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

[Bouzoo]

Just now, zMeul said:

Anand did only 10 tests, the dude on the forums did 26

And the difference should be almost 100% more? Tom HW showed 0% from Skylake to Kaby Laby yet it's supposed to be 5%?

[79wjd]

3 minutes ago, Bouzoo said:

And the difference should be almost 100% more? Tom HW showed 0% from Skylake to Kaby Laby yet it's supposed to be 5%?

These graphs don't show any improvement from Skylake to Kaby Lake (they don't have results for Skylake), @zMeul was just saying we expect 5% average per generation. 

 

But if you look through some of the benchmarks a sizeable number of them show a 20%~ improvement from Haswell to KL and at least one shows a 50% improvement. The other graphs likely benefit from architectural improvements/instruction set improvements/other and not simply just raw IPC. 

 

What's stands out about Stitl's results is that he's seeing a 14% increase in Cinebench, which seems high compared to everyone else's results.

[Sauron]

5 hours ago, zMeul said:

Opterons aren't mainstream either 

That's not what I was getting at - up until a couple of years ago you could get a xeon that was essentially a lower clocked i7 for i5 prices, now if you want to benefit from HT you need to fork out 350$ for the cpu or buy into a more expensive platform. The hyperthreaded i5 should help fill in that niche.

5 hours ago, Misanthrope said:

If you mean SMT is so badly implemented right now it would be detrimental to performance yes, small is the word I'd use to describe it. 

 

And yes I know it took intel a while to get it right with hyperthreading as well, but time isn't a luxury AMD has right now.

I was referring to the HT i5, not the ryzen, with that statement. I think it's good that intel is reacting and bringing something more value oriented to the table. Hopefully the same will happen at the higher ends of the spectrum, where intel's performance is threatened the most - beyond the point where game performance is no longer that relevant. Nobody in their right mind should buy a 6900k for games and a chip that can keep up with it in more compute oriented applications at half the price is a very enticing offer.

[79wjd]

1 minute ago, Sauron said:

The hyperthreaded i5 should help fill in that niche.

Isn't the 7640k rumored to be on X299? So presumably, the cost savings compared to a 7700k will be made up for in the cost of the more expensive board. 

[zMeul]

4 minutes ago, Bouzoo said:

And the difference should be almost 100% more? Tom HW showed 0% from Skylake to Kaby Laby yet it's supposed to be 5%?

Anand didn't do 10 test, when I looked it over their graph it contained multithreaded apps, they did 4-5 single threaded tests

and with 5 tests you won't get the best picture

[zMeul]

7 minutes ago, djdwosk97 said:

These graphs don't show any improvement from Skylake to Kaby Lake (they don't have results for Skylake), @zMeul was just saying we expect 5% average per generation. 

 

But if you look through some of the benchmarks a sizeable number of them show a 20%~ improvement from Haswell to KL and at least one shows a 50% improvement. The other graphs likely benefit from architectural improvements/instruction set improvements/other and not simply just raw IPC. 

 

What's stands out about Stitl's results is that he's seeing a 14% increase in Cinebench, which seems high compared to everyone else's results.

here's the thing tho, between the original Hawell and KabyLake there are many more CPUs:

• Haswell
• Haswell Refresh (Devils' Canyon)
• Haswell-EP
• Broadwell
• Broadwell-E
• SkyLake
• KabyLake

 

if you take all those into account it averages at 2.4% over each

[Sauron]

11 minutes ago, djdwosk97 said:

Isn't the 7640k rumored to be on X299? So presumably, the cost savings compared to a 7700k will be made up for in the cost of the more expensive board. 

If that's true then it will be a completely pointless product and I take every good thing I said about it back.

[79wjd]

14 minutes ago, zMeul said:

Anand didn't do 10 test, when I looked it over their graph it contained multithreaded apps, they did 4-5 single threaded tests

and with 5 tests you won't get the best picture

There are still discrepancies though: 

 

Cinebench R15: 

A 6700k is 17% faster than a 4770k while being clocked 10%~ higher, which makes a 6700k's IPC at around 7% higher. 

 

Vs. 

Which shows KL 13%~ faster. And it also shows both KL and Haswell scoring significantly lower than Anandtech's results (even after account for the clockspeed differences).

 

Cinebench R10:

A 6700k scores 9052, a 4790k scores 8785, which puts a 6700k about 3% ahead of Haswell. 

 

vs. 

And once again, the results don't match up.

 

5 minutes ago, zMeul said:

here's the thing tho, between the original Hawell and KabyLake there are many more CPUs:

• Haswell
• Haswell Refresh (Devils' Canyon)
• Haswell-EP
• Broadwell
• Broadwell-E
• SkyLake
• KabyLake

 

if you take all those into account it averages at 2.4% over each

Haswell Refresh, Devils Canyon, and Haswell-E are still based on Haswell, and Broadwell-E is still based on Broadwell. They don't add to the number of generations between Haswell and Kaby Lake. And again, Kaby Lake has the same IPC as Skylake, Kaby Lake is just a Skylake refresh. 

[zMeul]

8 minutes ago, djdwosk97 said:

Which shows KL 13%~ faster. And it also shows both KL and Haswell scoring significantly lower than Anandtech's results (even after account for the clockspeed differences).

weren't Anand tests done on W7 by any chance?

in his thread he explained something: W10 performs overall consistently worse than W7 - all the runs he did he did on Widows 10

 

Edited March 4, 2017 by zMeul

[79wjd]

3 minutes ago, zMeul said:

weren't Anand tests done on W7 by any chance?

in his thread he explained something: W10 performs overall consistently worse than W7 - all the runs he did he did on Widows 10

Andandtech tested on W7. If true, that would explain the lower scores, but that doesn't explain why Anandtech's 6700k scored around 3-7% faster than a 4770k/4790k after accounting for clockspeeds while Stilt shows it being 14% faster. 

[zMeul]

7 minutes ago, djdwosk97 said:

Andandtech tested on W7. If true, that would explain the lower scores, but that doesn't explain why Anandtech's 6700k scored around 3-7% faster than a 4770k/4790k after accounting for clockspeeds while Stilt shows it being 14% faster. 

Still tested Kaby, not Sky 

Kaby has no IPC improvements over Sky, but has architectural improvements that add to overall performance: https://www.pcper.com/reviews/Processors/Intel-Kaby-Lake-Performance-Surprising-Jump-over-Skylake/Content-Creation-and-Gam

like being able to shift the core clocks from IDLE to turbo boost much much faster

 

 

the same goes for every of the iterations, like Haswell to Devil's Canyon - architectural improvements that actually add up

[79wjd]

1 minute ago, zMeul said:

Still tested Kaby, not Sky 

Kaby has no IPC improvements over Sky, but has architectural improvements that add to overall performance: https://www.pcper.com/reviews/Processors/Intel-Kaby-Lake-Performance-Surprising-Jump-over-Skylake/Content-Creation-and-Gam

like being able to shift the core clocks from IDLE to turbo boost much much faster

Even looking at Kaby Lake, there are still discrepancies. 

[zMeul]

1 minute ago, djdwosk97 said:

what am I looking at?

[79wjd]

1 minute ago, zMeul said:

what am I looking at?

Cinebench R15 with a 7700k vs. a 6700k, both at 4.8ghz. Kaby Lake is .097% faster (e.g. margin of error). AND a 7700k is only 7%~ faster than a 4790k. So again, we're seeing a 7% single threaded advantage for Kaby Lake whereas Stilt was showing a 14% advantage. 

[zMeul]

29 minutes ago, djdwosk97 said:

Cinebench R15 with a 7700k vs. a 6700k, both at 4.8ghz. Kaby Lake is .097% faster (e.g. margin of error). AND a 7700k is only 7%~ faster than a 4790k. So again, we're seeing a 7% single threaded advantage for Kaby Lake whereas Stilt was showing a 14% advantage. 

I dunno mate, I just did a Cinebench R15 ST W10 test on my i5 6500 that runs at 3.2Ghz and got 149CB

something is very very wrong with Anand's results, my result is very much closer to Slit's Kaby that runs at 3.5Ghz 153CB

 

and it's backed up by EuroGamer's run from last year: 146CB http://www.eurogamer.net/articles/digitalfoundry-2016-intel-skylake-core-i5-6500-review

[79wjd]

4 minutes ago, zMeul said:

I dunno mate, I just did a Cinebench R15 ST run on my i5 6500 that runs at 3.2Ghz and got 149CB

something is very wrong with Anand's results, my result is very much closer to Slit's Kaby that runs at 3.5Ghz 153CB

Your 6500's scores are actually much closer to Anandtech's. A 6700k at 4.8ghz scores 206CB. Your 6500 turbos to 3.6ghz on one core. So 149CB @ 3.6ghz =~ 198.66 @ 4.8ghz. 

[zMeul]

4 minutes ago, djdwosk97 said:

Your 6500's scores are actually much closer to Anandtech's. A 6700k at 4.8ghz scores 206CB. Your 6500 turbos to 3.6ghz on one core. So 149CB @ 3.6ghz =~ 198.66 @ 4.8ghz. 

it turbos for shit because I observe load on all cores - it only turbos when rest of the cores are at IDLE

 

Edited March 4, 2017 by zMeul

[Ezio Auditore]

19 hours ago, Bouzoo said:

As a matter of fact I do.

 

 

You might notice this for instance is a game that benefits from more than 4 cores.

 

http://www.gamersnexus.net/hwreviews/2822-amd-ryzen-r7-1800x-review-premiere-blender-fps-benchmarks/page-7

Is watch dogs 2 actually scaling 16 threads cause there doesn't seem to be any improvement  with the 6900k overclocked to 4.4ghz.

Note: Ryzen SMT is fucked up.

[79wjd]

10 minutes ago, zMeul said:

it turbos for shit because I observe load on all cores - it only turbos when rest of the cores are at IDLE

I'm seeing a lot of discrepancies going on here. It looks like there's a pretty large variance in Cinebench results (up to around 10%). So with that in mind, I don't know how accurate any Cinbench results are unless they have done a few dozen runs and averaged them out. Stilt, presumably didn't do that, but he did average out almost 30 different benchmarks. Of course, that's not quite the same thing, but I wonder if Skylake/Kaby Lake does have a greater IPC advantage over Haswell than what has generally been assumed.  

Spoiler

R15: 

Anandtech: 

6700k @ 4-4.2ghz : 182 -> 218CB - 208CB @4.8ghz

6700k @ 4.8ghz : 206

4790k @ 4-4.4ghz : 181 -> 217CB - 197CB @ 4.8ghz

4770k @ 3.5-3.9ghz: 156 -> 192CB - 214CB @ 4.8ghz

 

Eurogamer: 

6600k @ 4.5ghz : 185 -> 197CB @4.8ghz

6500 @ 4.5ghz : 195 -> 208CB @4.8ghz

6100 @ 4.44ghz : 187 -> 202CB @4.8ghz

 

Stilt:

KL @ 3.5ghz : 153CB -> 210CB @4.8ghz

 

R11.5:

Eurogamer: 

4790k/4690k @ 4.6ghz : 2.0 -> 2.087

6600k @ 4.5ghz : 2.08 -> 2.22

6500 @ 4.5ghz : 2.20 -> 2.347

6100 @ 4.44ghz : 2.12 -> 2.29

 

Stilt: 

KL: 3.5ghz : 1.72 -> 2.359

Haswell: 3.5ghz : 1.50 -> 2.085

@Glenwing @Bouzoo @Sauron @Imglidinhere

 

 

[Misanthrope]

1 hour ago, Sauron said:

I was referring to the HT i5, not the ryzen, with that statement. 

Ooooh the i5 on a platform with motherboards that would be substantially more expensive than the chip itself? That's almost worst.

[MorganO]

2 hours ago, Sharkyx1 said:

Well just like any X intel chip or an nvidia titan. 

what?

[Sharkyx1]

36 minutes ago, KOMTechAndGaming said:

what?

•  

Quote

Don't get the 1800X, it is not worth it...

Translate to

Quote

Don't get the 6900K, it is not worth it...

or

Quote

Don't get the 6950X, it is not worth it...

or

Quote

Don't get the Titan X, it is not worth it...

 

 

Highest tier products are never worth it.

[Imglidinhere]

8 hours ago, Sharkyx1 said:

• Highest tier products are never worth it.

Yup! Even getting a 6800K is iffy. Technically it's the only CPU to really go for with the X99 platform.  

[79wjd]

Just now, Imglidinhere said:

Yup! Even getting a 6800K is iffy. Technically it's the only CPU to really go for with the X99 platform.  

20% more (I'm considering the higher motherboard cost as well) for 50% more cores makes sense and doesn't constitute the flagship tax. The 6850k, 6900k and 6950x, on the other hand, definitely do carry the flagship tax.

 

Of course, now that ryzen is out, it doesn't really make sense.

[tlink]

On 4-3-2017 at 0:52 AM, zMeul said:

there is nowhere to blossom

with CPUs, what you see is what you get

 

it's not like you release a new graphics driver and get couple of % perf gains in some titles

yes, AMD is capable of releasing micro-code updates to fix some of the shit, but don't expect to magically close the gap

 

AMD needs to get their asses in gear right now for Zen2, if they aren't already on it

hey look, huge performance differences between operating systems. gee maybe we shouldn't completely trample the idea before its even tested.
 

Quote

Windows 10 - 1080 Ultra DX11:

8C/16T - 49.39fps (Min), 72.36fps (Avg)
8C/8T - 57.16fps (Min), 72.46fps (Avg)

Windows 7 - 1080 Ultra DX11:

8C/16T - 62.33fps (Min), 78.18fps (Avg)
8C/8T - 62.00fps (Min), 73.22fps (Avg)

If this isn't a large performance gap ON THE SAME PROCESSOR than i don't know what is, this shows quite some room for performance improvements if its to be believed. for 8c/16t thats a difference of 26% for min fps, and 8% for the avg fps. thats massive. for 8c/8t its a min fps difference of 8% in min fps, and a 1% difference in avg fps. smaller but the minimum fps difference still is quite significant. hence why there should be more testing of things like this and optimization. like the post says there where other factors influencing it but a difference of 26% is really hard to explain by another factor tbh. not saying im taking this as proof, but more as hints that we shouldn't jump to conclusions too soon.