Current Generation Hardware Speculation with a Technical Spin [post GDC 2020] [XBSX, PS5]

And the "general shape" is grossly misleading. This is the reason I keep posting. It is unfortunately about the lack of openly published information on this but in the technical forum, published guesswork is a poor substitute.

None of this has any bearing here. You are trying to express power draw using a linearly scalable expression where draw is always a constant which isn't the case with FinFET semiconductors. If you don't understand this, I can't help you. This is way beyond my ability/time to reach somebody in a forum largely because of the point above. I spent many years attaining this discipline then almost another dozen years honing my craft in aerospace industry. You're either immersed in this level of engineering or you're not. It's not something you can learn in a few days or weeks. ¯\_(ツ)_/¯

With respect, this statement shows how little your know on this.

I'm leaving this here. You can either accept your'e strayed into a very complex area of engineering for which you don't comprehend how wrong you are, or you can just carry on.

 

I accept and know that there's way more than I understand or know about this topic. I've only a bachelors in electrical and that was years ago.
But grossly misleading and generalization are very different things. it brings into focus some objectivity where there is largely subjectivity.

If we cannot have discourse in this method, there is no value in discourse at all. There is no technical discussion, it's a discussion on how we interpret marketing words.

 

I take your point, but here is no subjectivity on measuring power draw of FinFET transistors. The makeup of the transistors and their relationship to each other, is calculable but the resultant equations are incredibly complex any equation would be valid only for for a specific design. If it were simpe, Sony wouldn't be using this paradigm They're not dong this because they think it's cool

 

I fully understand where you are coming from. But I don't like your answer even though your answer is correct, because I would rather rely on basic equations that are not fully representative of a chip, than to just black box things and believe that magic can happen that can sudden change the characteristics such that these basic equations are completely invalidated.
The basic checks are still useful to determine if we are entirely out of bounds at least in my eyes mind.

And I get I'm far from educated enough to talk about IC designs or any of this, it's been well over a decade since the last time I worked with mosfets and FPGAs but I don't believe that Sony has crossed any barriers or made any paradigm shifts with it's clocking mechanism such that we can't use historical information to apply it to what we expect from the behaviour of their chip.

I'm more than willing to wait to see what the results are in 2-3 months time.

 

Why are people so fixated on the GPU power consumption/clock.

I think its obvious that the driver for overall clock speed of the GPU will be AVX load , light 128bit vs dense 256bit 8x Zen2 core @ 3.6ghz doing dense 256bit FMA will consume 50watts. Doing 128bit mul ,add etc will be something like 15- 20.

That 30-35 watt difference is really what we are talking about here.
Im sure sony could have chosen just to throttle the CPU but that would be a very poor outcome ( see avx512)

 

Thanks for the example Function.
But let me ask:
Cerny sayd "might not want", or "will not"? They might be quite different.
But note Microsoft is talking about full clocks at 3.8 Ghz, and that is without SMT. Can they do it with SMT?
SMT can give you 30% extra performance, so reducing that clock speed in 30% will not result in the same performance? And if you cannot do it with SMT, isnt the choice of not using it not damaging overall performance , or can you active and deactivate SMT on the fly?
Also, could you point me to some link where they (Microsoft) make that claim about AVX. I'm not finding anything.

 

Just to be clear, I'm not invalidating the basic equations because they are what they are, what I'm saying is these linear equations are gross simplifications and not representative for determining power usage in the APUs in consoles. They're not only wrong, they're fundamentally misleading. My professor was one of those old school types who instilled into me that it's always better to acknowledge a complex issue than try to simplify it, particularly when it's the basis for higher-level discussion. Acknowledging complexity and your inability to express it results in better engineering decisions. To put this in perspective, the equations you posted enough are a fraction of the size of accepted equation for determining sub-threshold leakage of a single in FinFET SRAM transistor. That's one equation from about 10,000 somebody needs to accommodate to produce an equation to express the overall power draw.

On a more practical basis, ask yourself why when a CPU or GPU is virtually 'idle', is the power draw still high relative to high load power consumption? It's practically obscene, now why is this if there are few transistor state changes? It's almost like power draw isn't linearity scalable to transistor state changes

I get it's frustrating that there isn't more in open journals but that's just the way it is.

 

Man... thanks a lot for the time you took to explain this.
Going to read it carefully.

 

We just don't know. We do not even know how powerfull really the PS5 version or the xbox version is. Everything (even on developer conferences) is still heavily influenced by marketing . In the end, both companies just want to "sell" their hardware to developers so they develop for their plattform.

But like I wrote, I don't think sound will make a big jump, just because it is way to complicated to optimize there. To many different sound systems, ... make it almost impossible to optimize for something.

 

I still have not found anytany site with this claim but if this is true, then its a Microsoft change to Zen 2, since AMD claims Zen 2 will reduce frequency on AVX2 instructions.
https://www.anandtech.com/show/14525/amd-zen-2-microarchitecture-analysis-ryzen-3000-and-epyc-rome/9

 

Yes, but the difference is, frequencies are often above 4GHz and get reduced to under 4GHz. The console CPUs have generally lower frequencies. They don't have boost frequencies like their desktop counterparts so that might be the reason why it can hold it's frequency. E.g. if I remember correct even zen+ did not go under 3GHz with full AVX load. Intel CPUs on the other hand can go way below that.
And don't forget, 1 Core (2 Threads) are for the OS. So even if the game fully utilize 7 cores, there is still one core that might be very inactive.

 

If considering the extremely big case for PS5 and the liquid metal cooling patent, I will expect PS5 GPU operate more like the bottom graph, with range between 2.1~2.23 GHz.

And it is very interesting to see how multi-platform games will perform with narrow-fast GPU.

 

No disrespect but this entire MHZ talk regarding PS5 is meaningless and you guys are having a really, really hard time letting go of this concept!
For PS5, the number is debate-less.


1 - With a locked Total Board Power, the only thing the Mhz number is useful for, is to academically determine how good a piece of silicon is VS another identical piece of silicon (silicon lottery).

2- For PS5 its even more irrelevant because ALL PS5's will follow the same fixed pre-baked curve where (xTBP) = (xMHZ). In that curve you will never ever see (100% TBP) = (100% Mhz), so, never 2.23ghz when it matters unless the TBP limit allows for it! It doesn't! Cerny already said as much.

3- The kingdom Come graphic, only means the GPU is underutilized most of the time on PC. If that GPU was on PS5, a naughty dog game at the end of the life cycle would run bellow 1900mhz all the time.




The argument that it will run at 2.23ghz most of the time implies 2 things and 2 things only:

1 -The allowed TBP would allow for 2.23Ghz on true 100% GPU utilization (not going to happen. Cerny already said as much. It will drop even more as developers start squeezing every drop of performance).

2- The GPU will be under utilized = not as much work, which makes the argument pointless.



If this entire practice of keeping the MHz conversation alive is to have a baseline of comparison against XSX, you will have to look at the metric common to both that you can actually measure = performance per Watts!

 

Mark Cerny stated they had to cap the GPU frequency to 2.23Ghz to ensure that the "on chip logic operates properly" and he also said "we're able to run way over that [2.23Ghz]". So there is some form of practical speed limit because while the chip can be clocked much higher, the suggestion is that it's not reliable but there's no insight into what the specific issue is. It could be similar to the signal stability issue that Andrew Goossen mentioned when discussing Series X memory bus.

 

Take a look at oc'ed 5700xt's, they boost rather high too. 36CU, 448gb/s rdna product,

One can wonder why He and everyone even bothered with the whole thing, since its never going to happen anyway.

 

That's wrong. 2.23Ghz on true 100% GPU utilization could happen. 2.23Ghz with 100% GPU and CPU won't.
And you can lock frequency on devkit (of course CPU won't reach max performance then).

 

I have my doubts on that. Compared to the CPU the GPU is massive and is using lots of power. CPU = tens of W, GPU = hundreds of W.

 

You can’t have “100%” utilization in any practical scenario. I’ve seen devs/engineers chime in that the real number is around 30% for most games.

Cerny stated that you can reach 100% or close to it on the Tempest engine, and if you do, the amount of BW it needs is insane.

I also find it amusing that people talk the GPU curve as if it’s fixed forever. Sony can always push out firmware updates to adjust the curve.

 

I guess I was just upset that it felt like you were just trying to bury the conversation. I was hoping to get something for you that would indicate that with say finfet Pleakage is now >>>> Pswitching a reversal of how transistors used to run. A simple response may have been like The equation for Pswitching has not changed, that is still cubic but If you want to compare relative power draw you run into an issue if Pleakage is >>> than Pswitching and that i need to factor in now Pshort circuit on other items in there.

i don’t think anything you said was unreasonable, I’m not entitled to a full response, and You don’t have to give one if you don’t want to. We’re adults here and it’s not hard for me to read some books. I was just looking for direction on where to proceed and I think the topic would have been good to discuss anyway, it’s not a common topic I see coming up on forums at all.

in the end I think it would have made good discussion not bad.

 

Could you give a little more context about the "might not want" or "will not" part of your question? Road to PS5 is interesting but I sadly lack the time to keep going through it!

The reason MS talked about 3.8 being constant is because they were asked specifically about 3.8.

(Near the bottom) https://www.anandtech.com/show/1599...ft-xbox-series-x-system-architecture-600pm-pt

"09:35PM EDT - Q: Is link between CPU and GPU clocks? A: Hardware is independent.

09:36PM EDT - Q: Is the CPU 3.8 GHz clock a continual or turbo? A: Continual.

09:36PM EDT - Continual to minimize variance"

The clocks are continual. If AVX affected that, they wouldn't be. But we can demonstrate this! Earlier in the presentation (same link) MS stated that "AVX256 gives 972 GFLOP over CPU" (quoted from Dr Ian Cutress' transcription).

32 Flops/cycle * 8 cores * 3.8 gHz = 972.8 GFlops.

So we can say that:
- 972 GFlops is at 3.8 ghz
- The 3.8 gHz is continual to minimise variance. It is not a boost, and it is not affected by the GPU.

And this presentation wasn't from MS PR people, it was prepared by two members of the Azure silicon team. Legit experts. These people aren't clowns, and are every bit as professional as Cerny.

Btw, at 3.6 ghz, peak AVX2 output would be lower at 921.6 GFlops. It's one AVX256 instruction per cycle, per core. Which is not to say you couldn't squeeze in a tiny bit more work with HT enabled, but if you're hammering the AVX256 in a tight loop like a PC torture test there's not much time for anything else. And if you want to really burn in your chip and test the thermals and cooling you use some kind of AVX torture test.

There are no MS changes to Zen 2, and none are necessary to fix clocks.

What this article is describing is running to the limits of the server package based on temperature, voltage, power, whatever. You alter clocks to stay within limits.

If you have a fixed clock regimin where all operations (including AVX256) are below the power, temperature, voltage limits etc that the chip/system has then you know you will never need to downclock.

For example: the 3700X has a base clock of 3.6 (16 threads) and a boost clock of 4.4. Even with AVX256, if set up correctly, it won't drop below its base clock of 3.6. Now image if you turned boost off. You could run at 3.6 all day long, whatever you threw at it, and it'd be solid at 3.6.

That's basically what Xbox series X is doing.