Predict: The Next Generation Console Tech

[Gubbi]

For the future CPU: The CPU would likely be more future proof, if we had more cores / hw threads running at lower frequency (or even in-order) compared to high IPC / high clock CPU with just a few cores. 8 cores with SMT = 16 hw threads, would be a really good starting point. Next generation game engines should have no problem in splitting tasks to that many threads. More cores and less IPC seems to provide more total performance, assuming that the software fully utilizes all the cores.

IMO, the chance of seing another in-order design is small, - even ARM has gone out of order with Cortex A9 and A15. For a given level of (non-trivial) performance, you're better off going out of order. The only way for an in-order design to match the performance of an out of order is to lower data-dependency latencies, either by clocking higher or by burning more power on faster caches.

The PPE/XCPU was clearly influenced by the work at IBM that lead to Power 6, being deeply pipelined, quite narrow, multi-threaded and clocking super fast. Unfortunately performance didn't match the exotic power consumption. With Power 7 IBM has done a full 180 degree turn; it is wide, out of order and has much better single thread performance, - and each core supports 4 threads.

I expect MS to charter IBM to do a design again. I could imagine a core designed around the scheduler of the Power 7, - or something very similar. I expect 8 cores, each with 4 threads, clocked at 3-3.5GHz.

Cheers

 

[liolio]

IMO, the chance of seing another in-order design is small, - even ARM has gone out of order with Cortex A9 and A15. For a given level of (non-trivial) performance, you're better off going out of order. The only way for an in-order design to match the performance of an out of order is to lower data-dependency latencies, either by clocking higher or by burning more power on faster caches.

The PPE/XCPU was clearly influenced by the work at IBM that lead to Power 6, being deeply pipelined, quite narrow, multi-threaded and clocking super fast. Unfortunately performance didn't match the exotic power consumption. With Power 7 IBM has done a full 180 degree turn; it is wide, out of order and has much better single thread performance, - and each core supports 4 threads.

I expect MS to charter IBM to do a design again. I could imagine a core designed around the scheduler of the Power 7, - or something very similar. I expect 8 cores, each with 4 threads, clocked at 3-3.5GHz.

Cheers

On which width would you put your bet for the SIMD width? I feel like those questions are tied.
For example I feel like 4 cores with 8 wide SIMD would be than 8 with 4 wide SIMD. Scaling the perfs accordingly to the number of cores is costly too.
I was thinking (to my self as nobody found the hypothesis worse a talk) in a previous post about the relevance of a many-core taking distance with larrabee's failure as it's unclear if the concept is completely fucked up or while having clear disadvantage it also suffer from bad decisions and execution on Intel side. Anyway I won't go back to this.
My point is that Ms were to ask IBM a CPU I would favor them going with really powerful SIMD (like 16 wide) and a limited number of cores like 4. Scaling perfs across multiple cores is not trivial either Intel last architectures do very well in regard to AMD ones. If power consumption is a problem they might lower the clock speed and add a proper turbo which may tick on depending on the SIMD utilization.
Recycling the Power7 scheduler front end may be a good option.

 

[3dilettante]

The PPE/XCPU was clearly influenced by the work at IBM that lead to Power 6, being deeply pipelined, quite narrow, multi-threaded and clocking super fast. Unfortunately performance didn't match the exotic power consumption. With Power 7 IBM has done a full 180 degree turn; it is wide, out of order and has much better single thread performance, - and each core supports 4 threads.

Some of the circuit techniques that went into POWER 6 were tested with Xenon.
Microsoft did want to go OoO and IBM was willing. The problem was the time to market, which was too short to validate an OoO design.

I expect MS to charter IBM to do a design again. I could imagine a core designed around the scheduler of the Power 7, - or something very similar. I expect 8 cores, each with 4 threads, clocked at 3-3.5GHz.

That might be too big, and Power 7's design is more focused on server loads to start with.

 

[Blazkowicz]

so are the previous POWER chips written in all capitals.
a derivate of one such big chip was the PowerPC G5, which was an apt desktop processor (if a power wasting one).
a Phenom II is perhaps a "server processor" as well ; Bulldozer most certainly is.

I agree it's huge. a 4 core one with cruft taken out (such as decimal computing hardware) and much less L3, at 22nm may turn out excellent.
I'm sure IBM could make desktop CPUs as well as Intel and AMD, they do don't it because of market reasons not technical ones (with only the IBM PC compatible and Mac remaining, for over a decade)

 

[patsu]

Some of the circuit techniques that went into POWER 6 were tested with Xenon.
Microsoft did want to go OoO and IBM was willing. The problem was the time to market, which was too short to validate an OoO design.



That might be too big, and Power 7's design is more focused on server loads to start with.

3dilettante, what are the inherent advantages that make ARM more suitable than POWER in devices ?

 

[liolio]

so are the previous POWER chips written in all capitals.
a derivate of one such big chip was the PowerPC G5, which was an apt desktop processor (if a power wasting one).
a Phenom II is perhaps a "server processor" as well ; Bulldozer most certainly is.

I agree it's huge. a 4 core one with cruft taken out (such as decimal computing hardware) and much less L3, at 22nm may turn out excellent.
I'm sure IBM could make desktop CPUs as well as Intel and AMD, they do don't it because of market reasons not technical ones (with only the IBM PC compatible and Mac remaining, for over a decade)

Honestly I'm not sure, Intel impresses me more and more, cache density, cache latency, everything they used to call the "uncore", they provide next to unmatched single thread performance even the power consumption in regard to the provided perfs, to me they start ot look way ahead of every one. Not too mention they design their CPUs not for a "specific" use.

 

[Heinrich4]

About ARM A15 ..40% more powerful than A9:

http://www.itproportal.com/2011/03/14/exclusive-arm-cortex-a15-40-cent-faster-cortex-a9/

Interesting benchmark for A9 and others cpus:

http://www.coremark.org/benchmark/index.php

PowerVR 6 Rogue 20 * SGX5?

http://www.itproportal.com/2011/02/14/exclusive-imagination-technologies-series6-20x-faster-series5/

Microsoft Becomes An ARM Licensee.
http://www.itproportal.com/2010/07/23/microsoft-becomes-arm-licensee/

ARM Working With Microsoft On "Processor Architectures"
http://www.itproportal.com/2011/03/23/arm-working-microsoft-processor-architectures/

 

[(((interference)))]

In what resolution demo running?

1080p

IIRC huge part of the performance in the UE demo was soaked up by their new AA algorithm that was not optimized at all.

I don't know about that, they're only doing 4xMSAA.
http://www.nvidia.com/content/PDF/GDC2011/Epic.pdf

Didn't Epic stated that with a lot of optimizations they could get the thing to run on a single card? (If we speak of console there is rooms for extra optimizations).
The sad thing is that the GTX580 is a monster on any accounts and it most likely still be quiet a chip @28 nm.

Yes, they did say that it could be possible on just one card.

It will be disappointing if the 8th gen GPUs aren't better than what we have today - a step back from this gen with the likes of Xenos (and RSX+CELL)

 

[msia2k75]

1080p



I don't know about that, they're only doing 4xMSAA.
http://www.nvidia.com/content/PDF/GDC2011/Epic.pdf





It will be disappointing if the 8th gen GPUs aren't better than what we have today - a step back from this gen with the likes of Xenos (and RSX+CELL)

What do you mean by that?

 

[AlphaWolf]

I believe he means that if the next generation consoles (coming out 18 months to 3 years from now) aren't capable of performing as well as something running on a PC today, it would be disappointing.

I don't expect a console to be even with a bleeding edge PC at launch (not on all counts anyway), but better than what was bleeding edge 2 years ago probably isn't too much to ask.

 

[(((interference)))]

I believe he means that if the next generation consoles (coming out 18 months to 3 years from now) aren't capable of performing as well as something running on a PC today, it would be disappointing.

I don't expect a console to be even with a bleeding edge PC at launch (not on all counts anyway), but better than what was bleeding edge 2 years ago probably isn't too much to ask.

Yes, especially when you consider Xenos and the RSX+CELL combo were actually better than what was available on PCs back in 05/06.

 

[hoho]

I don't expect a console to be even with a bleeding edge PC at launch (not on all counts anyway), but better than what was bleeding edge 2 years ago probably isn't too much to ask.

Probably the biggest problem with targeting bleeding-edge is that there is a HUGE gap between high/midrange (single top-of-the line -1 level GPU) vs bleeding edge (3x highest GPUs, possibly dual GPUs each).

Yes, especially when you consider Xenos and the RSX+CELL combo were actually better than what was available on PCs back in 05/06.

Most definitely they weren't, at least when you didn't offload big part of your graphics work to CPU in the console

 

[Gubbi]

That might be too big, and Power 7's design is more focused on server loads to start with.

I didn't mean to imply that it would be a direct P7 derivative. There is no need for decimal floating point for example, a lot of the RAS features could be stripped off as well, and I don't see a need for 4 DP FP ops per cycle either, one or two VMX units would suffice.

A core built around a scheduler with four distinct queues for 4-way SMT just seems reasonable to me.

Cheers

 

[Arwin]

at least when you didn't offload big part of your graphics work to CPU in the console

But that was exactly his point.

 

[hoho]

But that was exactly his point.

So basically it all comes down to how much time you spend on optimizing for your target hardware. Raw performance-wise PCs were way ahead of consoles by the time PS3 launched, just no one did any serious optimizations for them.

 

[Shifty Geezer]

Yes, especially when you consider Xenos and the RSX+CELL combo were actually better than what was available on PCs back in 05/06.

IMPORTANT QUALIFICATION NEEDED!

Before this thread devolves into a console vs. PC mess, "PCs" is a very big and varied target, and "better" needs qualifying. Are you comparing to a PC of the same price? The same size? The very best PC you could build, no holds barred? Yadayada.

So basically, your remark isn't a sensible discussion point and you need to refine it. Or maybe just drop it given the topic. Suffice to say consoles have traditionally packed in a lot power for the price at launch to remain technically competitive for years in a world of ever improving technology, and you will be disappointed if next-gen this standard practice isn't maintained and the hardware launches well behind the cutting edge of what's possible.

 

[RudeCurve]

If Apple were to enter the console space in the next couple of years what kind of hardware do you think they'd include in the box?

And before someone mentions it, no the Bandai Pippin wasn't a very serious effort.

 

[Shifty Geezer]

Comparison of current gen consoles with launch-period PCs deleted for obvious reasons. See my above post if it's not so obvious to you.

 

[rekator]

If Apple were to enter the console space in the next couple of years what kind of hardware do you think they'd include in the box?

And before someone mentions it, no the Bandai Pippin wasn't a very serious effort.

Already enter, and the strategy is simple iPhone, iPad, and something for wireless streaming on TV.

 

[RudeCurve]

Already enter, and the strategy is simple iPhone, iPad, and something for wireless streaming on TV.

Those aren't consoles, they're cellphones and computer tablets that can also play games.