AMD - Piledriver chitchat

[Biscuit]

i wonder what MS will use for the next xbox, would be interested if that was based on a similar platform

[watercooled]

There seem to be very mixed opinions on the next Xbox. Originally, people suspected x86 like the PS4, then that leaked slide said ARM, but I've also heard POWER-based cores taped out a while back for the dev kits. Agreement seems to be it will be using an AMD GPU though.

[DanceswithUnix]

Charlie seems to think it is an x86 chip in the xbox.

http://semiaccurate.com/2012/09/04/m...rumors-abound/

Now he could be wrong, given the ties AMD and IBM have had in the past with things like SoI fabrication, if AMD can choose between ARM and x86 on an APU then there is no reason they couldn't stick a few cores of Power CPU in there as as option. It could still appear to be an AMD chip.

There seems to be another angle the journos and analysts are missing here as well. AMD said ages ago that they are embedding an ARM core to handle their encrypted boot, as phones have been able to do that for years. My guess is they have to do that else Sony won't want their APU for the PS4. So that gets you an x86-64 and an ARM chip in the same package. Sounds rather like the recent "announcement", yet no-one seems to think this is incremental at best to what was announced ages ago. Perhaps journalists drink too much and forget

[watercooled]

Never know, it could quite easily end up containing all three types of cores.
E.g. POWER for legacy support, at least on initial runs, ARM for security etc, and x86 for application.

Typical x86 cores lack the FPU grunt we've come to expect on console CPUs, but that could explain the whole APU thing.

As for the PS4, I've heard that AMD might have been implementing ARM cores since Trinity or possibly Llano, just they're not being used yet. They're small enough to not add much to die size/cost and would be very hard to detect on die shots. But console MFRs have always used modified designs; it might speed things up if AMD already have it implemented, but well within reason to expect Sony would add whatever extras they wanted anyway. I really can't see them using a part straight off-the-shelf, especially considering the extra requirements of console security systems, but I could be wrong. The original Xbox is the closest we've gotten to using an off-the-shelf part but the CPU was still a modified cross between Pentium and Celeron.

I'm just thinking 'out loud' here BTW...

[DanceswithUnix]

Typical x86 cores lack the FPU grunt we've come to expect on console CPUs, but that could explain the whole APU thing.

You think so? I have been quite unimpressed with my PS3 generally, and the whole "Cell architecture will take over the world" certainly didn't happen.

Little things like trying to use the ITV catchup service on the PS3 looks awful compared to what my Brazos E350 netbook can do.

[watercooled]

No I didn't mean it quite like that, and I agree about the Cell in general, but the 'PPE', similar to the core used in the 360, have high theoretical FP performance vs even modern x86 CPUs. This obviously doesn't translate directly into performance for all applications, but it's important for the games they run. If the next consoles are planning to offload a lot of the floats to the GPU, the BD uarch could be a good choice, and hopefully it will mean better optimisation for games on AMD CPUs and PCs in general.

[scaryjim]

The thing with consoles is the software is run much closer to the metal than on a PC. So PCs almost certainly don't get close to the peak FLOPS of any x86 CPU currently available. Sensible optimisations in software could comfortably see an A10-like APU running 1080p60 with high image quality, just by coding specifically for the hardware. After all, DIRT showdown is a fantastic example of what you can do - in both the positive (AMD) and negative (NV) lights - depending on optimisation of codepaths...

EDIT: also, another thought's just occurred to me: AMDs current architecture is meant to be entirely modular; so what if you made an APU with a 400+ GCN core GPU, and a separate block of 64 GCN cores specifically for running highly parallelisable FPU operations on (with HSA and unified address space to minimise memory bottlenecks)? Piledriver + GPU + FPU coprocessor == ???

[HalloweenJack]

watched and read a very interesting article about the Tian super computer:

http://www.anandtech.com/show/6421/i...idia-gpu-cores

18666 16 core bulldozers running on a totally optimized OS - aiming for the worlds fastest supercomputer

[scaryjim]

18666 16 core bulldozers

More to the point 18,688 GK110 cores - you know, the one that never made it to the consumer space! Apparently, "The K20 features 2688 CUDA cores".

That's 75% more raw compute units than GTX680. Imagine if they ever *had* brought GK110 to the GeForce line...

[watercooled]

It's a very heavily compute-oriented GPU from whet I've read; yields are apparently very poor and the frontend isn't likely beefed up that much vs. the shaders, so if it does make it to the consumer space, I'd expect a very expensive card with more cores disabled for yield (selling off bad Tesla dies essentially, and even K20 is part disabled) and a relatively slight gaming performance increase. I.e. based on what I know, not exactly a great value card...

However, considering how many K20's they're rolling out, they probably have tons of lower binned dies available to sell off. Even so, I thought they were planning a Kepler refresh soon, so I wonder where it would fit in?

[Biscuit]

More to the point 18,688 GK110 cores - you know, the one that never made it to the consumer space! Apparently, "The K20 features 2688 CUDA cores".

That's 75% more raw compute units than GTX680. Imagine if they ever *had* brought GK110 to the GeForce line...

So is that what the 680 was supposed to be, but they shelved it because the AMD chip was a bit of a 'disappointment'?

[Terbinator]

Are the K20's not just two GK104's in SLI? Or are those the K10's?

[watercooled]

So is that what the 680 was supposed to be, but they shelved it because the AMD chip was a bit of a 'disappointment'?

Read my post above yours, or search the net for more info on it, that's just not the case.

Edit: realised you may have been implying just that?

[scaryjim]

I'm not sure GK110 was ever meant to be a consumer chip. If you look at the computer performance of the GTX6x0 range they're all pretty ordinary. I suspect GK110 has all sorts of compute-specific features that make it more suitable for compute but less suitable for pumping pixels.

Interestingly SemiAccurate has some commentary on the specs today. It's a little confusing though, as both the reported specs, and the total number of compute unit at ORNL, are different in the German story SA are commenting on as compared to the Anandtech visit to ORNL. Go figure. To me the Anandtech figures look more sensible, but you never can tell.

Are the K20's not just two GK104's in SLI? Or are those the K10's?

The K10 is two GK104s. More information here: http://www.nvidia.com/content/tesla/...jul2012-lr.pdf The difference between single-precision and double-precision flops on the K10 gives you an idea of how compute-stunted GK104 really is. As a comparison, AMD's best single GPU compute card, the S9000, can deliver over 4x the double precision performance of K10...

Odd that they've shipped K20s to ORNL, a high profile HPC customer, but still haven't official announced the specs of it ... makes you wonder if ORNL have got cherry picked dies that can run more SMX than the "standard" K20s...

[DanceswithUnix]

No I didn't mean it quite like that, and I agree about the Cell in general, but the 'PPE', similar to the core used in the 360, have high theoretical FP performance vs even modern x86 CPUs. This obviously doesn't translate directly into performance for all applications, but it's important for the games they run. If the next consoles are planning to offload a lot of the floats to the GPU, the BD uarch could be a good choice, and hopefully it will mean better optimisation for games on AMD CPUs and PCs in general.

Understood, but just not convinced.

Consoles, and Playstations in particular, have a reputation of huge theoretical performance that you will never get near in real life.

AMD rate the Trinity 5800 CPU as 121.6 GFLOPs, slightly more than the entire cell CPU of which I think most devs could only really use the 25.6 GFLOPs of the main Power architecture unit.

BTW, AMD rate the GPU part of a 5800K at 800MHz as 0.8x384x2 = 614.4 GFLOP, only a factor of 5 more than the CPU. Am I the only one who thought it would be more unbalanced than that?

Edit to add: I know Trinity is new and Cell isn't, but AMD floating point hasn't really changed much in throughput for years as people on here keep complaining

[scaryjim]

... AMD rate the GPU part of a 5800K at 800MHz as 0.8x384x2 = 614.4 GFLOP, only a factor of 5 more than the CPU. Am I the only one who thought it would be more unbalanced than that?

I think the flexFPU can carry out more complex calculations than the VLIW4 cores, and of course they're bumped over 4x as fast.

Hmmm, by my calcs they reckon each FPU can theoretically push 16 operations per clock cycle. That must be a *lot* fused instructions - presumably made possible by FMA3/FMA4/AVX etc.