Playstation 5 [PS5] [Release November 12 2020]

[snc]

github was almost right but we got higher clocks and rdna2, so we have 1.2x advantage of raw performance for xsx but 2x faster ssd on ps5 and I think memory advatnage as 560vs448 is 1.25x so almost the same as performance difference but xsx has only 10 with full speed and ps5 16, generaly close consoles with edge for xsx

 

[Rurouni]

PS5s audio is based on modified AMD CU (not specified which specific version) with caches ripped out at least. I'm pretty sure XSX can sacrifice a CU or two for audio if needed, unless it has some audio DSP solution (or to add on top of it), there should be no reason they couldn't use TrueAudio Next if they wanted which does exactly that.

Some people says that PS5 will have more TF than XsX.... maybe if we count this audio CU thing...... secret TF!

 

[Janne Kylliö]

Honestly, adding this complex boost mode / throttling for some "small" clock adjustment wouldn't make any sense...

 

[Silent_Buddha]

Though possibly true, the realistic size of MS's SOC shows high and wide isn't that expensive, so why did Sony do something different? We have to wait and see what the final console BOMs are to know if there's some good rationale for the decisions or not.

Why thank you for giving me an opportunity to elucidate on something that came to mind after going through the PS5 GDC talk and the XBSX DF article.

It appears that Microsoft and Sony started development of their consoles with 2 different overriding guidelines.

Microsoft started off with the guiding principles of...​

• 12 TF no matter the cost in power (watts).​
  • I'm sure they were prepared to potentially have to release a 300+ watt console.​
• Reduce the effect of any bottlenecks that games are seeing.​
  • Storage is the most visible result of this.​

Because of that we see a XBSX case that may be significantly over-engineered for cooling because RDNA 2 turned out to be more power efficient than they expected.

Sony appears to have started with the guiding principles of...

• X power (watt) target.
  • And focusing on how to ensure that the console can hit that target no matter what conditions (within reason) the console is subjected to.
• What are the things that they can really focus on that wouldn't significantly affect the power target?
  • Storage is a really good starting point.
  • Audio, something Sony were already interested in is another low hanging fruit.
• How do they achieve BC so that Microsoft doesn't have a competitive advantage in terms of keeping people in their ecosystem?
  • Without as much abstraction as MS has or experience with supporting BC across a changing landscape of hardware, this meant the hardware had to do the lions share of it.
  • So, do they go with 36 CUs? Or 72 CUs? 36 is probably the more practical choice.

The result is that as they got closer to RDNA 2 (and thus the SOC) becoming a physical reality, the speed of the SOC maybe have increased over time due to RDNA 2 being more efficient than expected. So, it's possible that it started with a target of 9.2 GHz due to what they expected to be their max clock prior to silicon taping out. As AMD has done testing on it, they've been able to boost the clocks to where it is now.

This may be why we still haven't been shown the retail case for the PS5. I get the feeling that because MS started off not knowing (but expecting really high) what their power consumption would be or how hot their 12 Tflop SOC would be that they designed a robust case that would deal with a worst case scenario. IE - I get the feeling the case was designed before they had an SOC.

I get the feeling that the cooling for the PS5 may be more exotic and/or complex than the XBSX, so Sony's designers may still be looking at how they can make an attractive/striking case compared to MS which went with a more industrial/living room look. Cooling a narrow 2+ GHz chip is likely not trivial. The chip itself is going to have less surface area and 2.23 GHz may be beyond the power efficiency elbow of the power curve.

All that said, it's good that the boost is wattage limited and not thermally limited. However, high thermals can affect power consumption, so the thermal solution will have to keep the chip below X temperature at all times in all environments to ensure that the chip can always hit the expected clocks at the given power target.

That said, this does bring up some things that developers will have to keep in mind with PS5 performance that they don't have to with XBSX performance.

• If you have code that pushes the CPU and GPU equally hard, neither the CPU nor GPU will hit the boost clocks during that time.
  • Most of the time this shouldn't happen however, so the variable frequency shouldn't hurt it much.
• If Sony allows the use of SMT, the GPU may be negatively impacted.
  • The purpose of SMT is to allow fuller use of the CPU by allowing more threads to run on a CPU core to take advantage of when the core would be idle.
  • Basically by increasing occupancy on a CPU core, the CPU core will be idle less often, hence the CPU core will use more of the power budget at any given time.
    • This means that the CPU will be in contention with the GPU for the power budget far more with SMT than without.
  • IMO, because of the variable frequency of the CPU and GPU, I don't think Sony will expose SMT to developers.

I'm really excited that Sony focused as much as they did on Audio, I REALLY hope that audio takes off this generation, but a lot is going to depend on just how robust Microsoft's "3D audio hardware" is. If it ends up only being 1st party developers taking advantage, that's cool, but doesn't move the gaming landscape. I want the gaming landscape to get moved WRT to not just audio positioning but audio processing.

Regards,
SB

 

[PSman1700]

Why do they need that when adequate NVMEs can be used?

Because those expension cards looked very small and easy to connect. A bog standard NVME sounds more like a last minute solution, but perhaps cheaper?

this is a 9 teraflops GPU. No GPU in the world can sustain 2.23GHz.

This is how it is going to run the GPU at that speed. A 2080Ti has a much lower frequency...

They never said it can sustain that clockrate, its a peak performance clock. MS knew this, thats why they advertised so heavily on 'we have consistent performance clocks'

No, dont think we’ll get over expensive Sony branded cards.

Guess like PS3, PS4.. you can just buy third party M.2 PCIe 4.0 NVMM cards. They only have to be ok’ed by Sony, meaning they are fast enough.

Any card comes on top of the built-in 850 GB.

Aha.... well then, XSX solution seems the more friendly then. Why the odd 825gb though? Almost 200GB (two games?) less to install. Yes the SSD is faster, but also less in size.

Honestly, adding this complex boost mode / throttling for some "small" clock adjustment wouldn't make any sense...

GPU performance between 9 and 10TF at the max, depending on workload. A hot chip, explains, to an extend, the expensive cooling solution rumor from awhile ago.
The 2Ghz mark would seem as a realistic sustained clock for heavy workloads. A 60FPS game stresses the CPU more, so concessions will most likely happen there.

 

[Deleted member 11852]

Everything about Sony's PS5 has felt rushed and not all there tbh.

I can see how you may feel that way yet Sony first starting talking about PS5 almost a year ago, i.e. the first Wired interview. They've done a bit of comms with the public (two Wired interviews) and this presentation was for devs who don't care what the box looks like. Sony have always been a bit of a tease/troll about information they know folks want.

 

[Kaotik]

As far as I can tell, going of eurogamer, this is the "boost" implementation,

https://www.amd.com/en/technologies/smartshift

No, it's not.
SmartShift is only part of the puzzle.
The Boost is just like on GPUs and CPUs and APUs on the PC side of the fence, dynamic clocks based on different factors. In case of PS5, it's powerlimit, with CPU and GPU both having their specific portion of it by default. SmartShift allows you to give some of the power reserved for CPU over to GPU when needed, if you rather want higher GPU clocks than CPU clocks or your CPU is running half idle anyway (exaggarating).
late edit:
to clarify, it's not just that powershift, but the clocks are affected by what exactly you're doing, as certain loads consume more power than others. So let's say your load is slightly too powerhungry on the GPU side, you could sacrifice little TDP from CPU over to GPU and still maintain clocks, but as the load gets high enough it won't be enough (or maybe the CPU is running full throttle too, then there's nothing to give)

 

[disco_]

this is a 9 teraflops GPU. No GPU in the world can sustain 2.23GHz.

This is how it is going to run the GPU at that speed. A 2080Ti has a much lower frequency...

https://www.sil-tronix-st.com/webuploads/display/215/Tirage2_02_03.JPG

Just about every 2080 and Ti run >2000mhz. Nvidia uses low specs to show efficiency advantage but if you game on those cards, 2000-2100 is the norm. My 2070 super used to sit at 2125 out of the box. GPU Boost is a damn miracle worker(/s?).

I can see how you may feel that way yet Sony first starting talking about PS5 almost a year ago, i.e. the first Wired interview. They've done a bit of comms with the public (two Wired interviews) and this presentation was for devs who don't care what the box looks like. Sony have always been a bit of a tease/troll about information they know folks want.

I think MS surprised them. I'm pretty sure sony knew they were going for a lower spec, cheaper machine than MS but they underestimated what MS would do. The past few months has been sony trying to make up for that surprise deficit.

 

[PSman1700]

I can see how you may feel that way yet Sony first starting talking about PS5 almost a year ago, i.e. the first Wired interview. They've done a bit of comms with the public (two Wired interviews) and this presentation was for devs who don't care what the box looks like. Sony have always been a bit of a tease/troll about information they know folks want.

I dont think they initially rushed the PS5, they just started development earlier, with perhaps BC in mind (more of a hardware solution then MS does), and 'double Pro power', then Stadia came, and probably also cought wind of XSX specs somewhere down the line. From there, things like extreme clock speeds came into play for the GPU, to be able to claim 10TF (double digit). That it can't always reach that is another thing.

Just about every 2080 and Ti run >2000mhz. Nvidia uses low specs to show efficiency advantage but if you game on those cards, 2000-2100 is the norm. My 2070 super used to sit at 2125 out of the box. GPU Boost is a damn miracle worker(/s?).

True. Also, that's on PC then, where thermal and power limits aren't as much of a problem.
The thing i could see as a problem for Sony, with such high clocks, there must be higher voltages behind them. Ideally, you'd keep voltages as low as possible to extend chip life, and keep components cool. No idea about the final design, but consoles usually aren't so big. Going to be intresting how sony has fixed that.

 

[Kaotik]

Some people says that PS5 will have more TF than XsX.... maybe if we count this audio CU thing...... secret TF!

I'm pretty sure he said Compute Unit rather than Compute Units, I'm pretty sure secret extra Compute Unit won't give you that many flops

edit:
As for people who said this was rushed or whatever, just like MS's presentation this was meant to be presented at GDC

 

[disco_]

I dont think they initially rushed the PS5, they just started development earlier, with perhaps BC in mind (more of a hardware solution then MS does), and 'double Pro power', then Stadia came, and probably also cought wind of XSX specs somewhere down the line. From there, things like extreme clock speeds came into play for the GPU, to be able to claim 10TF (double digit). That it can't always reach that is another thing.



True. Also, that's on PC then, where thermal and power limits aren't as much of a problem.
The thing i could see as a problem for Sony, with such high clocks, there must be higher voltages behind them. Ideally, you'd keep voltages as low as possible to extend chip life, and keep components cool. No idea about the final design, but consoles usually aren't so big. Going to be intresting how sony has fixed that.

Your concern trolling is dully noted.

 

[Scott_Arm]

Did Sony talk about anything that matched up to what Microsoft called Sampler Feedback Streaming, which they suggested could reduce texture streaming from disc by 1/2 to 1/3 by loading only the needed parts of a texture vs the entire texture. It's a very different approach than what Sony is offering, which is to just load everything into ram as fast as possible, unless they actually do something like Microsoft's streaming as well.

 

[anexanhume]

Sure there isnt VRS and mesh shaders. They mentiond geometry engine and primitive shaders, the Vega thing. I hope at least this time they are functional.
By the way the tempest 3d block is not clear if isnt using CU shaders, would be the tip on the cake.
Not even a bespoke thing in the GPU except for the compatibility work.

Primitive shaders are mesh shaders.

VRS is a RDNA2 feature, as confirmed by AMD’s own RDNA2 slides on financial day.

 

[PSman1700]

Your concern trolling is dully noted.

2.23Ghz is going to scare some yes. Previously 2000mhz for the ps5 was seen as impossible and fantasy.

 

[function]

It'll be less pronounced because the GPU performance is within 15-20% range and bandwidth difference is within 25%. It's not as vast as the 4Pro's deltas of 42.85% GPU and 49.54% bandwidth.

EDIT: underestimated the 4Pro GPU deficit at 25%

I think the real world difference in BW available for the GPU will be larger than the theoretical total system BW.

For the PS4 there was the famous graph showing real world BW was well below theoretical, and that CPU access to memory disproportionately hurt real world BW for the GPU, due to CPU access being prioritised.

( ~10GB/s of CPU use is costing about ~30GB/s for the GPU)

So 560 / 448 = 1.25, a 25% advantage for the XSX on paper.

But assuming your memory controller is working in the 80%-of-theoretical ballpark, and that these fast new CPUs still have prioritised access that was costing, say 80 GB (pulled outta my ass) from the GPU in a stressful scene, the BW difference could turn out to be:

(560 * 0.8 - 80) / (448 * 0.8 - 80) = 1.32.

Obviously this is just an example, but the actual BW advantage for the GPU is larger than the raw bandwidth figures indicate. Plus a wider bus is supposed to allow for more efficient scheduling and higher effective utilisation and the XSX has that too.

The X1X "only" has 50% more BW than the Pro, yet there are games where the X1X has a 70%+ performance advantage. I'm guessing that in BW limited areas of games that XSX will be 30%+ faster than the PS5.

 

[Pressure]

Games are probably not optimized for SSD loading for now and the SXS demo is just to show improvement over the base HDD. Sony's quote is probably the best case scenario where the game doesn't even have to fully load to get playing and then most stuff is streamed in. Not something possible to demo right now. Will see more when people actually tests this but I have a feeling both will perform pretty similarly even with 2x the bandwidth quoted for the PS5 considering the 20x improvement from moving from HDD to SSD isn't remotely close to loading the game in a 20th the time.

It should be quite obvious that a normal SSD controller is not capable of doing what the 12-channel custom SSD controller does in the PS5.

 

[Deleted member 11852]

I think MS surprised them. I'm pretty sure sony knew they were going for a lower spec, cheaper machine than MS but they underestimated what MS would do. The past few months has been sony trying to make up for that surprise deficit.

Huh? By saying nothing?

 

[Kaotik]

Primitive shaders are mesh shaders.

VRS is a RDNA2 feature, as confirmed by AMD’s own RDNA2 slides on financial day.

Has the primitive=mesh been confirmed anywhere? since what AMD originally called primitive shaders isn't the same as mesh, only similar.

VRS is RDNA2 feature, but they've customized the GPU, it's possible they have dropped features too (personally I expect it to support VRS though)

 

[bgroovy]

They never said it can sustain that clockrate, its a peak performance clock.

Cerny said it hits that clock the vast majority of the time. Or are we back to saying he's lying about everything?

Did Sony talk about anything that matched up to what Microsoft called Sampler Feedback Streaming, which they suggested could reduce texture streaming from disc by 1/2 to 1/3 by loading only the needed parts of a texture vs the entire texture. It's a very different approach than what Sony is offering, which is to just load everything into ram as fast as possible, unless they actually do something like Microsoft's streaming as well.

He mentioned a memory mapping engine in the I/O section that sounded like it does a lot of this same stuff.

 

[dobwal]

The 2.23GHz GPU clockspeed is all well and good, but the variability makes me a little nervous. It's probably the best generation ever in which to take this approach, given the abundance of dynamic resolution scaling that we've seen take off in the current generation, but I'm still unsure of what this will mean in practice.

We'll see, but I don't want to get one of the shit ones, and be stuck with a console that only reliably hits 2.1GHz at the same rate that an ideal console hits this touted 2.23GHz. There are plenty of unknowns, and so I don't want to jump to any conclusions, but my gut says that I would've preferred 40CU's at a reliable 2GHz.

I don't believe Sony would allow for such a circumstance. AMD current RDNA 1 cards come with base, game and boost ratings. The base or lowest clocks is the lowest frequency the chip will run at under extreme duress like a power virus. The game clock is what most games will typically run at under most circumstances. Boost clocks are just max frequency.

The PS5 will probably boost to 2.23 GHz under circumstances where the gpu isn't being pushed too hard. So even if you end up with PS5 which can't normally boost that high (I doubt Sony would pass such a chip during QC), your PS5 will still have more than enough performance to manage the circumstance.

Devs are going to profile their games during the times of peak duress when the PS5 is least likely to provided that level of frequency. They will simply manage they like they always have and tone down effects or other aspects of performance to keep framerates at acceptable levels.