AMD Ryzen (Summit Ridge) Benchmarks Thread (use new thread)

Page 157 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.
Status
Not open for further replies.

hrga225

Member
Jan 15, 2016
81
6
11
Waiting for RyZen to officially lauch feels like being a child on Christmas Eve!

PLEASE SANTA when are you going to get here!:fearscream:
Spoken like a true enthusiast!

Most people after all news were like
kNyE-clapping-slow+clap-clap
giphy.gif

instead they should be more like
iu

oLvl.gif
 
  • Like
Reactions: prtskg

bjt2

Senior member
Sep 11, 2016
784
180
86
It's a pretty much a given it will be cheaper. AMD needs to break the inertia of people buying Intel. Also there is plenty of room for AMD to price the Ryzen for decent margins. There will probably be a top bin gold sample part that's going to be expensive. But I am convinced the main SKUs will be significantly cheaper than Intel.

If AMD was going after the $1k CPU market they would have included quad channel memory. Their intentions are clear, 8c/16t will be sold at mainstream prices. Around $500 or less.
I bought for work a 3820 with 2011 to have the quadchannel (i perform heavy and memory hungry calculations)... Newer notebooks with newer intel architecture are quite comparable in speed... With good prefetcher only simple calculations on many data are memory bound... I wrote also CUDA software (image filtering) that is L1 limited and not memory limited... So with enough L1/2/3 bandwidth, the memory performance will matter only for iGPU. But 2011 chips and Zen does not have iGPU...
 

KTE

Senior member
May 26, 2016
478
130
76
This is certainly becoming very entertaining...

Sent from HTC 10
(Opinions are own)
 

USER8000

Golden Member
Jun 23, 2012
1,542
780
136
Most of what you posted were GPU limited tests, We don't even know the details on how the games were tested by CanadrdPC, and which were CPU limited, etc, except to say there's clear scaling between the i5's with different frequencies, so, on average there is at least some CPU limitation.

Given most of the titles tested are quite old, we know by default they won't scale well with higher thread count, and it's plain obvious from the fact Skylake is soundly beating BW-E. Something that rarely happens in modern titles.

He also ignores the fact that the FX8370 and A12 9800 are tied in the gaming tests. The A12 9800 has core IPC improvements but is hampered by only 2MB of L2/L3 cache in total and also cannot hit maximum boost speeds due to TDP limitations. That hints at most of the games using at most 4 threads or less,and being more limited by single core performance.

Its literally staring people in the face. They are so desperate to prove Ryzen has not improved per core performance,that they ignore the bottom of the gaming graph.


tic6i57uob5y.jpg



A Haswell Core i3 dual core,is faster in 4 of the 6 games than an FX8350.

Gaming_08.png


IMG0039213.png

GRIDAutosport_proz_amd.jpg


fc_proz.jpg


The socket 2011 CPUs also have much larger caches,etc which do help a bit,just like the Broadwell CPUs with L4 cache which nicely helped with performance.

Either way,its obvious per core performance and indeed IPC have massively gone up.

In fact some of the posts here are approaching the same level of denial we saw when CanardPC leaked the first Athlon 64 scores.

Yet,those Athlon 64 ES chips only ran at 1.4GHZ,and the ones which were released ran at 1.8GHZ,and this looks like history repeating itself again.
 
Last edited:

psolord

Golden Member
Sep 16, 2009
1,875
1,184
136
Kinda offtopic, but just want to add, if we're talking about games, IMO, Starcraft 2 would be the most revealing game if you want to test a CPU's single core gaming performance. Starcraft 2 should be in every single game-based CPU benchmarks. Not to mention since the game is absolutely CPU limited, memory speeds also scale well with OCs.

You are quite correct thinking a Starcraft 2 benchmark as a good gaming cpu benchmark, but during my 2500k/860/Q9550 testing on the GTX 970, I have found some others that are even worse.

https://forums.anandtech.com/thread...rks-bonus-i7-860.2428699/page-2#post-37391692

Actually Diablo 3 seems to be the leader, since it maintained a very high delta for both the Q9550 and the 860. Thief is also a good candidate but it seems that newer cpus are mostly OKish.Unreal Engine 4 although multhithreaded, is very demanding on the cpu too.

Mafia 2's benchmarks with hardware accelerated Physx, does a number on the cpus too. And of course let's not forget GTA IV. I recently tested it on my GTX 1070 and I only got 59fps from 53fps with the 970, with very low GPU usage of course, indicating a huge cpu limit. This game is a timeless cpu test. Haven't tested any of the other top cpu limiters however to see what's what.

If Rysen is any good however, I will be testing them myself anyway because I will be having one, thank you very much! :D
 
Last edited:
  • Like
Reactions: jelome1989

superstition

Platinum Member
Feb 2, 2008
2,219
221
101
Its not about just setting some flags, you have to write/ refactor your code so that unrelated variables can be packaged together and executed with the same operation. Its so HARD that Compiler developers spend a massive amount of time & money on making their compilers do auto-vectorisation and guess what it still sucks and doesn't work well. Those lazy ass Compiler developers who do they think there are with their fancy Comp Sci based degree's and PHD's if they just stopped being so lazy :rolleyes:.
Remind me where I was talking about compiler developers being the central issue here. Rather, it is that the Blender devs didn't bother to use a compiler that uses modern instructions. But nice red herring.
I guess all computer game Dev's are lazy for not using AVX either, you realize you now expect the developer to maintain two complete different sets of code for one application. One factored in a way that allows the Dev to take advantage of AVX/AVX2/FMA operations and another to keep all the Pentiums and Celerons K10's westmere's etc usable.
And now we come to the false dilemma fallacy.

The Stilt's builds ran just fine on Lynnfield. Didn't you read my posts and see the data? In fact, his SIMD build was a bit more efficient than the recent stock builds.
You obviously dont understand, Compilers have many many flags that can be set, there are always trade offs to be made. There is then the version of the compiler which can have a big impact and the COST of the compiler can also be a factor. Choose the wrong optimization and things can go horribly wrong depending on how your code operates.
The Stilt's SIMD build runs faster on Lynnfield by a very small amount, vastly faster on Piledriver, and vastly faster on Haswell/Skylake.

I am seeing a false dilemma on your part.
all the things that tried to just be a benchmark get gamed to hell and back ( SPEC for example).
Another red herring.
All applications are benchmarks even if they are completely crap internally. They all tell you something if you take the time to understand.
Another red herring. Your argument and mine are not the same. Mine is that if you're going to use a benchmark, especially basically by itself, to claim something then you should be certain it's up to the task. The Stilt's builds show that the stock Blender builds aren't.

Unless, of course, you want to get in a time machine and go back to 2010 or so.

It was The Stilt himself who chided me about my complaint that AVX-2 might be abused in an upcoming Cinebench release to make Zen look bad. He said modern instructions should be used. Maybe you should talk to him about that.
Blender is telling us that Bulldozer has internally bottlenecks that other cores don't have.
No. The stock Blender builds are telling us that they're not using modern instructions and are therefore anachronistic.
No it isn't, its the complete point that you choose to ignore and dream up other crap like its the app and the Dev's fault. Those lazy dev's (shakes fist).
Unless it took The Stilt a lot of effort to make those builds your point is invalid. And, Blender has been in development for a long time. It makes perfect sense to assume that the effort put into coding the program is far beyond the effort it takes to recompile it with a decent compiler.
Bulldozer has serious technical issues why do you think it's performance is so bad per clock! Go look at an ARM A73 core its a narrower design yet smokes it in performance per clock, why?
If you want to harp about Bulldozer (which is particularly amusing since I benchmarked Piledriver NOT an 8150) then do as I ask and post the data progression with The Stilt's builds. The SIMD build would suffice.

Then we'll see how relevant Piledriver's flaws are in the big picture. One thing we already know is that it won't post an unfair advantage like Lynnfield does because it will actually have the benefit of the modern instructions that increase its performance dramatically.
The point is its not just setting a flag you have to make your code fit the model the execution units use. if you go use the 2.75 code base and compiled it with AVX,AVX2,FMA/etc you would see that. Then what does refactoring your code like that do to products that don't support that optional instruction set.
The vast speed increase on a processor design from 2012 (has anyone tested the 8150 yet?) that also is seen on current Intel processors is clearly worthwhile. Fixing any bugs relating to using a clearly superior compiler is clearly worthwhile.
You realize Blender is already using 128bit SSE operations its not like those lazy developers didn't know how to use SIMD:rolleyes:
Another red herring. The stock Blender builds are super slow unless you're still running Lynnfield.
the FACT is bulldozer has issues with large amounts of FP operations in flight
the FACT is bulldozer has issues with a large amount of FP stores in flight
the FACT is bulldozer/piledriver has to round Robbin its instruction decode between both cores
the FACT is if you vectorise something (pack 4 32 bit ops into 1 128bit op) you reduce:
the amount of instruction Decode from 4 to 1
the amount of scheduled ops from 4 to 1
the amount of stored data from 4 to 1
Great!
There is no conspiracy here if bulldozer see's a larger gain from the REFACTORING of code then other products then Bulldozer has a bottleneck in a space where no other Core does! Now given one of the Zen architects explicitly called out this issue with Bulldozer.........
Bulldozer Bulldozer Bulldozer! Repeat it three times and maybe WInona Ryder will pop out of a cake.

Let's completely ignore the huge speed increase posted by people who have tested Haswell and Skylake.

Once again, it's clear that the big speed increase The Stilt's builds provide to the current Intel CPUs is due to the deficiencies of the Piledriver (Oh, sorry... I mean Bulldozer) architecture.
This is going to be my last post on the matter, im not wasting anymore time.
I appreciate that.
 

itsmydamnation

Platinum Member
Feb 6, 2011
2,743
3,074
136
There is no false dilemma, thats just you failing to see the forest through the trees. Your missing the big picture and just ignored what i said. Here let me point you at it:

The better question is what actually changes between SSE build and AVX/AVX2 build?
Is the compile creating more friendly code thax to 3 operand?
are some of the new optimization in the code base that allow such a big uplift removing a bottle neck?
Is the improvement from using a different/newer compile that handles bulldozer better?

Just because whatever refactoring they have done hasn't hurt older CPU's doesn't mean it's anywhere near universally true. If you view a program as having logic (if/or/loop/jmp/etc) and maths. SIMD is all about executing that maths quicker/cheaper. Now if you refactor your logic (taking a performance hit) to get an even greater gain from using a new instruction what just happened to everything that doesn't support that instruction?

I dont know how many more ways i can politely tell you your wrong? Maybe go post your hypothesis (whatever that actually is other derp derp blender) on stackoverflow / realworldtech or something and see how you fair.

Simple question why did AMD feel the need to completely re-engineer their FPU unit in Zen if bulldozers is fine and its just blenders lazy devs fault?

I actually find it mildly offensive that you seem to think i would get stuck is such a simple logic paradox, i've supplied a massive amount of logic to backup my position, you actually haven't tried to understand anything.
 
  • Like
Reactions: CatMerc and cytg111

Lebdnil

Junior Member
Jul 13, 2016
11
8
81
He also ignores the fact that the FX8370 and A12 9800 are tied in the gaming tests. The A12 9800 has core IPC improvements but is hampered by only 2MB of L2/L3 cache in total and also cannot hit maximum boost speeds due to TDP limitations. That hints at most of the games using at most 4 threads or less,and being more limited by single core performance.

Its literally staring people in the face. They are so desperate to prove Ryzen has not improved per core performance,that they ignore the bottom of the gaming graph.


tic6i57uob5y.jpg

Let's use your logic with data shown in the review. You claim that games don't scale beyond 4 threads (just ignoring for a bit than FX8370 does 73.6% in games vs 69.8% for A12-9800 and you take that for "tied"). Very well, just here we have just 3 datapoints: Core i5-6400 (3.1GHz all core turbo), Core i5-6500 (3.3GHz all core turbo) and Core i5-6600. All of them have exactly the same architecture, probably same mainboard and so on, so they just differ in frequency. Presto, some Excel (calc, btw) and you get a beautiful linear regression to correlate frequency with the score in games (don't adjust well, so here you have your first tip than games don't scale linearly with frequency).

1- First, just linear regression: [Freq in MHz] = 66.72972x[Score] - 2994.1337. Now put "98.6" in score and you get 3598MHz as frequency needed to reach that score. Ignoring for a bit than zero score points to -2994MHz frequency ;-), let's input 118.2, which is the score of the Core i7-6700K. You get 4892MHz, so, assuming games don't scale beyond 4 threads, this really proves than Sylake core in Core i7-6700K has a 16% higher IPC (4892/4200) than Skylake core in Core i5 family. Right, sure.

2- Well, as that linear regression don't cross 0,0, let's use a second order polynomial regression: [Freq in MHz] = 0.33466x[Score]^2 + 3.3795x[Score]. Again, let's check with "98.6" and this time gets 3587MHz as the frequency needed for that score. And now using "118.2" we get 5075MHz as the frequency needed to reach that score; this time, with a regression that goes to zero frequency when score reach zero, Skylake core in Core i7-6700K has, amazingly, 21% higher IPC than Skylake core in Core i5 series.

So, sorry but games scale beyond 4 threads, as just you could verify using Skylake cores only. Now a small disclaimer: just because Skylake core scale more or less with those regressions don't mean than OTHER cores scales the same. We don't have more datapoints beyond Skylake (A12 and FX8370 use different cores and differ in core count, i7-6800K and i7-6900K differ in core count, and there are just one entry for Zen and Haswell quadcore), so we couldn't know how every other core scales in those games. And yes, I do think Zen has made a huge leap in IPC vs Con cores (just compare to FX8370), but also think this review points at Zen reaching an IPC just like Broadwell or a bit behind (rather behind in games, unless Zen scales much better with frequency).

Byes
 
  • Like
Reactions: Sweepr

USER8000

Golden Member
Jun 23, 2012
1,542
780
136
He also ignores the fact that the FX8370 and A12 9800 are tied in the gaming tests. The A12 9800 has core IPC improvements but is hampered by only 2MB of L2/L3 cache in total and also cannot hit maximum boost speeds due to TDP limitations. That hints at most of the games using at most 4 threads or less,and being more limited by single core performance.

Its literally staring people in the face. They are so desperate to prove Ryzen has not improved per core performance,that they ignore the bottom of the gaming graph.

You claim that games don't scale beyond 4 threads

So, sorry but games scale beyond 4 threads

No,you are claiming that I claim that,by lying through your teeth.

I said MOST(not ALL) of those games tested,ie,4 of those 6 games tested,don't scale well with more cores and threads. They are far more sensitive to single core performance,cache and don't really make any real use of above 4 cores.

But you would know that if you knew anyone who played those games and conveniently editing out those other charts is not helping you.

You are trying to help your mate out by trying to sound clever.The damage limitation is getting serious now!! Basically ignore all evidence about the games being tested and the AMD CPUs being tested.

Gaming_08.png


FX8350 has no advantage over the FX4350. Core i5 4670K and Core i7 4770K are the same. An overclocked Pentium is almost the same.

ARMA3 is very single core performance limited and it needs two to three strong threads - it does not really scale well beyond 4 threads. Even going back to ARMA2,cache also helps too - more L3 cache does improve performance.



IMG0039213.png


Oh wait,lets look what HT is doing for the Core i7s against the Core i5s - nothing. What about the FX4000 against the FX8000 -nothing. Any minor differences are explained by clockspeeds and cache amounts. The six core Phenom II X6 is loosing to a quad core Phenom II X4.

AMD-Vishera-FX-8350-Review-Anno-2070.png


Looks like even a Core i3 3240 beats an FX8350. HT makes no difference. The difference between even the Phenom II X6 and X4 is down to actual boost clockspeeds(Phenom X4 960T is slightly higher than the 1055T and the 1090T is higher than both). FX8150 and FX4170 tied due to similar boost clocks. HEDT Core i7 has far more cache which helps in gaming over bog standard cores. Example:

http://imgur.com/a/aZ7Ji

ARMA2 is very lightly threaded(its quite old),but you can see the extra cache helping there.


GRIDAutosport_proz_amd.jpg


Wow,mega thread scaling there - that is why the Core i7s are dong so much better than the Core i5s.

Plus look at the difference between the Core i7 3970X and Core i7 2600K(Sandy Bridge). So much mega scaling happening there,which is easily explained by the extra cache,and extra memory bandwidth of the platform. A whole 10% - LMAO.

The FX8350 also tends to maintain higher clockspeeds than the FX6300,due to a combination of a better stock cooler and a higher base clockspeeds - I know this from friends with the said CPUs. So the difference is basically explained by the higher boostspeeds of the FX8350.

Its the same here:

http://images.anandtech.com/graphs/graph9320/74991.png

74991.png


Once it hits 4 cores it does not really scale well and does not use HT- the game is also very sensitive to cache too. The Broadwell Core i5 has a load of L4 cache and does not boost veru highly too.

The top 3 CPUs have either a ton of L3 or L4 cache.

fc_proz.jpg


The Core i5 4670K>>Core i7 4770K. FX8350 is slightly faster than a FX4300 which again is a clockspeed difference. Core i7 5960X also has more cache,which helps performance.

CPU_w_600.png


The FX9590 is clocked 1GHZ higher than the FX6300.

The A12 9800 is TDP and cache limited(only 2MB of L2 cache)- you can see this with the Anandtech review of the X4 845:

http://www.anandtech.com/show/10436/amd-carrizo-tested-generational-deep-dive-athlon-x4-845/8

It uses the same Excavator core as the A12 9800,and it actually regresses in gaming benchmarks compared to the previous CPUs at the same clockspeed.

In reality looking at Cinebench scores single core performance is quite close.

Also,in 4 of the 6 games tested,a Core i3 4330 bests an FX8350. Games like ARMA 3 are the poster child for games which need two to three fast cores and a decent amount of cache. Yet in any game like BF1 and Watch Dogs 2 which do scale well to more cores,the FX8350 is closer to a Skylake Core i5.

b1_proz_12.png


wd2_proz_2.png


That is what games which scale well beyond 4 cores look like - an FX8350 destroying a Core i3.

But,apparently in your world games in which an FX8350 looses to a Core i3 4330 use above 4 cores - riiiiiight!

Now,I know you on purpose misrepresented my post to say I said all games,I said most - 4 of the 6 games have an FX8350 being destroyed by a Core i3 4330,and the FX8350 being tied over its FX4000 cousins or doing slightly better due to higher clockspeeds.

You can massage it all you want - the Core i3 4330 dual core is faster.

Now,BF4 does actually does use more than 4 cores:

bf4_proz_2.jpg


That will easily help out the 8 and 16 thread CPUs and push the advantage to them.

You see all that pointless maths,when it only takes one game out of the six to skew the scores slightly and all for misrepresenting what I said.

Sometimes look at the games tested first before wasting your time - look at what is actually being tested.

The last game is The Witcher 3:

witcher3_1080p_cpu-test.jpg


r_600x450.png


The last graph is with a GTX1080. The game is very GPU limited.

The games is massively bottlenecked if you have less than 4 fast cores,but once you reach that extra cores do not make much difference.

Out of the six games tested,4 are very single core performance limited . They also scale poorly with HT. The Witcher 3 is mostly graphics limited as a game once you hit 4 fast cores.

I know some of you are not happy with the results,but the only way the 8C/16T Ryzen chip can get within 10% of a Core i7 6900K whilst being 10% lower clocked - is by having similar single core performance in 4 of the 6 games tested,which don't scale well with SMT and are more lightly threaded.

The whole reason,why some want the games to be all massively threaded(like your mate),is to twist the results so to push AMD has not having hit BW-E IPC in gaming and this why many of you just conveniently ignore the Core i7 6900K results and try and only talk about the Core i7 6700K or one of the Skylake Core i5.

It does not work that way.

But even if all of those games scaled well to more cores,an 8C/16T Ryzen CPU is so close to a higher clocked Core i7 6900K it pretty much means per thread performance is very close,ie,IPC would be very close.

Now,we could argue AMD has better SMT than Intel which would change things,and push AMD IPC estimates down.

However,since 4 of those games tested,don't really use SMT well,it can only mean that AMD core performance and IPC level is close to BW-E level.

So at this point - you and your mates have not really proved anything.

I expect you will ignore all this and make something else up to try or something,but I have been long enough lurking on forums to remember the same response when CPC leaked the Athlon 64 results,etc.

Anyway have a happy festive holiday,I expect it will be another 100 pages of people trying to make the results look as negative as possible for AMD. I have wasted too much time with people who seem annoyed if AMD can have a CPU with BW-E level IPC.

I cannot fathom the whole point of the negativity by some - anybody would want AMD to do well,so it leads to more competition and better pricing and products for everybody.

Even Jay2cents in his latest video about Ryzen said we need it to be good for the sake of all enthusiasts.

I think some of you younger enthusiasts have forgotten how things were back 15 years ago,when we had both companies pushing each other.
 
Last edited:

jelome1989

Junior Member
Jan 30, 2010
24
2
71
You are quite correct thinking a Starcraft 2 benchmark as a good gaming cpu benchmark, but during my 2500k/860/Q9550 testing on the GTX 970, I have found some others that are even worse.

https://forums.anandtech.com/thread...rks-bonus-i7-860.2428699/page-2#post-37391692

Actually Diablo 3 seems to be the leader, since it maintained a very high delta for both the Q9550 and the 860. Thief is also a good candidate but it seems that newer cpus are mostly OKish.Unreal Engine 4 although multhithreaded, is very demanding on the cpu too.

Mafia 2's benchmarks with hardware accelerated Physx, does a number on the cpus too. And of course let's not forget GTA IV. I recently tested it on my GTX 1070 and I only got 59fps from 53fps with the 970, with very low GPU usage of course, indicating a huge cpu limit. This game is a timeless cpu test. Haven't tested any of the other top cpu limiters however to see what's what.

If Rysen is any good however, I will be testing them myself anyway because I will be having one, thank you very much! :D
Wow that is a hihgly interesting and impressive piece of work. Thanks, looking forward to it! And since you now have a 1070, if you wouldn't mind, try adding Arkham Knight into the mix. That game is more CPU intensive than what most people think it was.
 
  • Like
Reactions: psolord

Lebdnil

Junior Member
Jul 13, 2016
11
8
81
No,you are claiming that I claim that,by lying through your teeth.

I said MOST(not ALL) of those games tested,ie,4 of those 6 games tested,don't scale well with more cores and threads. They are far more sensitive to single core performance,cache and don't really make any real use of above 4 cores.

But you would know that if you knew anyone who played those games and conveniently editing out those other charts is not helping you.

You are trying to help your mate out by trying to sound clever.The damage limitation is getting serious now!! Basically ignore all evidence about the games being tested and the AMD CPUs being tested.

Ok, I try to not drag into this much, more so when you don't reply about anything I post, including THIS:
but also think this review points at Zen reaching an IPC just like Broadwell or a bit behind (rather behind in games, unless Zen scales much better with frequency)

But you really do a by-the-book ad-hominem post, so it wouldn't be fair to leave as is. Yeah, I edited your post leaving out those wall of images, but just for the sake of readability; then, you repeat again that. Well, my failure.

However, I fail to understand why those wall of images are relevant: you add a Arma2 bench (Arma3 used in the review), Watch Dogs 2 (not used) and Battlefield 1 (also not used). Not to mention those Anno 2070 result using Win7 SP1+ and a GTX580, which I'm pretty sure that magazine don't use in their Zen benchmarks. Yes, yes, I don't need a third reply with the same wall of images: there are to prove than those games don't scale "MOST" beyond four threads. The problem I see, well, those only review just publish ONE number for all the games: probably a geometric mean or something along the lines: yes, I don't know that publication before this, so I don't know their methods. So you could just point me that methodology instead of those ad-hominem attacks
by lying through your teeth

But you would know that if you knew anyone who played those games and conveniently editing out those other charts is not helping you

You are trying to help your mate out by trying to sound clever

So at this point - you and your mates have not really proved anything

I think some of you younger enthusiasts have forgotten how things were back 15 years ago

Beautiful. Btw, my first computer was a Pentium 100 (circa 1995, so you couldn't think I just buy it third hand or something), so I know what was happening 15 years ago, "grandpa". But I digress: my main and only point was than clock scaling alone could NOT explain Core i7-6700K result; taking into account than the only two differences between Core i7-6700K and their Skylake i5 sibblings are HT and clocks, if clocks alone don't explain that result, HT should have a notable impact. Unfortunately, we (and I mean all "we" in the community) don't have a second data point for Zen, either with less cores (or more, if you prefer) or at a different frequency; having so, we could calibrate frequency/core count scaling and have a much better grasp of the IPC, scaling and so on for Zen.

We just have one data point, whereas Zen do really much better than Piledriver, so within AMD Zen is really a superb step forward, but as you yourself say:
8C/16T Ryzen chip can get within 10% of a Core i7 6900K whilst being 10% lower clocked

I think I have proved (enlighten me if not) that those games as-a-whole (because we only have those lonely number provided by that Zen review, summing up performance on all games tested) don't scale very well with frequency: so, Zen doing 10% lower than i7-6900K clocked 10% lower doesn't mean Zen has reached the same IPC as Broadwell-E, unless (as I also said myself in my post, you could have read it a bit more calm and reach that part) Zen scales much better than Skylake with frequency, thing that we really don't know. At least, I really don't know, probably is best not to generalize in this.

Again, you provide just another datapoint validating my theory:
The FX9590 is clocked 1GHZ higher than the FX6300
(and gain very little, allow me to add). So pure linear frequency scaling is NOT the norm. Then, I have not used those bench you pasted because I sincerely doubt they use exactly the same settings than that review of Zen, so I don't think that results could be extrapolated as they are to those in the review. Feel free to prove me wrong extrapolating one of the results in the Zen review (not Zen, whatever other known CPU there) with your bunch of benchmarks; I promise I would be really impressed and acknowledge so in public. But until that, please allow me to be a bit sceptic about perfect clock scaling of Zen architecture (10% gain in those average of games for 10% more frequency, as you claim).

And yes, merry Christmas for all ;)
 

dfk7677

Member
Sep 6, 2007
64
21
81
No review about gaming performance of a CPU is valid in my eyes, if the CPU isn't 100% and the GPU lower. That is why many reviews fail, being GPU bottlenecked. If you are testing the CPU you should lower the settings as low as you have to, so as to get the CPU bottlenecking the system.

Games are going to follow the lead of BF1 (and apparently, Watch Dogs 2, GTA VI) and make use of more cores/threads. This is inevitable as CPU manufacturers seem to be having a hard time to increase single thread IPC much.

In BF1 i7s gain ~50% clock for clock over i5s due to HT.
 
  • Like
Reactions: KTE

mikk

Diamond Member
May 15, 2012
4,111
2,105
136
I expect you will ignore all this and make something else up to try or something,but I have been long enough lurking on forums to remember the same response when CPC leaked the Athlon 64 results,etc.


Your problem is that you have no clue about anything what you write, your analysis won't add anything useful because it is so flawed. You don't even consider GPU limits and that it makes any scaling above 4 cores very hard. You also don't take in mind that the french magazine could have tested in a more professional way, using a faster GPU and may be turned AA Off or even going down with the resolution so that the CPU scaling isn't hardly bottlenecked by the GPU. Even the benchmark sequence could make differences to the scaling. As we numerous times said even a 10% scaling above 4 cores will effect the gaming index, but you hardly won't see it in any of your links. And there is definitely a scaling above 4 cores just by looking at the results, your wall of text cannot disprove anything int this regard. It is only a waste.
 
  • Like
Reactions: Sweepr

superstition

Platinum Member
Feb 2, 2008
2,219
221
101
Yesterday
This is going to be my last post on the matter, im not wasting anymore time.
Today
There is no false dilemma, thats just you failing to see the forest through the trees.
Next time either mean what you say or don't post flaming posturing things like promises you're not going to continue a debate that you chose to turn into flaming.

I am not going to waste anymore of my time and you can count on that.
 

majord

Senior member
Jul 26, 2015
433
523
136
Let's use your logic with data shown in the review. You claim that games don't scale beyond 4 threads (just ignoring for a bit than FX8370 does 73.6% in games vs 69.8% for A12-9800 and you take that for "tied"). Very well, just here we have just 3 datapoints: Core i5-6400 (3.1GHz all core turbo), Core i5-6500 (3.3GHz all core turbo) and Core i5-6600. All of them have exactly the same architecture, probably same mainboard and so on, so they just differ in frequency. Presto, some Excel (calc, btw) and you get a beautiful linear regression to correlate frequency with the score in games (don't adjust well, so here you have your first tip than games don't scale linearly with frequency).

1- First, just linear regression: [Freq in MHz] = 66.72972x[Score] - 2994.1337. Now put "98.6" in score and you get 3598MHz as frequency needed to reach that score. Ignoring for a bit than zero score points to -2994MHz frequency ;-), let's input 118.2, which is the score of the Core i7-6700K. You get 4892MHz, so, assuming games don't scale beyond 4 threads, this really proves than Sylake core in Core i7-6700K has a 16% higher IPC (4892/4200) than Skylake core in Core i5 family. Right, sure.

2- Well, as that linear regression don't cross 0,0, let's use a second order polynomial regression: [Freq in MHz] = 0.33466x[Score]^2 + 3.3795x[Score]. Again, let's check with "98.6" and this time gets 3587MHz as the frequency needed for that score. And now using "118.2" we get 5075MHz as the frequency needed to reach that score; this time, with a regression that goes to zero frequency when score reach zero, Skylake core in Core i7-6700K has, amazingly, 21% higher IPC than Skylake core in Core i5 series.

So, sorry but games scale beyond 4 threads, as just you could verify using Skylake cores only. Now a small disclaimer: just because Skylake core scale more or less with those regressions don't mean than OTHER cores scales the same. We don't have more datapoints beyond Skylake (A12 and FX8370 use different cores and differ in core count, i7-6800K and i7-6900K differ in core count, and there are just one entry for Zen and Haswell quadcore), so we couldn't know how every other core scales in those games. And yes, I do think Zen has made a huge leap in IPC vs Con cores (just compare to FX8370), but also think this review points at Zen reaching an IPC just like Broadwell or a bit behind (rather behind in games, unless Zen scales much better with frequency).

Byes

A couple of things.

1. You should really be using the 6600K vs 6700K for the calculations regarding HT yield, as they're both 91w, hence why the former performs slightly better.

2. That aside yes, the 6700k's score is high enough that you can conclude at least some games (or maybe even only one) is scallng beyond 4 threads.. but it's not by much. ~10% by your own math (Adjusted to use the 6600K) , and both this point, and my point is backed up by the fact Intel's own BW-E cpu's at similar clock to the i5's are beating them, decpite an IPC deficit... by a little. So that's fine, but it's minor, and messy.

So the only reasonable approach to take is ignore the Skylake results, and focus on Zen 8 core vs BW-E 8 core. Then you can take the Frequency scaling data from the i5's , (since as you point out they're quite useful being all the same uarch and core count). and apply it to this comparison, which is approximatly 50% according to the tests, which is handy..

We now know for a fact Ryzen is launching at 8% higher base clock than this ES.. so with that scaling, we should see 4% higher performance in gaming:

97.4 * 1.04 =

101.2.

107.4 (6900K) / 101.2

= 1.06

BW-E = 6% faster clk/clk

so this is Quote: rather behind in games To you?

Ironically, if you apply rhe same logic to applications (apply frequency scaling data of i5's to calculate IPC) , you'll see gaming is actually a strong point.

I will point out one flaw in my own logic here.. 8% higher base clock may not translate into an 8% higher max boost clock, (with a knock-on effect at intermediate clocks) but we're looking at single digit figures here even if you assume the worst.
 

Lebdnil

Junior Member
Jul 13, 2016
11
8
81
Thanks. It has been refreshing to get a proper response, majord

To expand a bit further:

1- I've not used Core i5-6600K because that pesky "multicore enhancement". I have a i5-6600 myself and I know it clocks at just 3.6GHz under all-core-load (clock ratios hardcoded in non-K Skylake chips), while K-series could go to single core turbo even under all cores loaded, depending on that BIOS/UEFI setting. My Asrock (Z170Extreme4) defaults to "multicore enhancement" enabled, but do nothing in my non-K Skylake ;) So, Core i5-6600K could score better than i5-6600 because has more power headroom... or because is running at 3.9GHz. Thats why I skipped it.

2- Yeah, as we only know the total score in gaming, we couldn't discern in what way each game contributes to that global score. Then, I'm not sure you could extrapolate Skylake scaling to other architectures, but yes, as a first approximation it should work. So let's say that Ryzen get 6+/-2% less IPC than BW-E (taking into account differences between architectures and differences in max boost clock); yep, as is, it's really close, but then, recent architecture scaling from Intel has been less than stellar: a 6% difference is almost the jump between Sandy and Ivy Bridge, or Haswell vs Skylake. Therefore my "rather behind".

Thats all. I hope we get a full review soon; Intel's new cores have been mostly incremental advances, so it's really refreshing to look at a new architecture that also looks really promising (Bulldozer was truly a new architecture, but also a flop).
 
Status
Not open for further replies.