Predict: The Next Generation Console Tech

[Rangers]

Bitcoin mining is a nice "comparison" between the flops of Nvidia and AMD... at least it was, half a year ago, as I am not sure of todays numbers. Back then, even the highest end NVidia GPUs couldn't hold a candle against my meager 5650 in my laptop. Though, the calculations needed for Bitcoin mining probably do favor AMD to a large margin, hence they are much faster.

Does this translate to better gaming performance? Hell no, but neither is a motorcycle a good choice for a Stock Car Race...

For that matter something like Furmark (which, is a big furry rotating donut aka real world shading in action aka somewhat game applicable) is also dominated by AMD.

There are definitely examples in the real world that AMD cards just had (have?) a lot more brute math force.

 

[TheWretched]

Was that due to speed or performance/watt/$?

Performance... my 6870 in my desktop can reach about 300Mhash/s, whereas a 580 barely reaches half that. And given that a 580 also sucks a LOT more power, it's also less profitable^^ (not that I am doing it... I was just playing around with a bit of OpenCL some months ago and since Bitcoin mining is mostly open source, it was an interesting case study).

 

[MrFox]

Performance... my 6870 in my desktop can reach about 300Mhash/s, whereas a 580 barely reaches half that. And given that a 580 also sucks a LOT more power, it's also less profitable^^ (not that I am doing it... I was just playing around with a bit of OpenCL some months ago and since Bitcoin mining is mostly open source, it was an interesting case study).

Is a hash computation even relevant to floating point performance?
Folding at home is very heavy in actual floating point ops.
http://folding.stanford.edu/English/FAQ-flops

 

[TheD]

Workload conditions (compute versus rendering) have a lot to say about this. The needs for many compute problems are quite different from a raster problem. A VLIW4 design at say 28nm 200mm^2 versus a GCN/Fermi on the same process size would see quite different performance (the former probably being better for most games up to this point as it has a high overlap for utilization, the later better for future work where compute becomes an important aspect and where previous GPU designs were not robust enough to maintain performance in these specialized programs).

What I am saying is that one can not just claim that because of console optimizations an AMD GPU would be able to claim a performance advantage over an Nvidia GPU in rendering workloads due to somehow unlocking its full FP performance.

AMD having better performance in furmark just proves my point, it is a PC program and thus proves that AMD cards can use the FP advantage they have and are not being unduly held back by PCs.

 

[ultragpu]

Charlie tweeted yesterday in regards to his recent ps4 and xbox720 rumor articles.
About PS4 using x86
https://twitter.com/#!/CDemerjian/st...37612361875459

Just got two independent confirmations about my PS4 x86 article.

About the xbox720 use of PowerPC:
https://twitter.com/#!/CDemerjian/st...50408038346752

99+% sure it is, but not 100%

 

[Acert93]

How ironic! For those of us rightfully/wrongfully hoping for relatively advanced consoles both would be encouraging rumors.

 

[Crossbar]

Could it just be graphics companies are reserved to the fate of stacked memory and interposers? Seeing as AMD had some sort of sample they showed SA in late 2011 indicates they are seriously looking at addressing memory issues outside of the traditional format.

Yeah, this is probably the case. Improving standard memory units beyond GDDR5 likely involves differential signalling. But since Rambus seems to hold some strong patents for that, the DRAM giants may just bypass that evolutionary step and go straight to stacked memories. The profit margins are low and competition is tough in the DRAM business, Elpida just went belly up.

The benefits of stacked memories using TSVs are twofold as I see it, first you can widen the bus without being pad limited and second you can lower the signal voltage and thereby lower the TPD significantly. I remember reading that it is common that about 20% of the TDP of performance ICs comes from the IO circuitry drving signals via the PCB.

GPUs seems like a perfect fit for silicon interposers and stacked memorys using TSVs. The hi performance version of the Wide IO memory standard may be what they are waiting for. If we can get the first samples of such memory during this year, it could possibly make its way into some console.

 

[V3]

So which x86 is a realistic candidate for PS4 ? Time frame, power, size, cost and all ? Is it 100% certain the x86 is from AMD ?

 

[Kaotik]

So which x86 is a realistic candidate for PS4 ? Time frame, power, size, cost and all ? Is it 100% certain the x86 is from AMD ?

If it's x86, it's quite definately from AMD (you know, package deal, buy one GPU, get a CPU free )
Depending on the timeframe when it comes, either Piledriver or Steamroller based CPU cores probably

 

[Rangers]

Charlie tweeted yesterday in regards to his recent ps4 and xbox720 rumor articles.
About PS4 using x86
https://twitter.com/#!/CDemerjian/st...37612361875459


About the xbox720 use of PowerPC:
https://twitter.com/#!/CDemerjian/st...50408038346752

I dont think Charlie knows anything...

Before his PS4 x86 article he wrote on twitter "Trust me on this one guys, after all I nailed the Xbox 720".

I'm thinking umm, what proof is there of that?

The bad thing is I sometimes kinda write in my head Charlie's correction articles when all his next gen reporting doesn't come true later. i've seen enough of Charlie's excuse articles lol.

"Well, sources tell us even though the PS4 doesn't use stacking technology, that was indeed the PLAN until Sony's last minute change (take my word on it!)". Followed by more archaic and detailed "explanation" charlie style, so of course he was never wrong.

 

[Rangers]

If it's x86, it's quite definately from AMD (you know, package deal, buy one GPU, get a CPU free )

Depending on the timeframe when it comes, either Piledriver or Steamroller based CPU cores probably

Well Bulldozer would be terrible in a console, and piledriver will be based on it.

Color me skeptical of all this stuff. I'd even wager a good chance Nvidia is the PS4 GPU.

 

[Rangers]

What I am saying is that one can not just claim that because of console optimizations an AMD GPU would be able to claim a performance advantage over an Nvidia GPU in rendering workloads due to somehow unlocking its full FP performance.

AMD having better performance in furmark just proves my point, it is a PC program and thus proves that AMD cards can use the FP advantage they have and are not being unduly held back by PCs.

I think it does prove it. You must understand because there's two PC GPU vendors, games generally do not target hardware. Hardware is molded to software on PC, the opposite is true on console.

AMD doing well on furmark and shader limited PC games shows they have a shader edge, that in most games goes idle. On console, nothing goes idle. Real simple. That extra shading power (being demonstrated in furmark) WILL be maximized, with more shaders on a console, whereas in PC the game can be texture, bandwidth, or who knows what else limited first, and nobody is going to code better shaders just for high end AMD cards on a PC game, because PC games are sloppy as hell to start with as they're targeting infinite hardware variations.

 

[Kaotik]

Well Bulldozer would be terrible in a console, and piledriver will be based on it.

Color me skeptical of all this stuff. I'd even wager a good chance Nvidia is the PS4 GPU.

Would it? Bulldozer isn't really that bad as long as you actually use all the possible 8 threads, and both Piledriver & Steamroller are going to improve various weaknesses of the Bulldozer core.
Think of it like Phenoms - the Phenom "1" was quite bad after all, but Phenom II fixed a lot of it's weaknesses, and while couldn't compete with the fastest Intels, it wasn't bad anymore.

 

[Blazkowicz]

Well Bulldozer would be terrible in a console, and piledriver will be based on it.

Color me skeptical of all this stuff. I'd even wager a good chance Nvidia is the PS4 GPU.

a two module, "fixed" bulldozer, without L3 and with a more mature process, would be a lot smaller and a bit better than the current pig reminiscent of R600 and Barcelona.

 

[hoho]

Would it? Bulldozer isn't really that bad as long as you actually use all the possible 8 threads

Now what if you calculate performance per transistor/watt? The CPU alone would take humongous chunk of transistor and power budget if it really was BD-based.

 

[Blazkowicz]

on a console you will get optimal scheduling and use of the FPU.
dunno if a three-module one without L3 would work out well.
not only there's L3, but a lot of uncore fat that AMD can cut.

 

[TheD]

I think it does prove it. You must understand because there's two PC GPU vendors, games generally do not target hardware. Hardware is molded to software on PC, the opposite is true on console.

AMD doing well on furmark and shader limited PC games shows they have a shader edge, that in most games goes idle. On console, nothing goes idle. Real simple. That extra shading power (being demonstrated in furmark) WILL be maximized, with more shaders on a console, whereas in PC the game can be texture, bandwidth, or who knows what else limited first, and nobody is going to code better shaders just for high end AMD cards on a PC game, because PC games are sloppy as hell to start with as they're targeting infinite hardware variations.

No, you must understand that games are not going to be made for a console with graphics that have the same workload as a program design to do nothing more than push ALUs to breaking point!
Not unless you want a game to look like furmark!
And even if you did want a game like that you are unlikely to get it with most games being multiplatform.

The fact is that to get an AMD GPU into a ALU limited scenario you would have to cut down on everything else that will limit it first and thus make a very below par looking game.

 

[-tkf-]

Very interesting, on the CPU side it will be something Console proven against something with a very bad reputation from the PC world.

Maybe someone with more knowledge can breakdown this: http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2 mostly i am wondering about this:

Compared to Sandy Bridge, Bulldozer only has two advantages in FP performance: FMA support and higher 128-bit AVX throughput. There's very little code available today that uses AMD's FMA instruction, while the 128-bit AVX advantage is tangible.

Is that something that a console could use to it´s advantage?

Right now i am thinking that the PS4 would really need something special from the GPU to make up for the AMD CPU.

 

[hoho]

Maybe someone with more knowledge can breakdown this: http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2 mostly i am wondering about this:



Is that something that a console could use to it´s advantage?

FMA is pretty logical.

Higher AVX128 throughput comes from AMD having the FPU shared and if two cores in a module only need 128bit instructions then both cores can run them simultaneously. If they use 256bit ones only one core can use the FPU at a time giving it similar 256bit throughput than what Intel gets at half the core count.

Basically I don't think that the shared FPU and higher 128bit throughput would be much use vs Intel offerings while FMA is obviously awesome. Then again I don't think Intel CPU would be a possibilty in a console anyway

 

[McHuj]

Very interesting, on the CPU side it will be something Console proven against something with a very bad reputation from the PC world.
.

While Bulldozer gets a really bad rap on the PC side, I don't think it's that bad of a design. A fairer indicator of the performance will the upcoming Trinity APU. Coincidentally, Fudzilla today claims that the top Trinity SKU (likely April launch) will have to 2 piledriver modules clocked at 3.8 GHz within a 100W TPD. We'll see how well piledriver cores address some of the issues and how good the performance/watt really is.

Perhaps at the eventual 28nm SOI shrink, they'll be able to increase performance another %25 or so and keep the same TPD (or push the TPD a little higher ~125 for the SOC, sub 200W for the whole system)