Support NeoGAF

[sonomamashine]

Not to , PC to Sony.

Bruuuuuuuh

It achieves a 20,000 firestrike which is similar to that which was quoted for Gonzalo - so - in short - it means that next gen might be roughly equivalent to a Ryzen 3700 + Radeon 5700xt combo - not much better, nor not much worse

Wich i think its a good thing compared to what we had before, slightly better than the xt or slightly worst.

 

[alabtrosMyster]

Bruuuuuuuh

I am funny, ain't I??

 

[sonomamashine]

I am funny, ain't I??

I love it, i can read it all day long keep it up.

 

[TLZ]

Thanks for all your hard work DemonCleaner

 

[SonGoku]

slightly better than the xt or slightly worst.

slightly worst

 

[sonomamashine]

slightly worst

Im talking about rx 5700xt not gonzalo
Edit:
Ryzen 3700 + Radeon 5700xt combo - not much better, nor not much worse

 

[SonGoku]

Im talking about rx 5700xt not gonzalo
Edit:
Ryzen 3700 + Radeon 5700xt combo - not much better, nor not much worse

Gonzalo (assuming it exists) has the GPU equivalent power wise of an underclocked 5700xt. So slightly worse than a stock 5700xt

 

[sonomamashine]

Gonzalo (assuming it exists) has the GPU equivalent power wise of an underclocked 5700xt. So slightly worse than a stock 5700xt

I know gonzalo is underclocked xt, I meant ps5 would be Ryzen 3700 + Radeon 5700xt even, slightly worst or slightly better.

 

[Arkage]

A 20k score is great for a 400-500 price range.

 

[DemonCleaner]

First things first:

• I'm not saying Gonzalo is PS5, it might be!
• I'm not suggesting this simulated Gonzalo is equivalent to PS5's power in games, this thread is about how much computational power you can put in a console size box and not how efficient you can use that power.

going by some comments, i felt like this clarification was needed in the OP

 

[DemonCleaner]

60 CU chip with a 320bit bus would be ~380mm2 on 7nm DUV, with 2DCUs disabled leaves 56CU @1600 = 11.4TF
or you meant if 1.8GHZ is even possible on a console on DUV?
btw can you try undervolting and clock to 1680mhz like the one posted on ree

sure, from an economical perspective im with you. it's doable. the question to answer is, if it's even possible considering the power requirements. going lower in clocks won't scale indefinitely.

btw can you try undervolting and clock to 1680mhz like the one posted on ree

yeah. i'll be doing some kind of frequency / wattage chart. that will take some time.

->

yeah, i will do something like that definitely (it's basically the whole goal of this exercise to get a feeling for the power/perf sweetspot of RDNA). it's a lot of work and im still figuring out what's the best way to do it (probably will set step up the power limit in incements. it won't work to step up the frequency as soon as you run into a power limited scenario. what makes it even harder is that you have to find a stable voltage to complement the coresponding clock rate. too many variables...). hope i get there on the weekend but no promises.

 

[SonGoku]

if it's even possible considering the power requirements. going lower in clocks won't scale indefinitely.

5700XT consumes very little undervolted at 1680Mhz (ree post)
A 56CU GPU in the 1550-1600Mhz range should hit the sweet spot in perf/watt with added benefit of better yields and fine tuning
A 200W SoC should be manageble in a console form factor with decent cooling like vapor chamber

A 20k score is great for a 400-500 price range.

If they where 1080p/1440p consoles it would be great
For 4k gaming is terrible

 

[DelireMan7]

Me entering this thread thinking "Stimulating Gonzalo" is a game for PS5 :

 

[Slings and Arrows]

This is some jeffrigby level shit

I love it.

 

[henau212]

A 20k score is great for a 400-500 price range.

You have to remember: this is an engineering sample. They might run the chip with all CUs enabled for testing purposes. Also, as others said, 1800 clock is probably too much for a console. It might be lowered to 1650 in the final console.

That said:
1) it is not clear if Gonzalo ends up in the PS5
2) it is also not clear if the engineering sample is the final combination of enabled CUs and clock. Both might be lower to reach the yields and TDP

 

[TeamGhobad]

you're doing a helluva job there brownie

 

[DemonCleaner]

So ok, that power thingy tends to confuse people. Let's give it a go.


So here again is the power measurement from the wall in Fire Strike.

Ok, 208W for Gonzalo, so what do we actually make of that?

We have a second measurement, the GPU die power over time which gets a mean of 125W in said test. I also know the power efficiency of my PSU on 1/3 load. It's 90%. And I know the difference between die power and TBP of the 5700XT.

From that framework we can derive quite a bit:

*

So black are measured values. Yellow are estimates (Whew! only one). And the rest [red] can be calculated or derived. Granted that's no excact science, but i bet it wont be to far off what you would measure if you had the means.


So if we mutate the actual computing components into one hypothetical APU we see that we get around 150W of TDP. Thermal Design Power. The stuff that when it eventually turns to heat you have to cool away. So keep in mind that there are some redundant components in GPU and CPU like memory controllers that an APU would have less of. Those need power the APU wouldn't need. Also moving data between RAM CPU VRAM and GPU also consumes lots of juice, that wouldn't be needed if big parts of the data would just hang around in on-die cache.

Long story short, an acutal console APU would probably be even more efficient for said reasons. If i had to pull a figure from where the sun doesn't shine, i would say like another 20W.

So this combination of ingredigents would fit nicely in a power budget last gens consoles were offering. Launch PS4 was around 140W i think.



*edit: i repeated the testing for the 1.8Ghz data point like 4 times and I looks like i had a little fluke in the first result. Furthermore the estimate for MB&Aux seemed a little low thinking over it. I uped that to 15W in the new graph.

 

[alabtrosMyster]

A 20k score is great for a 400-500 price range.

No 20k SUCKS big time, you can't play any game with that kind of specs, especially if you don't have an intel 9900k that gives you that sweet 6% benefit over Zen2CPUs!

Think about it, that will be in 2020, by then the average PC will score at least 45k in that same test.

 

[DemonCleaner]

So while strating the power scaling tests yesterday i realized something strange. I underclocked the 5700XT to 1700Mhz and was expecting that it would hit that clock most of the time, beacause it wouldn't run into a powerlimit anymore compared to the previously shown 1800Mhz underclock. As you can see in the results, i was kinda wrong about that:

It just swings around 1650Mhz instead of the target frequency of 1700Mhz.

Compared to the previous 1800Mhz target testing you can clearly see that it much more stable now:

Confirming that it doesn't run into it's 125W power limit or some other limitation anymore. You can also see that in the die power diagramm:

First i thought that might have something to do with the lower voltages. But after testing lower und higher max frequencies (with out doing those graphs mind you), to me it pretty much looks like the target value you put into wattman is 50 Mhz higher than the real frequency target the card is trying to archieve. No idea what's the purpose of this offset. I think this means, i must repeat the Gonzalo leak spec testing with 1850Mhz in Wattman. But i don't expect too much of a change, because of it reaching the power limit there quite often.

 

[DemonCleaner]

lol, you can undervolt the 3700X even further than 1000mV at gonzalo settings. Btw. I haven't used Ryzen Master in years, but it's really awesome nowadays. AMD software stack's getting better and better.

 

[xool]

It just swings around 1650Mhz instead of the target frequency of 1700Mhz.

Do clocks run go up in 50MHz steps ? .. probably the difference between "<" (less than) and "<=" (less than or equal to) somwhere in the code

 

[DemonCleaner]

Do clocks run go up in 50MHz steps ? .. probably the difference between "<" (less than) and "<=" (less than or equal to) somwhere in the code

Good idea, but it's much more fine grained. 1 or 2Mhz steps. You can actually see that on the frequency charts.

 

[DemonCleaner]

Fucking hell this was a bunch of work....


So first results for the power scaling (respectively frequency/voltage scaling) profile of the 5700XT. I tried to undervolt every datapoint, but i didn't go through the hassle of finding the lowest stable undervolt for each one.

So here's system power drawn from the wall over target frequency in Wattman in relation to graphics score scaling in Fire Strike (All is done with the 3700X @3200Mhz & 1000mV):

As you can see, power draw looks exponential and really starts firering up above 1.8 Ghz target frequency. All that while the Fire Strike score raises not even quite linearly. The yellow dot is the Gonzalo simulation.


Another way to show this, is to compare scaling factors of Fire Strike and Wallpower with the Gonzalo simulation as a base (100%):

Like i said, there is still some room left for further undervolting, but the result shouldn't look to different from that.

 

[sonomamashine]

Good work .

 

[xool]

[edit = Good work .. is good to see real numbers]

Any idea what is going on with memory clocks at the same time - I read that 5700XT uses 14Gb/s GDDR6 which means the WCK clock is 7GHz, and it takes 4 of these WCK clocks for seamless memory transfers ie 1.75 GHz (see https://www.micron.com/-/media/client/global/documents/products/technical-note/dram/tned03_gddr6.pdf page 5 fig.5)

So above 1.75GHz GPU optimal clock matching with the memory will be lost .. maybe this is why the power/performance curve takes off at just over 1.75 (1.8GHz)

[disclaimer - I'm not saying this is right - just putting it out there] .. if my guess is right with 1.6Gb/s GDDR6 that threshold coud be raised to ~2.0GHz GPU ..

 

[_sqn_]

1800mhz sweet spot, same clock has gonzalo 9.2tf ?

 

[R600]

So Navi does not actually profit of really low clocks (1200-1400MHZ) so this makes predictions or 64CUs at 1500-1600MHZ rather unbelivable

 

[vpance]

So Navi does not actually profit of really low clocks (1200-1400MHZ) so this makes predictions or 64CUs at 1500-1600MHZ rather unbelivable

I don't think the tests are a clear indicator of that.

To determine that we need to compare cards with more or less CUs (or disable in bios?) and do further undervolting for lower clocks.

 

[Shin]

Should also have posted that famous programmer tweet saying consoles have usually twice the performance as an identical PC setup (just to stop the people from complaining that this setup will be weaker than a PC with the same specs).

I'm not sure who it was though.

I got you covered fam.

 

[Futaleufu]

Too optimistic. They are not going to put a 200W PSU in a next gen launch console. Launch consoles used 112W (Xbox One) and 134W (PS4) according to Google, while XB1X uses 245W and PS4 Pro uses 165W. They are going to save the 200W+ power increase for the mid-gen refresh.

 

[sonomamashine]

Too optimistic. They are not going to put a 200W PSU in a next gen launch console. Launch consoles used 112W (Xbox One) and 134W (PS4) according to Google, while XB1X uses 245W and PS4 Pro uses 165W. They are going to save the 200W+ power increase for the mid-gen refresh.

lol what ?????

 

[jonnyp]

Too optimistic. They are not going to put a 200W PSU in a next gen launch console. Launch consoles used 112W (Xbox One) and 134W (PS4) according to Google, while XB1X uses 245W and PS4 Pro uses 165W. They are going to save the 200W+ power increase for the mid-gen refresh.

If they can put components inside the box that require 200+ Watts and are able to launch the console in the 4-500 dollar range, they will do that.

 

[_sqn_]

Fucking hell this was a bunch of work....


So first results for the power scaling (respectively frequency/voltage scaling) profile of the 5700XT. I tried to undervolt every datapoint, but i didn't go through the hassle of finding the lowest stable undervolt for each one.

So here's system power drawn from the wall over target frequency in Wattman in relation to graphics score scaling in Fire Strike (All is done with the 3700X @3200Mhz & 1000mV):

As you can see, power draw looks exponential and really starts firering up above 1.8 Ghz target frequency. All that while the Fire Strike score raises not even quite linearly. The yellow dot is the Gonzalo simulation.


Another way to show this, is to compare scaling factors of Fire Strike and Wallpower with the Gonzalo simulation as a base (100%):

Like i said, there is still some room left for further undervolting, but the result shouldn't look to different from that.

Great work, do you have firestrike overallscores result? We would know minim tf number that gives above 20k result with zen2 3.2ghz.

 

[sonomamashine]

Great work, do you have firestrike overallscores result? We would know minim tf number that gives above 20k result with zen2 3.2ghz.

8Tflopish.....

 

[_sqn_]

8Tflopish.....

I thin it would be below 20k

 

[R600]

Someone posted 5700 with 3600x boosted to 4.2GHZ scoring 20k.

I wonder if additional RT hardware will bring some additional uplift in FS score, therefore making it not 1:1 case with PC. Who knows, we'll see!

 

[DemonCleaner]

Great work, do you have firestrike overallscores result? We would know minim tf number that gives above 20k result with zen2 3.2ghz.

i couldn't do the FS overall for every test yet, because i need afterburner running to gather the GPU die power data. the problem is, when you have afterburner running Fire Strike crashes at the physics test EVERY TIME. that's also the reason why i only show the over time graphs for graphics tests 1 and 2.


i did an overall run at target frequency 1750Mhz yesterday, thought: it gives just aroound 20K:

1750Mhz target is a real mean frequency of 1700Mhz with very little variance.

 

[_sqn_]

i couldn't do the FS overall for every test yet, because i need afterburner running to gather the GPU die power data. the problem is, when you have afterburner running Fire Strike crashes at the physics test EVERY TIME. that's also the reason why i only show the over time graphs for graphics tests 1 and 2.


i did an overall run at target frequency 1750Mhz yesterday, thought: it gives just aroound 20K:

1750Mhz target is a real mean frequency of 1700Mhz with very little variance.

thx very much, so if gonzalo is 1800mhz and over 20k it's probably 40cu as 36cu is 8.3tf and score below 20k

 

[sncvsrtoip]

Yesterday there were benchmark leaks of new amd apu flute that looks similar to gonzalo and it turns out cpu is slower than 3.2ghz zen2:
https://www.userbenchmark.com/UserRun/18618484#PROCESSOR amd flute
https://www.userbenchmark.com/UserRun/18619396 zen2 3.2ghz
https://www.userbenchmark.com/UserRun/18598362 2700x 3.2ghz

 

[xool]

Yesterday there were benchmark leaks of new amd apu flute that looks similar to gonzalo and it turns out cpu is slower than 3.2ghz zen2:
https://www.userbenchmark.com/UserRun/18618484#PROCESSOR amd flute
https://www.userbenchmark.com/UserRun/18619396 zen2 3.2ghz
https://www.userbenchmark.com/UserRun/18598362 2700x 3.2ghz

Absolutely.

already having this discussion at https://www.neogaf.com/threads/next...lation-analysis-leaks-thread.1480978/page-173

don't want to discuss this twice

 

[sncvsrtoip]

Absolutely.

already having this discussion at https://www.neogaf.com/threads/next...lation-analysis-leaks-thread.1480978/page-173

don't want to discuss this twice

If gonzalo has similar cpu then taking as cpu 3700x3.2ghz is not proper

 

[WhiskeyDiver]

Why is it everyone else's firestrike works just fine for them...and gives me shit like this?

Spoiler

(Image spoiler tagged for convenience as it's size is a bit of an eyesore. )

Every other benchmark on 3dmark works just fine. Firestrike itself? Noooo siree bob. That fucking benchmark hates my computer.

It's a 2070 with a +100/750 oc and a 9600k at 4.8ghz if anyone's interested in what I tested it with. I can go a bit higher on the gpu, but that seems to be the sweet spot for guaranteed stability.

I'll be honest. I'm just happy that my graphics score on firestrike is at least similar to a 5700xt in some cases. This will no doubt probably change when the 5700xt partner boards come out with much beefier cooling allowing for higher and more aggressive overclocks.

Edit: I wonder...if I can get a firestrike score that isn't zero, if an overclock on my ram, currently at 2666mhz, to 3000 or 3200mhz with decent timings would add a nice lil bit to my score. Any marginal gain would bring me a bit o happiness. I just like to see benchmark improvements whenever I tweak my oc settings and manage something better. I know I could do a bit better with a 5ghz oc. Again, that would be pointless unless Fire Strike would cooperate.

 

[DemonCleaner]

Why is it everyone else's firestrike works just fine for them...and gives me shit like this?

for some reason it quits the combined test. most of the time that overclock instability. try running your setup on stock and see if it posts a overall score. if not,reinstall 3d mark.

 

[WhiskeyDiver]

for some reason it quits the combined test. most of the time that overclock instability. try running your setup on stock and see if it posts a overall score. if not,reinstall 3d mark.

I'll give it a go. It probably needs to be fed a bit more vcore voltage. I might have set it a bit too low last time I tweaked my settings.

 

[DESTROYA]

Your work is getting some attention

PlayStation 5 'Gonzalo' PC hardware simulation throws up interesting performance and power consumption results, but it's not really an apples-to-apples comparison

A PC hardware simulation of the upcoming Sony PlayStation 5's purported hardware specs offers some insights into the likely performance and power draws that can be expected from the console. The OP undervolted an AMD Ryzen 7 3700X and the Radeon RX 5700 XT to simulate the PlayStation 5 specs as...

www.notebookcheck.net

 

[DemonCleaner]

Your work is getting some attention

PlayStation 5 'Gonzalo' PC hardware simulation throws up interesting performance and power consumption results, but it's not really an apples-to-apples comparison

A PC hardware simulation of the upcoming Sony PlayStation 5's purported hardware specs offers some insights into the likely performance and power draws that can be expected from the console. The OP undervolted an AMD Ryzen 7 3700X and the Radeon RX 5700 XT to simulate the PlayStation 5 specs as...

www.notebookcheck.net

yay, i'm famous now thanks for sharing.


i would have loved if they added the updates, especially the power over frequency curves from post 73.

and i think this part is a litted bit misquoted:

"The OP notes that the clocks had to be lowered below 1750 MHz to achieve stability and notes that Wattman has a 50 MHz offset i.e. 1750 MHz in Wattman actually translates to 1700 MHz in real life. "

the first part sounds like the GPU would crash at a target of 1800Mhz. it's just that at a target frequency of 1700Mhz the frequency over time chart is pretty much nailed to a narrow frequency band. at 1800 you get much more variation. but the gpu can still be run stable at the pruposed underclock.
on the second part: im pretty sure that's not a glitch in wattmann, but that you get that offset of 50Mhz in every other OC tool (like afterburner) also.

 

[StreetsofBeige]

So what kind of TF comes from 20,000 Firestrike?

 

[Dr.D00p]

So what kind of TF comes from 20,000 Firestrike?

Depends whether your working in 'old teraflops' or shiny 'new teraflops'....the official exchange rate hasn't yet been determined by the marketing departments in Redmond & Tokyo.

 

[Racer!]

So what kind of TF comes from 20,000 Firestrike?

About 8TF

 

[Kenpachii]

Some god tier topic you made here kudo's for that.

I got you covered fam.

So what kind of TF comes from 20,000 Firestrike?

It's basically a stock 1080 gtx. LIke people expected with new stuff on it.

Honestly 20k score in total aint bad.