a look at (ry)Zen's IPC - from Excavator to KabyLake

[zMeul]

Just now, N3v3r3nding_N3wb said:

The 6950x is slower than the 6700k in gaming.

 

Mate, much logic!?

mate, you missed the point and ended up in the same trap

 

[N3v3r3nding_N3wb]

Just now, zMeul said:

yes, Sky over Kaby doesn't have a IPC gain, but tech in Kaby does make it faster - specifically the SpeedShift v2; this was covered quite a bit on PCPer , and I think Anand did an article on it also

Yes, in very specific use-cases, Kaby Lake is better than Skylake clock-for-clock. That's how Intel got away with claiming a 15 percent performance increase (in SysMark).  This is irrelevant, however, because we're discussing IPC, and the IPC advantage of Kaby Lake over Skylake is, of your own admission, nothing.

[zMeul]

Just now, N3v3r3nding_N3wb said:

Yes, in very specific use-cases, Kaby Lake is better than Skylake clock-for-clock. That's how Intel got away with claiming a 15 percent performance increase (in SysMark).  This is irrelevant, however, because we're discussing IPC, and the IPC advantage of Kaby Lake over Skylake is, of your own admission, nothing.

yes, but you cannot extract pure IPC numbers do you

[N3v3r3nding_N3wb]

4 minutes ago, zMeul said:

mate, you missed the point and ended up in the same trap

 

I will admit that I took it out of context, but your conclusion that Ryzen 5 will automatically be worse than Intel's i7s because the R7 lineup, even with more cores, is slower than the i7s was faulty.  I was highlighting the fact that more cores do not make a CPU better at gaming -- in an admittedly snarky and smart-assed way.

[zMeul]

Just now, N3v3r3nding_N3wb said:

I was highlighting the fact that more cores do not make a CPU better at gaming -- in an admittedly snarky and smart-assed way.

that's exactly my point, only didn't added /s

[N3v3r3nding_N3wb]

Just now, zMeul said:

that's exactly my point, only didn't added /s

Ah, well, sorry for taking you out of context, I'm just not willing to read 13 pages just so I can argue over one post.  I normally don't do that and, again, apologize.

[N3v3r3nding_N3wb]

13 minutes ago, zMeul said:

yes, but you cannot extract pure IPC numbers do you

Well, going by that logic, the Ryzen processors can't definitively be said to have Haswell's IPC or even to have worse IPC than Intel's current processors because (insert extremely cherry-picked benchmark).  Technically, IPC depends on the application, but in general, you can match processors up with each other based on the vast majority of benchmarks.  

Edited March 14, 2017 by N3v3r3nding_N3wb
Included cherry-picked benchmark

[Spenser1337]

People forget it literally nuts on the 6900k in workloads

 

its half the price

 

seems good to me 

 

[zMeul]

 

system in the left: 1800X; and in the right i7 5960X

sumary: same 4Ghz, same Premiere project

 

export time: 2min30sec slower on the Ryzen system - witch confirms, once again, the lower IPC of the Ryzen CPUs compared to Haswell

 

[The Sloth]

9 minutes ago, zMeul said:

 

system in the left: 1800X; and in the right i7 5960X

sumary: same 4Ghz, same Premiere project

 

export time: 2min30sec slower on the Ryzen system - witch confirms, once again, the lower IPC of the Ryzen CPUs compared to Haswell

 

 

windows also needs to step up 

 

[Cheddle]

These results are woefully inconsistant with what is being shown in the other 100+ reviews floating about the place. 

[Daniel Z.]

On 2017/3/4 at 6:48 PM, zMeul said:

here's the thing tho, the i7s are cheaper (up to a point)

and the i5s are cheaper still

 

why would you pay premium for less performance? it doesn't make any sense

*cough* youtubers who can't afford X99. I am looking into a used X99 or new Ryzen because for me, gaming is 2nd and folding/rendering/editing is 1st

[Valentyn]

On 21/03/2017 at 1:33 AM, zMeul said:

system in the left: 1800X; and in the right i7 5960X

sumary: same 4Ghz, same Premiere project

 

export time: 2min30sec slower on the Ryzen system - witch confirms, once again, the lower IPC of the Ryzen CPUs compared to Haswell

 

Ignoring that the Intel System has 32GB RAM vs 16GB on Ryzen, and two GTX 1080's vs two Maxwell Titans.

The 1080's alone would make a difference in rendering as they're faster than the Maxwell cards.

A better test would be using the same RAM, and GPUs on both systems as well.

Not double the ram, and essentially two GTX 1070's vs 1080's.

Or if he wanted to only do a proper Encoding test, both systems with the same RAM, and then use something a GT 730 graphics card, and the software render instead of the Default CUDA one it uses when it detects an NVIDIA GPU.

[zMeul]

2 hours ago, Valentyn said:

-

not saying his test is perfect, but he is showing real world workloads and not CPU benchmarks

CUDA is to be used because it's faster and the whole point behind paying more for these Ryzen CPUs is that when CUDA/OpenCL/OpenGL acceleration is available it doesn't really matter

 

look at Puget's benchmarks I posted a while back

[Valentyn]

36 minutes ago, zMeul said:

not saying his test is perfect, but he is showing real world workloads and not CPU benchmarks

CUDA is to be used because it's faster and the whole point behind paying more for these Ryzen CPUs is that when CUDA/OpenCL/OpenGL acceleration is available it doesn't really matter

If he wants to test correctly, he needs the same variables at least, not just the same clock speeds.
Having faster CUDA cards, and double the RAM would give the 5960X an advantage that 1800X didn't have. 
The only way to remove that issue is a pure CPU encode and render while having the same RAM for his project.

For Instance in a small 4K project I can easily use 22GB of RAM in Premiere Pro.

 

 

 Jay didn't even mention or show how much his projects were using, which I found quite the oversight really.

 

On the note of the CPUs don't matter, that's completely false. The Encode is primarily CPU bound, while rendering can be accelerated by Graphics cards.

 

The fact that there's a 8% difference overall makes it leans more towards the GTX 1080's and possibly the RAM being the deciding factor there; not solely the processors.

 

From Puget, we can see how the GPUs make a difference, and the 1080 is faster than the 1070/Maxwell Titan

 

https://www.pugetsystems.com/labs/articles/Premiere-Pro-CC-2017-GeForce-GTX-1080-Ti-11GB-Performance-912/

[juri-han]

i got the feeling ryzen is going to fall under the same 'fine wine' meme as their cards, techspot pretty much did the same thing linus did looks like the simulated 1600x and 1500x were on par or better in game then i5s but were going to have to wait for windows to get optimized before ryzen can hang with i7s in raw FPS... but the 1600x is pretty much on par with the 1800x for clock speeds and $100 less then a 7700k   

 

http://www.techspot.com/review/1360-amd-ryzen-5-1600x-1500x-gaming/

[zMeul]

1 minute ago, Valentyn said:

The fact that there's a 8% difference.

I would really love if people stopped using percentages as actual differentiating factor - it's bullshit!

the diff is 2min30sec on a ~25min workload - that's how it should be presented!

[Valentyn]

10 minutes ago, zMeul said:

I would really love if people stopped using percentages as actual differentiating factor - it's bullshit!

the diff is 2min30sec on a ~25min workload - that's how it should be presented!

 

And that somehow invalidates my argument?

 

Even Puget found in Premiere Pro the GTX 1080 is 6.5% faster than the Maxwell Titan X.

 

Asking for consistency in testing is hardly a radical notion.

 

Quote

The GTX 1080 is about 4% faster than the GTX 980Ti and about 6.5% faster than the Titan X.

 On this system, the GTX 1070 was about 8.5% faster than the older cards. The GTX 1080 was a even faster, beating the older cards by about 10%

 GTX 1070 and 1080, one thing we did find was that dual GPU configurations can often work really well for Premier Pro. We didn't see much of a gain when exporting to 4K (only 2.5-10% better performance), but exporting to 1080p and rendering previews was anywhere from 20% to 50% faster with two video cards versus just one. 

https://www.pugetsystems.com/labs/articles/GTX-1070-and-GTX-1080-Premiere-Pro-Performance-810/

 

 

 

[zMeul]

Just now, Valentyn said:

 

if you're going to continue and use percentages, I'm done

 

ps: Puget doesn't allow hotlinking

[Valentyn]

6 minutes ago, zMeul said:

if you're going to continue and use percentages, I'm done

 

So PugetSystems, that you heralded earlier as proof for Premiere Pro CPU tests are now useless for metrics because they use percentages. 

 

Yes, you are done.

[zMeul]

5 minutes ago, Valentyn said:

Yes you are done.

percentages without context means utter shit

yes, Puget's methodology is quite baffling when measuring differences

 

the workload in Puget's testing is literally ~1minute - you are literally debating seconds in differences

 

100 FPS vs 106 FPS - that's 6%, but without context it means shit

 

 

[Mr.Meerkat]

2 hours ago, zMeul said:

not saying his test is perfect, but he is showing real world workloads and not CPU benchmarks

CUDA is to be used because it's faster and the whole point behind paying more for these Ryzen CPUs is that when CUDA/OpenCL/OpenGL acceleration is available it doesn't really matter

Real world workloads=unfair tests...right...

RAM capacity intensive scenario, uses less RAM

Potentially Cuda intensive scenario (depending on settings)

, uses worse cards

 

Comparisons should be apple to apple, not apple to apple fruit salad.

 

It's like comparing the RX480 8GB versus the 1060 6GB however the RX480 has an i7 and 8GB of RAM while the 1060 only has an i5 and 4GB of RAM. That doesn't seem fair does it?