouldn't ATI already be following it in the PC market? i.e. much larger / faster cache at the expense of CU's and memory bus width?

Are these performance increases dependant on using this SRAM for the framebuffer and thus it has to be sized to a known maximum framebuffer size (1080p)? But then it's usage would obviously be completely different to the usage of the L3 in the 7970.

I must say all this talk of "super flops" does sound a bit dubious. I totally understand the futility of using flops as a single measure of power. Clearly there are plenty of other factors to consider (fill rate, texturing performance, geometry/tessellation performance, memory setup etc..) but all this talk of a 1.2 TFLOP GPU performing like a 3-4 TFLOP PC GPU because of some special sauce begs the question; how could AMD (and NV) be so incompetent to build these massive expensive GPU's with all this shader/texture/ROP/geometry power connected to big wide expensive buses and the fastest GDDR memory available when it's possible to build a GPU with 1/3rd the theoretical power which is cheaper and draws far less energy but performs the same.

I just can't see AMD/NV designing something so horribly unbalanced that results in a 4TFLOP GPU performing like a 1.2 TFLOP GPU using a different and cheaper memory configuration. From a common sense point of view it doesn't seem to add up.

I am not claiming that such a setup would win in most situations, but that - despite being cheaper - might win in several workloads that Microsoft might have considered relevant for next generation engines (e.g., sparse voxel octree traversing). So the talk of "super flops" might only make sense for some specific workloads, which are not necessarily the same as the workloads for which the current high-end PC GPUs are designed.