[hexus.net]AMD claims it will power another gaming device

[III-V]

So it's just a gaming processor that should be a very powerful single core. Got it.

Yep. That's totally what I was saying.

Wish we had higher standards for our users. It's unfortunate that moderation has gotten so lax in this forum. Such childish straw men arguments have no business here.

 

[ShintaiDK]

So then all processors should be a single, extremely powerful core, right?

Gaming wise yes.

A 15Ghz singlecore Haswell would run in circles around a 4790K in games.

The problem is to make the 15Ghz singlecore Haswell...hence more cores is the tradeoff rather than the solution.

 

[AnandThenMan]

Yep. That's totally what I was saying.

Wish we had higher standards for our users. It's unfortunate that moderation has gotten so lax in this forum. Such childish straw men arguments have no business here.

Then spell out your point. You say game code is highly serialized so an 8-core CPU is no good. Correct me if wrong but are you not saying this? So how many CPU cores is ideal exactly? One would seem to be the best going by your statement. In fact ShintaiDK seems to be saying exactly this, yet you're tossing out an insult towards me and calling out the moderation here?

A 15Ghz singlecore Haswell would run in circles around a 4790K in games.

No it would not because games have to run under an OS which uses processing cycles, multi core is perfect for this. And it also depends on the game code. There are a lot of background tasks going on especially with some games, were are not talking about an unconnected box here.

 

[III-V]

Then spell out your point. You say game code is highly serialized so an 8-core CPU is no good. Correct me if wrong but are you not saying this? So how many CPU cores is ideal exactly? One would seem to be the best going by your statement. In fact ShintaiDK seems to be saying exactly this, yet you're tossing out an insult towards me and calling out the moderation here?

The world is not black and white. Simply because I have pointed out that implementing one end of the spectrum is sub-optimal, does not mean that I think the complete opposite end of the spectrum is optimal.

I shouldn't have to spell it out -- I was hoping my audience (not you) -- would be smart enough to know the above. Perhaps I should have spelled it out, to prevent garbage responses from people such as yourself.

Games do benefit from multiple cores to some extent (and certainly not marginal one), but to say what is ideal for supercomputing is also ideal for gaming consoles is patently false.

Shintai has gone off the deep end, honestly. I endorse few of his opinions these days, and consider him to be part of the problem (just on the opposite faction). He seems to view the world as being black and white much as you do. In this case, he is mostly correct, but he is ignoring the costs associated with context switching.

In regards to the current subject at hand, a console would benefit best, at minimum, from two cores. I would personally guess that the ideal core count would be between 2 and 4 strong cores (leaning 3 or 4, with 4 being my personal choice given that it's a core count found in far more systems). Prioritizing core count such that you have to use 8 anemic cores, rather than 2-4 strong ones, seems to be a bad choice. 1 is certainly too few, as is a supercomputer's worth of cores.

No it would not because games have to run under an OS which uses processing cycles, multi core is perfect for this. And it also depends on the game code. There are a lot of background tasks going on especially with some games, were are not talking about an unconnected box here.

The more area you devote to running these individually-trivial background tasks, the farther you get away from what should be the primary task at hand -- that you have frames to render.

 

[AnandThenMan]

The world is not black and white. Simply because I have pointed out that implementing one end of the spectrum is sub-optimal, does not mean that I think the complete opposite end of the spectrum is optimal.

I agree. But you made a black and white statement about game code being highly serialized. At one time yes it was, not anymore a game engine is doing more than just throwing pixels at a screen with the imagery being an object driving on a rail. I've heard several devs say that past 5-6 cores things get hazy. So actually 8 cores sounds about right because you need a couple of them fo the OS housekeeping and whatever other background stuff is always going on. The current consoles are not just game machines remember that.

I shouldn't have to spell it out -- I was hoping my audience (not you) -- would be smart enough to know the above. Perhaps I should have spelled it out, to prevent garbage responses from people such as yourself.

Games do benefit from multiple cores to some extent (and certainly not marginal one), but to say what is ideal for supercomputing is also ideal for gaming consoles is patently false.

The trend is to more cores, I think you'd agree with that. So maybe right now 4 cores is ideal, going forward I'm willing to bet that 8 cores is better. More ideally 8 very strong cores would be nice, but heat/power/cost in a game console is extremely important. I would have liked to see Sony/MS clock the APU higher, I'm willing to bet that 2Ghz would be a fairly easy target. But they went conservative so they must have had their reasons.

In regards to the current subject at hand, a console would benefit best, at minimum, from two cores. I would personally guess that the ideal core count would be between 2 and 4 strong cores (leaning 3 or 4, with 4 being my personal choice given that it's a core count found in far more systems). Prioritizing core count such that you have to use 8 anemic cores, rather than 2-4 strong ones, seems to be a bad choice. 1 is certainly too few, as is a supercomputer's worth of cores.

Essentially here you're saying use an Intel CPU rather than AMD. That's fine except there is no Intel CPU that contains the graphics capability needed. So what you rather have, 8 slower cores and good graphics, or 4 better cores and and a GPU section that would be completely unusable.

The more area you devote to running these individually-trivial background tasks, the farther you get away from what should be the primary task at hand -- that you have frames to render.

The "background" tasks are not trivial they are part the overall system and experience. The PS4 for example multi tasks extremely well have you played around with one? Plus the games are doing all manner of things like AI, voice, network, physics. This sounds more like a supercomputing load than single purpose game code that is only there to chuck pixels on the screen.

 

[Lepton87]

Gaming wise yes.

A 15Ghz singlecore Haswell would run in circles around a 4790K in games.

The problem is to make the 15Ghz singlecore Haswell...hence more cores is the tradeoff rather than the solution.

It would provided that it would juggle 4-8 threads at once so it wouldn't be a HW. With just two threads it would have problems in some games with context switches. Some games just don't run with 2 threads these days without stutters.

 

[Cerb]

2 threads would be enough, for gaming, to handle stalls and context switches. We have games running 4+ threads because they need to to make use of modern CPUs, not because it would be ideal, if faster CPUs existed. Concurrency does not require many cores. If we had such CPUs, those badly coded games would have other problems, instead of stutters on duallies, because they would not have been developed and tested with more threads as their baseline.

 

[III-V]

I agree. But you made a black and white statement about game code being highly serialized. At one time yes it was, not anymore a game engine is doing more than just throwing pixels at a screen with the imagery being an object driving on a rail. I've heard several devs say that past 5-6 cores things get hazy. So actually 8 cores sounds about right because you need a couple of them fo the OS housekeeping and whatever other background stuff is always going on. The current consoles are not just game machines remember that.

That's not a black and white statement at all. The word "highly" does not mean "entirely."

It of course, would be ideal to have as many cores as there are threads that you'll ever come across, if we ignore a lot of realities of economics, power, and such, and it seems you recognize this, seeing as you state as much below.

The trend is to more cores, I think you'd agree with that. So maybe right now 4 cores is ideal, going forward I'm willing to bet that 8 cores is better. More ideally 8 very strong cores would be nice, but heat/power/cost in a game console is extremely important. I would have liked to see Sony/MS clock the APU higher, I'm willing to bet that 2Ghz would be a fairly easy target. But they went conservative so they must have had their reasons.

Right, I don't disagree with any of this.

Essentially here you're saying use an Intel CPU rather than AMD. That's fine except there is no Intel CPU that contains the graphics capability needed. So what you rather have, 8 slower cores and good graphics, or 4 better cores and and a GPU section that would be completely unusable.

Not at all. Intel obviously wasn't an option. I think it's laughable that you are constantly making the case for AMD, but are seemingly ignorant of their own offerings.

I actually think AMD had the best product portfolio... but choosing 8 Jaguar cores over 4 Steamroller cores was a mistake, in my opinion.

 

[Cerb]

The "background" tasks are not trivial they are part the overall system and experience. The PS4 for example multi tasks extremely well have you played around with one? Plus the games are doing all manner of things like AI, voice, network, physics. This sounds more like a supercomputing load than single purpose game code that is only there to chuck pixels on the screen.

All of these would run faster with a fast core with a small number of threads. They are split out to other threads because those threads and cores exist in hardware. Older games did all the same things, but on one core, usually in one thread, without all the concurrency/sync issues of more threads. They are using more because we don't have those faster cores, and no way to get them on the horizon.

 

[jpiniero]

For MT perf/watt, esp at a low TDP, Jaguar is way better.

 

[Cerb]

I actually think AMD had the best product portfolio... but choosing 8 Jaguar cores over 4 Steamroller cores was a mistake, in my opinion.

I can't find the actual core sizes, ATM, but an entire Kabini is a bit under half the size of an entire Kaveri, and even Kaveri would be impossible to cool efficiently. The Jaguar cores+cache would need to average, what, <10mm^2/ea? A full size fat core would be too much. While some devs have been stepping on their cranks, overall, I'd say it was a fairly good choice, given that most current console games came out looking and behaving pretty decently, while that took 2-3 years for the prior gen, and they still had issues with map/level sizes, except for Crytek, NPC counts, etc., etc..

If lucky, they might have been able to include 1 BD/PD/SM module, but I think the more Jaguars was definitely better (taking 2/6 cores for the OS, and 3GB RAM, was stupid, though, especially since they knew they'd be able to do better, over time).

 

[turtile]

Jaguar is 3.1mm2 per core.

 

[Cerb]

Jaguar is 3.1mm2 per core.

Is that including the cache management block (forgot what they call it) and 1/4 of the cache slices?

 

[itsmydamnation]

No, it's not valid. Supercomputers compute problems that are easily parallelizable. Game code is largely serial, not parallel, in nature.

Go tell that to the actual next gen game developers maybe people like sebbbi on B3D who directly contradicts your "statement" and have actually released titles on the X1/PS4.

Games that scale have moved to asynchronous job systems, the latest ones being built don't even have a "main" thread and have QoS for different job types etc.

The other annoying thing is calling jaguar weak, it isn't. its FP is narrow, but its integer performance is very strong, insanely stronger then the previous generation of consoles. But that means nothing for throw away marketing numbers. But it means a lot for complex branching code for things like AI. It easy for the keyboard warriors to say otherwise but chances are if a 2 module PD/SR was going to be faster for there expected workloads then they would have picked it.


the last gen consoles had very high peak numbers with very big glass jaws ( you have to do software based prefetching yourself!) . The new consoles have much more modest peak numbers with a much lower delta between peak and average and only really one glass jaw ( latency between cpu clusters). Unfortunately those numbers are easily manipulated to ones agenda.......

 

[turtile]

Is that including the cache management block (forgot what they call it) and 1/4 of the cache slices?

Just the core. The entire 4 core module is 26.2mm2.

 

[Cerb]

Go tell that to the actual next gen game developers maybe people like sebbbi on B3D who directly contradicts your "statement" and have actually released titles on the X1/PS4.

The two are not the same. Between long-running tasks and the dependency tree created, there will be fairly low upper bounds. The scale of each is radically different (off by multiple orders of magnitude), and the supercomputer doesn't care about latency. Games can run multiple subtasks concurrently, but there's only so far it can be split out, many of those tasks cannot themselves be split up, and the overhead of doing the splitting is not insignificant. Tasks given the GPU are those that may be super-computer-like. Most everything else needs fast cores, however many threads it may go out to. Compared to other consoles of days past, the Jaguars are pretty fast, even if they are weak from a PC PoV.

Such job queue systems for concurrent tasks is just something some crazy people at Ericsson would have say, "told you so," about, having done it well 20 years ago. It's not parallelism is the same sense as the work on pixels is, FI, that can basically scale out to infinity.

Just the core. The entire 4 core module is 26.2mm2.

Steamroller is around 30mm^2, IIRC, so not far off, w/ 1 module v. 4 cores. Given that even Kaveri would be too power-hungry per module, and there aren't competitive ARM or PPC cores to use, that's an easy choice.

 

[turtile]

Steamroller is around 30mm^2, IIRC, so not far off, w/ 1 module v. 4 cores. Given that even Kaveri would be too power-hungry per module, and there aren't competitive ARM or PPC cores to use, that's an easy choice.

Kaveri was never a choice because it was not ready in time for the consoles. Richland was 32nm while the GCN architecture was based on 28nm. Jaguar made perfect sense at the time because it was the only CPU architecture designed for 28nm.

 

[Cerb]

Kaveri was never a choice because it was not ready in time for the consoles.

Yes, but I couldn't be bothered to look up how much worse prior fat APUs were, power-wise, after trying to find Jaguar 4-core total size . But if Kaveri isn't up to it, they definitely weren't.

 

[III-V]

Kaveri was never a choice because it was not ready in time for the consoles. Richland was 32nm while the GCN architecture was based on 28nm. Jaguar made perfect sense at the time because it was the only CPU architecture designed for 28nm.

AMD could, and would have easily accelerated their development of Steamroller to meet the console launch deadline.

The other annoying thing is calling jaguar weak, it isn't.

http://dictionary.reference.com/browse/relativity

See definition 1.

It easy for the keyboard warriors to say otherwise but chances are if a 2 module PD/SR was going to be faster for there expected workloads then they would have picked it.

I'd be happy if their 8 core Jaguar configuration were the optimal one -- that would mean that the consoles would have the best feasible hardware they could have received. I just personally believe that a quad core Steamroller configuration would have been more optimal (as far as Steamroller's readiness, see my response to turtile). There is no need to be reacting in the manner you have.

In some fantasy world, a quadcore Haswell setup alongside either GCN or Kepler (or really, Maxwell -- had it been ready in time -- would have been very nice) would be the ideal configuration, but of course that's not exactly something that could have happened. It's possible a POWER design would have been better as well (Intel vs IBM is a battle that I have not seen a clear answer as to who the "winner" is), although this would kill porting advantages.

 

[pw257008]

As far as perf/clock goes, Jaguar is not far behind AMD's fat core. It can't clock as high (at least while maintaining efficiency), but at 2G and below, it's definitely more power efficient than AMD's big cores.

 

[III-V]

As far as perf/clock goes, Jaguar is not far behind AMD's fat core. It can't clock as high (at least while maintaining efficiency), but at 2G and below, it's definitely more power efficient than AMD's big cores.

I don't disagree.

 

[itsmydamnation]

the lack of turbo in jaguar helps flatter other CPU's relative to it, but in a console you don't want variable performance. Also remember instruction latencies are very low on jaguar especially the AVX instructions, so that can help serial code.

As far as perf/clock goes, Jaguar is not far behind AMD's fat core. It can't clock as high (at least while maintaining efficiency), but at 2G and below, it's definitely more power efficient than AMD's big cores.

clock for clock a SR with L3 cache would likely win significantly in games, but compared to PD in the APU's its pretty close, SR is ahead.

 

[krumme]

Jaguar was highly synthesized - pd wasnt. Nothing but highly synthesized cores is economically compettitive for this and future consoles.
Look at the team sizes. Besides using highly synthesized cores they reused existing building blocks and ip from either amd or inside. Eg sound/security.

 

[SPBHM]

The other annoying thing is calling jaguar weak

Digital Foundry on their latest article doesn't sound so impressed by the Jaguar CPU
http://www.eurogamer.net/articles/digitalfoundry-2014-xbox-one-vs-ps4-year-one

basically blaming it for the lack of innovation on the next gen games, so I think you would find a lot of people considering jaguar weak, specially when combined with such a strong GPU on the PS4.

 

[MisterLilBig]

No, it's not valid. Supercomputers compute problems that are easily parallelizable. Game code is largely serial, not parallel, in nature.

What game code can't be made parallel? Or easier, what game code must only be serialized?

Games do benefit from multiple cores to some extent (and certainly not marginal one), but to say what is ideal for supercomputing is also ideal for gaming consoles is patently false.

Why?

The more area you devote to running these individually-trivial background tasks, the farther you get away from what should be the primary task at hand -- that you have frames to render.

It's part of the design of these consoles.


You know for all the comments that AMD had better cores to choose from, it's as if no one saw/read that test that Jaguar has like 95% of the IPC of one of the big cores, at same clocks of course.