Support NeoGAF

[SlimySnake]

He increased a 1725MHz GPU to 2150MHz

Isnt that what i said though? I really dont understand this. Are you reading my posts? I literally did the math and pointed out how the 7.9 tflops stock 5700 was turned into a 9.7 tflops GPU and offered better performance than a 10.1 tflops 5700 xt. What in the world are you even saying here?

 

[SlimySnake]

Just had to click on ignored messages to see who u were replying to and I wasn't shocked it was him .

He is on a crusade to put out fake info on ps5 while having zero understanding of most simple things tech related. Its best to put him on ignore and let him lie to himself about ps5 .he might feel better playing x1 games on his new machine

Im gonna give him one more chance but yeah, i was definitely thinking about putting him on ignore after he ignored all the benchmarks i posted and continue to bring up power consumption. As if hardcore gamers give a shit about power consumption now. Just tell that to any PC gamer going out there to buy 750w to 1000w power supplies that their GPU consumes a lot of power if you overclock it and see if they laugh at your face. as long as you are getting performance from your system, and its quiet and doesnt overheat who gives a shit.

They are doing the same shit they did back when the rumors of the PS5 overheating came out. Utter nonsense and they keep bringing it back up after claiming 18% overclocks offer 5% more power even when presented actual benchmarks that prove the complete opposite.

 

[TheGreatWhiteShark]

you are just making shit up now. 18% overlock for 5% gain is just made up shit.

We have realtime benchmarks that show that overclocks do not offer a small fraction. here you see a 19% boost in clocks offer a performance boost of 13% in actual benchmarks. Like I said earlier, it's not 1:1, but its not a small fraction here.

In the firestrike benchmarks i posted above, the higher clocked 5700 outperformed the lower clocked and higher tflops 5700xt anniversary edition.

and stop with the heat and noise nonsense. you dont know if the ps5 is gonna overheat. you dont know how noisy its going to be. it has fuck all to do with performance. Absolute nonsense.

I am talking about real performance, like in games.
you want some overclock examples for 5700 XT?
from 1800 to 2050 overclocked , you get 5% on battlefield, 6% on RE2 , 4% on AC:Odyssey, 6% on fortnite, 3% on forza horizon, 6% on tomb raider, out of more than a dozen games tested, only one managed 10%, the division2.
and that was even at 1440p, not even 4k.
result: for a ~7% performance increase on a 18% overclock, you get a ...~40% consumption increase, with all assorted things this brings.
oh, and that was on liquid cooling too.

so, I'm not "making up shit", and also, I've had enough conversation with you.
thank you

 

[Dnice1]

Lol dude at least look up formula given by amd and ms in hot chip

4× cu × gpu clock in mhz

Ps5 321 giga rays
Xsx 380 giga rays

15% difference

Yeah, I could see it being more than just a 15 percent difference. You have to remember when doing these calculation for XSX you are dealing with fixed cpu/gpu clocks. I find any formula involving PS5 and the variable CPU/GPU clock to be very, lets just say theoretical. If ray tracing is already gpu intensive then the gpu won't be running @ 2.23Ghz. As Mark Cerny said, "Not all games will run at 3.5 and 2.23Ghz. When that worst case game arrives it will run at a lower clock speed." I would think AAA game with ray tracing would be one of those worst case scenarios.

 

[SlimySnake]

I am talking about real performance, like in games.
you want some overclock examples for 5700 XT?
from 1800 to 2050 overclocked , you get 5% on battlefield, 6% on RE2 , 4% on AC:Odyssey, 6% on fortnite, 3% on fora horizon, 6% on tomb raider, out of more than a dozen games tested, only one managed 10%, the division2.
and that was even at 1440p, not even 4k.
result: for a ~7% performance increase you get a ...~40% consumption increase, with all assorted things this brings.

so, I'm not "making up shit", and also, I've had enough conversation with you.
thank you

where are these benchmarks? All the stock 5700xt benchmarks ive seen run at 1.9 ghz average. You have to look at the average in game clocks by monitoring it to see where it actually sits. Using AMD's advertised game clocks isnt good enough because they could be 1.4 ghz or 2.0 ghz, you have to take an average.



So going from 1.9 ghz to 2.05 ghz is only a 7% increase. So if you are getting 5-6% from a 7% increase, great.

 

[TheGreatWhiteShark]

snip

here they are, you earned a spot in my ignore list

Overclocking the Radeon RX 5700 XT: 2.1 GHz on Liquid Cooling

After testing the AMD's new Radeon 5700 GPUs and Nvidia's RTX Super answer, we are particularly happy about the value offered by the latest Radeons. The $400...

www.techspot.com

 

[Gaz]

Remember the Xbox One and Mister X secret sauce stuff? That was hilarious. Now its PlayStation's turn and I'm like

 

[TheGreatWhiteShark]

Remember the Xbox One and Mister X secret sauce stuff? That was hilarious. Now its PlayStation's turn and I'm like

you know, this is what I am trying to tell some people, to avoid being in that ridiculous position, and I'm saying it in a very gentle and good will way, only to get bad responses and swears in return.
maybe its better to just get my popcorn and laugh like you do, too.

 

[MasterCornholio]

Remember the Xbox One and Mister X secret sauce stuff? That was hilarious. Now its PlayStation's turn and I'm like

Honestly the Mr X stuff is all about making FUD to make the PS5 appear like a last gen system. I haven't seen any games to additional GPUs in the power supply that makes the PS5 more powerful than the XSX.

 

[raul3d]

here they are, you earned a spot in my ignore list

Overclocking the Radeon RX 5700 XT: 2.1 GHz on Liquid Cooling

After testing the AMD's new Radeon 5700 GPUs and Nvidia's RTX Super answer, we are particularly happy about the value offered by the latest Radeons. The $400...

www.techspot.com

Haha.. at this rate he has basically everybody ignored that does not agree with his half-technical oppinion.

After the raytracing nonsense he spread about GT7 and he got called out, I stopped reading his posts. Your time is better spent not reading them.

 

[TheGreatWhiteShark]

Honestly the Mr X stuff is all about making FUD to make the PS5 appear like a last gen system. I haven't seen any games to additional GPUs in the power supply that makes the PS5 more powerful than the XSX.

ps5 is not a last gen system. its pretty much very next gen. its the effort to disqualify the differences with xbox, and the xbox in general that is hilarious.
personally I care only that it is a well mannered and silent machine. it will only play its exclusives for me, so I dont really mind for anything else.

 

[MasterCornholio]

its the effort to disqualify the differences with xbox, and the xbox in general that is hilarious.

Depends what you're talking about. Some people are creating FUD where they claim the differences are massive between them. So much so that when the XSX plays next gen games at 4K 60FPs the PS5 will be doing the same at 1080P 30FPs.

The differences between the two are alot smaller than that.

 

[Agnostic2020]

where are these benchmarks? All the stock 5700xt benchmarks ive seen run at 1.9 ghz average. You have to look at the average in game clocks by monitoring it to see where it actually sits. Using AMD's advertised game clocks isnt good enough because they could be 1.4 ghz or 2.0 ghz, you have to take an average.



So going from 1.9 ghz to 2.05 ghz is only a 7% increase. So if you are getting 5-6% from a 7% increase, great.

He is lost on the BS he wants to say . Just put him on ignore.

Remember the Xbox One and Mister X secret sauce stuff? That was hilarious. Now its PlayStation's turn and I'm like

Lol no one is as delusional as mister x media .

Here is his recent fake info chart saying xsx has 80 times better raytracing when it was proven its merely 15% not 8000%(321 gigarays vs 380 gigarays). The delusional team is still the same. They r still delusional unfortunately. This time they r trying to make ps5 look last gen delusionaly . It is what it is . People dont change I guess

 

[Gaz]

you know, this is what I am trying to tell some people, to avoid being in that ridiculous position, and I'm saying it in a very gentle and good will way, only to get bad responses and swears in return.
maybe its better to just get my popcorn and laugh like you do, too.

Honestly you're better off getting the popcorn out and enjoying the show. One company has been awfully transparent about its specs and the results will speak for themselves later on this year.

 

[onQ123]

Remember the Xbox One and Mister X secret sauce stuff? That was hilarious. Now its PlayStation's turn and I'm like

PS5 don't need any secret sauce the real specs are better in most gaming situations.

 

[TheGreatWhiteShark]

He is lost on the BS he wants to say . Just put him on ignore.

I provided the article where a test was conducted on a liquid cooled 5700xt, from 1800 to 2100.
for 18% overclock, around 7% median gains, with ~40% rise on consumption. heat? remove the liquid cooling and use your imagination.
the article describes the results as "disappointing"
I have said many times that the rdna way to increase performance is through CUs, not through clocks.
Ignore list it is for you too.

 

[Boglin]

Clock speeds do not change performance linearly. If you were to underclock the 5700xt by 18% instead of overclocking it you will lose more than 7% performance. 5700xt is already at a threshold where increasing the clocks does not help much but that does not say anything about RDNA 2 and its optimal clock range. It might work amazingly at 2200.

Nobody here knows what the optimal clocks are for RDNA 2 GPUs at any given CU count. We need to wait for benchmarks.

 

[TheGreatWhiteShark]

Clock speeds do not change performance linearly. If you were to underclock the 5700xt by 18% instead of overclocking it you will lose more than 7% performance. 5700xt is already at a threshold where increasing the clocks does not help much but that does not say anything about RDNA 2 and its optimal clock range. It might work amazingly at 2200.

Nobody here knows what the optimal clocks are for RDNA 2 GPUs at any given CU count. We need to wait for benchmarks.

it could. although cerny's talk when he said that they couldn't have it at 2.0ghz constant, as well as microsoft's choice , along with the recent past can be some kind of indicator. 5700xt is on 7nm, not some dinosaur

 

[SlimySnake]

He is lost on the BS he wants to say . Just put him on ignore.

Lol no one is as delusional as mister x media .

lol that guy put US on ignore.

His last post is interesting though. It took him 10 posts to finally post some benchmarks after i literally provided four different benchmarks, but he got there eventually so props to him. It does seem like the clocks are not offering large boosts in game beyond a specific point, though I am still not seeing the average clocks in these benchmarks. The link he posted just looked at one game F1 which sat at around 1800 mhz while I literally posted a gears 5 benchmark showing the 5700xt averaging 1,900 mhz. We dont know how high these average clocks were and how high they were after the boost.

The games are also bottlenecking somewhere else because even though the percentage difference isnt really that great, the overclocked card is managing to keep pace with 2070 super and even the 2080 in some cases which they simply shouldnt be able to do so clearly some of these games arent pushing the rtx cards either.

Of course, he didnt post the actual benchmarks which show how well it actually performs regardless of the smaller bump in percentage.

F1- Better than 2070 Super
Battlefield - Better than 2080.
Resident Evil 2 - Better than 2070 Super.
Forza Horizon is better than the rtx 2080 ti. lol And thats the problem with ingame benchmarks, they differ by game engines and we really need to look at where the clocks sit for each game. Where they were before and after because its entirely possibly that Forza Horizon was already pushing the 5700xt to 1.9 ghz before they overclocked it. Assuming that every game started out at 1.8 ghz or every game topped out at 2.05 ghz is simply incorrect. The delta simply isnt accurate.

The firestrike, timespy and other benchmarks the JayzTwoCents channel did show an amazing improvement in the 5700 because they offer a realtime account of average clocks which is the truest indicator of perf per clocks.

 

[FritzJ92]

I've just been reading everything and enjoying the show!

However, I must add, what is the relative importance of proving that an overclocked 5700 can outperform a higher TF 5700xt when they only have a 4 CU difference? XBSX and PS4 have a 16 CU difference. Thank you and I will return to eating my popcorn and reading.

 

[Boglin]

it could. although cerny's talk when he said that they couldn't have it at 2.0ghz constant, as well as microsoft's choice , along with the recent past can be some kind of indicator. 5700xt is on 7nm, not some dinosaur

I just don't have a lot of confidence in using those as indicators.

Microsofts choice put them at a much higher CU count and they may have chose to run the Series X at a lower frequency for something as simple as wanting to keep the console running quiet and below 300W.
It's hard to judge PS5's frequency against the Series X when it has a drastically higher CU count while still using a relatively low amount of power.

Cerny said 2ghz looked unreachable using a fixed frequency, but that just means for their cooling budget and power usage target.
Still, the PS5 ended up being a behemoth of a console for cooling reasons so I wonder how much power it's actually going to end up using to reach those 2.23ghz speeds.

I'm glad we don't have to wait too much longer to get the full picture.

 

[Godfavor]

From Cerny himself "to reduce power by 10 per cent it only takes a couple of percent reduction in frequency, so I'd expect any downclocking to be pretty minor".

If we put it the other way around, going for 2% more frequency takes a toll of 10% power increase.

Seems to me that the chip is over the optimal range between power consumption and performance.

Correct me if I am wrong.

 

[TheGreatWhiteShark]

Seems to me that the chip is over the optimal range between power consumption and performance.

Correct me if I am wrong.

Well, at least I now have the satisfaction of knowing that at least one person understands.

And consumption brings heat. Much heat.


In all the overclock examples I've seen on 5700xt, their aim with watercooling is to prevent the gpu from spiking down from heat.
Thats the bigger success than the upper limit reached.

....

Edit: if you wanted to be pendatic about it, there is a cube relationship between voltage and frequency, so the reverse example "for a further 2% increase you need a 10% increase in voltage" is not correct. You would need more than 10% increase in voltage to achieve a further 2% in clockspeed. More like 13-15%.
This only furthers your understanding that it is clocked well above optimal

 

[JonnyMP3]

From Cerny himself "to reduce power by 10 per cent it only takes a couple of percent reduction in frequency, so I'd expect any downclocking to be pretty minor".

If we put it the other way around, going for 2% more frequency takes a toll of 10% power increase.

Seems to me that the chip is over the optimal range between power consumption and performance.

Correct me if I am wrong.

Depends on what your 100% mark is. 10% power or 2% of frequency drop leaves 90% or 98%... but you can't go 110% power or 102% frequency within the stable parameters.

 

[Panajev2001a]

From Cerny himself "to reduce power by 10 per cent it only takes a couple of percent reduction in frequency, so I'd expect any downclocking to be pretty minor".

If we put it the other way around, going for 2% more frequency takes a toll of 10% power increase.

Seems to me that the chip is over the optimal range between power consumption and performance.

Correct me if I am wrong.

I appreciate the thought, but disagree with the conclusion. Nobody is arguing with the effects increasing voltage to increase frequency and the inverse bring (for example). You are stating a fact and making a statement that does not seem to quite follow from it or at least that it less novel than you may think (the frequency of both chips is capped).

There is also the missing big pink elephant in the room that is the actual work being done by the chip having different power requirements. Workload being run, as in which exact instructions you are flipping transistors with so to speak, matters for power consumption a great deal. Code can be architected to take that into account too (you may shift some vector crunching code to the GPU and use AVX2 instructions a bit less and/or focus on AVX codepaths and take a bit longer processing your work if you have extra CPU time budget available... frames processing tends to have CPU and GPU being off sync for some tasks where the CPU may be ahead by a frame).

This also disregards efficiency/utilisation percentage of HW (Cerny’s comment about 100% utilisation in many modern PS4 games not being close to the reality at all and for some of them the result would be HW failures due to thermal issues) people have profiled a lot of very high performance external and especially internal games (the latter which have been optimised for PS5 for a good while) to determine efficiency or how much useful work (without stalls, under utilisation, etc...) is done every clock cycle by either CPU or GPU decreasing the chances a downclock is needed, making it more likely power “budget” can be transferred across the units allowing only one to downclock, etc...

 

[Boglin]

From Cerny himself "to reduce power by 10 per cent it only takes a couple of percent reduction in frequency, so I'd expect any downclocking to be pretty minor".

If we put it the other way around, going for 2% more frequency takes a toll of 10% power increase.

Seems to me that the chip is over the optimal range between power consumption and performance.

Correct me if I am wrong.

Seems correct to me when looking at the relationship between power consumption and performance. The PS5 most likely isn't very energy efficient which seems kinda obvious considering it's the less powerful machine but it's also considerably larger in volume. I'm curious if it will end up using an external power brick. :edit: Sony electing to keep the standard console form factor also plays a large role in determining the volume of the console if they're trying to keep the noise levels low.

That doesn't say anything about the frequency to performance ratio though.

 

[Godfavor]

I appreciate your replies, I am not a tech expert at any means. I just rephrased what Cerny said about power consumption and clock speed.

It is true this does not tell the whole story as there is no indication how clock speed translates into performance.

The soc might emit a lot of heat (it doesn't matter if the cooling solution is good enough).

Power consumption is also questionable as we do not know where this thing is capped. We might be able to do some EU research in home appliances in this matter, as EU forces all the companies to be up to a certain limit of power consumption for environmental reasons. So Sony should have been taken this into consideration of they want to sell in EU. The same goes for MS as well.

I just took a guess of how this gpu's behave in relation with power consumption. It does not matter if it is 99% or 101% mark. The chip does not have a defacto number in Mhz or watt. The scaling is what it matters.

 

[BluRayHiDef]

AMD will soon be on 5nm and in a few years 3nm.

We really don't know how much better Ampere is in perf/watt versus RDNA2 yet. I suspect the gap has shrunk quite a bit with RDNA2.

I'm not a technical wizard, but isn't 5nm the smallest node that's free of issues pertaining to the laws of physics? Doesn't some material other than silicon need to be found in order implement node sizes smaller than 5nm?

 

[jonnyp]

I'm not a technical wizard, but isn't 5nm the smallest node that's free of issues pertaining to the laws of physics? Doesn't some material other than silicon need to be found in order implement node sizes smaller than 5nm?

TSMC 3nm is planned for mass production in 2022.

 

[BluRayHiDef]

TSMC 3nm is planned for mass production in 2022.

So, at what node size will silicon be unusable?

 

[jonnyp]

So, at what node size will silicon be unusable?

TSMC is researching 2nm it seems but there is a lot of uncertainty beyond 3nm. Have a look at:

Transistor Options Beyond 3nm

Transistor Options Beyond 3nm Complicated and expensive technologies are being planned all the way to 2030, but it's not clear how far the scaling roadmap will really go.

semiengineering.com

Quite interesting little read

 

[onQ123]

Part 2

 

[TheGreatWhiteShark]

Part 2

ok, let alone everything else, at around the 7 minute mark, this guy says:
"microsoft let amd take the lead designing cpu & gpu, and only asked for a few customizations here and there for certain functionality, like backward compatibility"

this guy now officially has entered my "stupid persons who don't have a clue on what they are talking about" list

 

[geordiemp]

So, at what node size will silicon be unusable?

Why do you think the gate material is silicon just because the wafer is ? High K gate exotic materials been used for ages now.

The pitch on 7m is actually 30 nm, thats the density, its the width of the gate that is 6 nm on TSMC, but they call it 7....

And they are changing the FinFET to be vertical...

 

[geordiemp]

Part 2

So, the elephant in the room, has nobody noticed yet.

Big navi leaked to have L1 cache feeding 10 CU shader array
Ps5 same, 10 CU shader array

XSX has 4 x L1 feeding 4 shaders arrays, and they are 14 CU which is much larger. Is this layout optimised for running 4 instances of XB1s and blade server. The shader arrays are almost 50 % larger than PC RDNA2 and Ps5,....

Answers on a postcard.

 

[FritzJ92]

Part 2

This was terrible on so many levels... he is usually better than this, but seems like the hot chips break down was the limit of his epxertise, so he jus read everything off and repeated rumors and speculations.

 

[MarkMe2525]

here they are, you earned a spot in my ignore list

Overclocking the Radeon RX 5700 XT: 2.1 GHz on Liquid Cooling

After testing the AMD's new Radeon 5700 GPUs and Nvidia's RTX Super answer, we are particularly happy about the value offered by the latest Radeons. The $400...

www.techspot.com

I don't understanding announcing that you are adding someone to your ignore list. Do you think they give a damn? He is also just debating with the same amount of zeal as you are.

 

[funking giblet]

So, the elephant in the room, has nobody noticed yet.

Big navi leaked to have L1 cache feeding 10 CU shader array
Ps5 same, 10 CU shader array

XSX has 4 x L1 feeding 4 shaders arrays, and they are 14 CU which is much larger. Is this layout optimised for running 4 instances of XB1s and blade server. The shader arrays are almost 50 % larger than PC RDNA2 and Ps5,....

Answers on a postcard.

14 CU's doesn't mean "more cached items" but just "more cache bandwidth". So we need to know if it's fully utilized for it to be a bottleneck. Otherwise, more CUs will = faster.

This is from my own personal caching experience with distributed systems, unless you hit the bandwidth limit, adding more horizontal scaling will improve things, if the jobs can truly be loadbalanced.

 

[geordiemp]

14 CU's doesn't mean "more cached items" but just "more cache bandwidth". So we need to know if it's fully utilized for it to be a bottleneck. Otherwise, more CUs will = faster.

This is from my own personal caching experience with distributed systems, unless you hit the bandwidth limit, adding more horizontal scaling will improve things, if the jobs can truly be loadbalanced.

I read it as L1 cache feeding 14 CU at a slower clock will just feed the shader CUs slower, of course we dont know the size of the L1 cache for ps5,, XsX and Big Navi PC parts (which will have same L1 feeding L0 of l0 CU) to compare things properly

All we have is the RDNA1 white paper, I assume XSX L1 bandwidth feeding 14 CU will be effectively slower ?

 

[onQ123]

I read it as L1 cache feeding 14 CU at a slower clock will just feed the shader CUs slower, of course we dont know the size of the L1 cache for ps5,, XsX and Big Navi PC parts (which will have same L1 feeding L0 of l0 CU) to compare things properly

All we have is the RDNA1 white paper, I assume XSX L1 bandwidth feeding 14 CU will be effectively slower ?

Like always I tried to talk about this in a thread

40 CU 9.75 TFLOPs RX 5700 XT vs 64 CU 12.6 TFLOPs RX Vega 64 ( Different architectures but probably the best way to compare PS5 & Xbox SX for now)

Why RX 5700 XT vs RX Vega 64 & not RX 590 vs RX Vega 64? My reason for this is because RX 590 is ROP limited with 32 ROPs while RX 5700 XT has the same 64 ROPs as the RX Vega 64. RX 5700 XT is a newer architecture & benefit from that but at the same time I'm comparing it to the RX Vega 64 &...

www.neogaf.com

 

[funking giblet]

I read it as L1 cache feeding 14 CU at a slower clock will just feed the shader CUs slower, of course we dont know the size of the L1 cache for ps5,, XsX and Big Navi PC parts (which will have same L1 feeding L0 of l0 CU) to compare things properly

All we have is the RDNA1 white paper, I assume XSX L1 bandwidth feeding 14 CU will be effectively slower ?

Oh slower clock will affect bandwidth, of course, but increasing CU count from a pure performance aspect, just means that the bandwidth has to be sufficient.

In that example above... The 40CU model has a lot of headroom in cache, so performance will excel where the bandwidth becomes a limiting factor, which usually happens with high hitrates.

The poison chalice with high hit rates becomes that local caches (ie: shared memory) are preferable to L1, so it becomes an optimization exercise to use these more and reduce data structure size, than trying to saturate L1 cache.

After all, caching is essentially not doing the same thing twice or more, even reading cache itself!