TPU Investigate thier own Zen 3 gaming results, some interesting finds

[humbug]

Back when Zen 3 first came out like everyone else TPU reviewed them in games, but their results were quite different to all other large review outlets, they were the only one who put Intel ahead in gaming performance, if only by a few %.

This led to them being accused of Intel bias, some people put Anand Tech slides on their forum in which AMD was consistently faster than Intel, anything from 10 to 30% faster.

So the thing is TPU used a 2080TI, where as everyone else used a 3080 or a 3090.
The 2080TI is not fast enough to get bottlenecked by a Zen 3 CPU, not even at 1080P.
That doesn't explain why Intel had the gaming lead in TPU's charts.

To be fair to TPU they realised this so they investigated.
What they found was that on Turing (RTX 2080TI) at a low GPU load of 0 to 40% Zen 3 was 20% faster, between 90 and 100% GPU load the performance would stall ever so sightly on Zen 3 giving Intel a slight lead of a few %, given that TPU's testing was basically with GPU loads of 90% and higher that's what resulted on their charts.

At this point TPU theorised it was something to do with Zen 3 idle state switching, Zen 3 is particularity aggressive with this, it will put unused cores in to sleep mode at the slightest opportunity, its incredibly fast at waking them up again but no matter how fast it is if its not instant you will lose cycle time on that core.

The Turing vs Ampere effect.

They tested it again but this time with an RTX 3090 and an RTX 3070, the latter an Ampere GPU with the equivalent performance of a 2080TI.
This time at 0 to 50% GPU load Zen 3 was 25% faster and critically at no point did Intel ever score higher than Zen 3.

So its probably not the aggressive idle state of Zen 3, its something in Turing.

The two important charts are here......

https://www.techpowerup.com/review/intel-10900k-vs-amd-5900x-gaming-performance/

 

[Cyber-Mav]

I don't understand? Whats being described is simple cpu bottlenecking.
At low gpu load you get more cpu usage and higher fps. When gpu load is high there is less emphasis on cpu power and both intel and zen3 are similar.

Whats this thread about?

 

[humbug]

I don't understand? Whats being described is simple cpu bottlenecking.
At low gpu load you get more cpu usage and higher fps. When gpu load is high there is less emphasis on cpu power and both intel and zen3 are similar.

Whats this thread about?

Have you read it? you shouldn't be asking that if you had read it.

 

[Cyber-Mav]

I didnt read article im on my phone. Will chek it when i get bak home on pc

 

[Joxeon]

The floor in the testing is though is the 10900k is running stock yet overclocking benefits Intel more than AMD and also they use the same 3800/16 which again plays into AMDs favour as its pretty much the best AMD can run while the 10900k can run 4400mhz.

 

[Cyber-Mav]

going by the charts intel maxing out around 540fps. could be more cache on zen3 keeping cpu better fed.

 

[jigger]

Definitely interesting. Have they asked nvidia to comment?

 

[humbug]

The floor in the testing is though is the 10900k is running stock yet overclocking benefits Intel more than AMD and also they use the same 3800/16 which again plays into AMDs favour as its pretty much the best AMD can run while the 10900k can run 4400mhz.

The floor in the testing is though is the 10900k is running stock yet overclocking benefits Intel more than AMD

The Flaw? the 10900K boost to single core 5.3Ghz and 4.8Ghz all core stock, overclocked they run about 5 to 5.3Ghz all core, at best that's 10% and probably more like 5%, you're not catching Zen 3 stock per core performance unless you're running about 7Ghz, so no.

and also they use the same 3800/16 which again plays into AMDs favour as its pretty much the best AMD can run while the 10900k can run 4400mhz.

This has been tested time and time again and doesn't make the blindest bit of difference beyond about 3800Mhz, there is a reason for that, about 3800Mhz the IMC runs at half speed. Rocketlake is the first Intel CPU to have an option force 1:1 Ratio, like Zen 2 and 3, and like Zen 2 and 3 you're lucky if you can run a 2000Mhz IMC at a 1:1 ratio let alone 2200Mhz.

Definitely interesting. Have they asked nvidia to comment?

No, doesn't really matter much anyway.

 

[stooeh]

I think Joxeon does have a point with regards to the RAM and some games do benefit like Warzone.

 

[Grim5]

Back when Zen 3 first came out like everyone else TPU reviewed them in games, but their results were quite different to all other large review outlets, they were the only one who put Intel ahead in gaming performance, if only by a few %.

This led to them being accused of Intel bias, some people put Anand Tech slides on their forum in which AMD was consistently faster than Intel, anything from 10 to 30% faster.

So the thing is TPU used a 2080TI, where as everyone else used a 3080 or a 3090.
The 2080TI is not fast enough to get bottlenecked by a Zen 3 CPU, not even at 1080P.
That doesn't explain why Intel had the gaming lead in TPU's charts.

To be fair to TPU they realised this so they investigated.
What they found was that on Turing (RTX 2080TI) at a low GPU load of 0 to 40% Zen 3 was 20% faster, between 90 and 100% GPU load the performance would stall ever so sightly on Zen 3 giving Intel a slight lead of a few %, given that TPU's testing was basically with GPU loads of 90% and higher that's what resulted on their charts.

At this point TPU theorised it was something to do with Zen 3 idle state switching, Zen 3 is particularity aggressive with this, it will put unused cores in to sleep mode at the slightest opportunity, its incredibly fast at waking them up again but no matter how fast it is if its not instant you will lose cycle time on that core.

The Turing vs Ampere effect.

They tested it again but this time with an RTX 3090 and an RTX 3070, the latter an Ampere GPU with the equivalent performance of a 2080TI.
This time at 0 to 50% GPU load Zen 3 was 25% faster and critically at no point did Intel ever score higher than Zen 3.

So its probably not the aggressive idle state of Zen 3, its something in Turing.

The two important charts are here......

https://www.techpowerup.com/review/intel-10900k-vs-amd-5900x-gaming-performance/

This unfortunately is a risk people take

when you see a cpu/gpu review - those scores is for THAT system not YOUR system. So many thing affect performance, you buying the gpu/cpu being reviewed doesn't mean you'll get the same performance

 

[alec]

This is almost a year old.

11900K review came after that, and there nobody can accuse TPU of one bit of Intel bias. Quite scathing if I recall.

• Lower gaming performance than AMD Zen 3 and Core i9-10900K
• Application performance can't even match Ryzen 5800X
• High power consumption

 

[humbug]

This is almost a year old.

11900K review came after that, and there nobody can accuse TPU of one bit of Intel bias. Quite scathing if I recall.

Yeah, its not the accusation i'm making, i just think their findings are interesting.

 

[Cyber-Mav]

Its how they came to that result which is interesting. They used a custom static scene in unreal engine so there is no game physics or enemy scripts going on. Hardly a real game scenario not sure what to make of it and odd why intel seemed to cap itself around 530fps. Its like it doesnt like light loads.

 

[HangTime]

It's not odd, it's just the cpu bottleneck kicking in. Reducing the GPU load at that point makes little difference which is why it appears 'capped', if they reduced the cpu load it would go up.

 

[Cyber-Mav]

On op's charts the intel cpu hardly rises in fps going from 60% cpu limited to 0% cpu limited. Looks like a cap of some sort or maybe certain part of the cpu holding it back like integer or fpu or something else?

 

[alec]

On op's charts the intel cpu hardly rises in fps going from 60% cpu limited to 0% cpu limited. Looks like a cap of some sort or maybe certain part of the cpu holding it back like integer or fpu or something else?

this is bad labeling on the chart. X axis shows gpu load. So at 0% it is most CPU limited, at 100% most GPU limited. Thus between 0 and 60% gpu load, CPU is the main FPS limit for Intel, that is just how fast it is.

 

[Cyber-Mav]

this is bad labeling on the chart..

now it makes sense. they should sort out chart labels better.