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

Status
Not open for further replies.
Sony's approach is too extreme and whatever exotic cooling solution they employ, it will impact yields and acoustic, but as a concept it's sound and something that I propose for some year now.

Hey fehu, what happened to this thing?

https://forum.beyond3d.com/posts/2084023/

One of the two console will be composed by:
- single soc: on basic 7nm process at TSMC, not 7nm+ or 6nm
- cpu: 8 Zen2 core @2.2GH with SMT enabled, they will call them "custom", but it will just normal plumbing work to adapt the cache to the soc
- gpu: based on navi, hybrid between rdna and rdna2, @ about 7TF, frequency not finalized, hardware RT (no info if proprietary or amd's technology)
- memory: 24GB GDDR6 on a 384bit bus, bw not finalized, but expected at around 800GB/s
- ssd: 2TB soldered
- usb: 3.2 Gen. 2, type C, 1.5A
- size: comparable to actual xbox one x
- new very low power mode with less than 2 seconds from power button push to dashboard
- controller: current gen's evolutionary iteration, rechargeable with off the shelf cables
 
Going to provide some background info on clocking in general; will need some help clarifying or correcting information:
https://en.wikichip.org/wiki/intel/xeon_gold/5120


As you can see from their Xeon chipset from Intel The Y axis show the instruction levels as well as the base clock. The numbers listed are the turbo clocks. You can see as more cores are used for processing the boost continually drops. All of this is due to power consumption and cooling.

There are 3 levels for Intel Chips. L0 which is normal workloads, full boost. L1 which is heavy AVX2 workloads and some light AVX512. And L2 which is sustained heavy AVX512 instructions.

You can see the power profiles listed. Nothing drops below L2 setting in this case, is 1600 Mhz or 1900 Mhz depending on how many cores you have.
This showcases how dramatic more cores and workloads can significantly eat into power budgets. You can see even under standard workloads (non AVX) 8 Core utilization will drop the megahertz down from 3200 to 2900.

on a Core i9-9900
we see similar behaviour (https://en.wikichip.org/wiki/intel/core_i9/i9-9900) for standard workloads
1 Core: 5,000 MHz (5 GHz, 5,000,000 kHz) +
8 Cores: 4,600 MHz (4.6 GHz, 4,600,000 kHz) +

So I want to make an argument towards why I think fixed clocks are important here that they can be sustained even under load. If XSX is able to keep its frequency despite all workloads and all cores being held, that is a pretty large feat and provides a very stable baseline for developers to program against.
 
Last edited:
Going to provide some background info on clocking in general; will need some help clarifying or correcting information:
https://en.wikichip.org/wiki/intel/xeon_gold/5120


As you can see from their Xeon chipset from Intel The Y axis show the instruction levels as well as the base clock. The numbers listed are the turbo clocks. You can see as more cores are used for processing the boost continually drops. All of this is due to power consumption and cooling.

There are 3 levels for Intel Chips. L0 which is normal workloads, full boost. L1 which is heavy AVX2 workloads and some light AVX512. And L2 which is sustained heavy AVX512 instructions.

You can see the power profiles listed. Nothing drops below L2 setting in this case, is 1600 Mhz or 1900 Mhz depending on how many cores you have.
This showcases how dramatic more cores and workloads can significantly eat into power budgets. You can see even under standard workloads (non AVX) 8 Core utilization will drop the megahertz down from 3200 to 2900.

on a Core i9-9900
we see similar behaviour (https://en.wikichip.org/wiki/intel/core_i9/i9-9900) for standard workloads
1 Core: 5,000 MHz (5 GHz, 5,000,000 kHz) +
8 Cores: 4,600 MHz (4.6 GHz, 4,600,000 kHz) +

So I want to make an argument towards why I think fixed clocks are important here that they can be sustained even under load. If XSX is able to keep its frequency despite all workloads and all cores being held, that is a pretty large feat and provides a very stable baseline for developers to program against.

Without taking into account TDP relations to turbo boost and Intel's PL levels and duration windows, the above generally conveys it well although mainstream and HEDT are generally a better comparison. Keep in mind that mainstream doesn't have AVX512 and only in cascade lake X, they added AVX512 in the HEDT only so they can claim some advantage over AMD threadripper.

Overall, the point you're making is very sound. Having fixed clocks is a very real advantage and anyone who tinkers with hardware knows that. You ideally want to hold max frequency across the board in any condition for maximum performance.

"optimistic boost" which is what you'll see in the PS5 since it's an AMD chip and their boost properties are well known and documented. This means that you'll see max frequency under light loads. As the load picks up, the frequency will drop in steps for both cpu and gpu. The good thing is that AMD's boost algorithm is very good, meaning it can take very granular steps in boost and do it quickly (Intel relatively sucks at this but nvidia is similar to AMD in this regard). BUT when using boost, you will drop frequency, of that there is no doubt. For me, the most curious nature of the PS5 are around what base clock profiles will ultimately be.
 
I mean, Cerny can't have fucked up this bad.
Pretty sure he's the guy the guy who not only OK's Way of the Warrior, he also signed ND to a multi-game contract based on WotW. Having played Way of the Warrior...
It's a good thing Naughty Dog turned out alright in the end.

PS4 they made poor thermal estimates or devs got so good at beating the apu that cooling sometimes went into overdrive.
As a guy who's owned both a launch model PS4 and launch model Pro, I would assume they made poor estimates. The fan spins up pretty loud in both of those just sitting on the home screens.

Can we please agree as a community to put the discussion about Downclocks and Stability and Overheating on hold until Sony at least show's the retail case and cooling?
Nobody is going to read a statement like this unless it's accompanied by a .gif
 
Pretty sure he's the guy the guy who not only OK's Way of the Warrior, he also signed ND to a multi-game contract based on WotW. Having played Way of the Warrior...
It's a good thing Naughty Dog turned out alright in the end.

God damnit, I knew this part of hustory, but you colored it it much better than I did hahahha.
 
You do realise it’s harder to spot potential talent vs proven talent?

Nothing anyone does will be perfect or work every time, however he’s not stupid and he’s not going to make a really poor error- especially on hardware which is the foundation of your work.
 
You do realise it’s harder to spot potential talent vs proven talent?
That's an interesting point. In regards to Way of the Warrior, what were the limiting conditions of its creation? Did it have a stupidly low budget? Or development timeline? The game is a bit of a turkey and a very poor man's Mortal Kombat looking at videos, but if it was made for $100 in one weekend, it'd show amazing talent and you'd think, "imagine what those guys could do with some real support."
 
That's an interesting point. In regards to Way of the Warrior, what were the limiting conditions of its creation? Did it have a stupidly low budget? Or development timeline? The game is a bit of a turkey and a very poor man's Mortal Kombat looking at videos, but if it was made for $100 in one weekend, it'd show amazing talent and you'd think, "imagine what those guys could do with some real support."

A similar example might be that the last 2 jobs I got were not what I applied for, the companies both looked at my CV and said ‘you’re better than this, you can do this job for us’ and offered me a different role with a higher wage...I didn’t even know I was capable of those roles!
 
Do you really think the guy who made Knack is going to concede a 30% power advantage to Microsoft?

I think the guy who made Knack, the original arcade Marble Madness, the guy has worked with/for SEGA, Crystal Dynamics and Universal, the guy who recruited Naughty Dog and Insomniac long before they were on Sony's radar, the guy who has since worked on some of PlayStation's most notable titles (Spyro, Crash, Jak & Daxter, Ratchet & Clank, Uncharted, Resistance, God of War), the guy who designed the Vita, the PS4 and PS4 Pro, I think that guy is less concerned with stratosphere performance than he is about making a platform that is easy to develop for by removing technical barriers.

In his 'the road to PS4' presentation (GameLabs, June 2013) he said that before he began designing PS4 he visited a lot of developers and asked what they wanted. They didn't want powerful architecture that was crazy or even raytracing (at that time), they wanted a straightforward architecture that was easy to extract performance from he delivered that with PS4 with that goal. In his 'the road to PS5' talk presentation developers asked for SSD so that was the focus of PS5.
 
Last edited by a moderator:
I have been looking at 5700XT peak power draw vs average power draw during gaming and I noticed that peak power draw was only ~10W higher then average.

Maybe my math is a bit of, but I cannot see how they arrived at base 2.23GHz with variable frequency, up from not being able to lock it at 2.0GHz in old way of doing things.

If I had to guess, I think PS5 will vary in frequency quite significantly when everything is running at full pelt. Something like :

lvl0 - 2.23GHz
lvl1 - 2-2.2GHz
lvl2 - 1.9-2.0GHz

How much of its time will PS5 run at max frequency is to be seen, but Mark Cerny was careful with his words. He said they expect it to run at or close to that frequency most of its time.

How close is "close" and how much is "most of its time" remains to be seen, but with Cerny saying they have expectations, rather then guarantees, we may only know after devs spill the beans.

On top of this, I also wonder does pushing clocks so high and making it variable make sense given bandwidth allocation PS5 has? I have to assume 12% upclock on same bandwidth will not yield linear upshift in performances, while it will push TDP curve way up there for limited performance gain.

Depending on how variable clocks work, I can see developers locking CPU at 3.5GHz and leaving GPU be, instead of micro managing which ops bring higher power draws and result in frequency drops. It might be the case of it being much easier to scale a setting or two down a notch then go through managing entire thing for few % higher performance.
 
optimistic boost" which is what you'll see in the PS5 since it's an AMD chip and their boost properties are well known and documented. This means that you'll see max frequency under light loads. As the load picks up, the frequency will drop in steps for both cpu and gpu.

This contradicts what Cerny said when he stated the belief that both would spend most of their time at max clocks. We simply don’t know, and I don’t know why we’re acting like we can make definitive statements without real-world examples.
 
developers asked for SSD so that was the focus of PS5.

Can't imagine devs didnt beg for a real CPU this time around. SSD was kind of a given, imagine if they where tied to mechanical drives still, for another 7 years. Yes, there were speculations prior 2019 that the next generation would be on HDD again.
 
"he stated the belief that both would spend most of their time at max clocks..
That's the problem, what is most of the time? I know the majority of time spent on my console is probably watching Hulu/Netflix :LOL:. There's also a ton of very popular games right now that aren't exactly taxing for consoles to run. This is a pretty safe statement and it leaves a lot of room for interpretation in my opinion. Based on this statement, If you want to believe PS5 clocks run near max all the time you can. Based on this same statement you can also decide to believe "the majority of time" means they can't in the most demanding games. That's why we need more information.
 
Last edited:
I have been looking at 5700XT

The 5700XT is RDNA 1.0.

Cerny said clocks are deterministic, so if a developer wants the full 10.28 TF all the time at 2.23Ghz, it will be achieved. He specifically said he wants the same performance across all consoles sold.
 
I have been looking at 5700XT peak power draw vs average power draw during gaming and I noticed that peak power draw was only ~10W higher then average.

Maybe my math is a bit of, but I cannot see how they arrived at base 2.23GHz with variable frequency, up from not being able to lock it at 2.0GHz in old way of doing things.
The 5700XT and the entire RDNA 1 architecture is irrelevant now, you can't use them as the basis of estimating power consumption at a given clocks or CU count, I have been saying that a GPU with 52 CUs @ 1825MHz would consume some where around 280W to 300W with RDNA 1, which remains true, the problem is consoles use RDNA 2, which is more power efficient than RDNA 1 (any where from 30% to 50%). Now we have no real ground to stand on in estimating power consumption of console GPUs, RDNA 2 remains uncharted territory now.
 
Status
Not open for further replies.
Back
Top