Crysis 3 CPU scaling, hyperthreading vs AMD cores

[BoilerGamer]

Hey Guys, Just saw a few Crysis 3 CPU performance charts online:





Looks like this is a very mutli-thread centric game and also seem to favor AMD cores over Intel HT(8350 beating 3770K). I wonder if the game is not as well coded for HT as it is for AMD cores. Otherwise Intel need to considering bringing hexacore down to mainstream to maintain its gaming edge in the future, assuming this game is a trend and game developers favor AMD cores due to AMD octa core consoles.

[Firestarter]

note thate pcgaminghardware.de used a Titan (GK110) for the benchmarks

[Arclight]

note thate pcgaminghardware.de used a Titan (GK110) for the benchmarks

Did they also travell into the future cause the date says 21 February 2013....if it's indeed from a .de website then it's 10:10 pm at the time of writing
http://wwp.greenwichmeantime.com/time-z ... n/denmark/

Edit:
Review of Crysis 3 single player

[JohnC]

Eh... Hyperthreading was always giving a very negligible performance increases... Anyawy, it would be interesting to see the actual core utilization screenshots, not just fps numbers.

[Firestarter]

Eh... Hyperthreading was always giving a very negligible performance increases... Anyawy, it would be interesting to see the actual core utilization screenshots, not just fps numbers.

it would be interesting but not very relevant right? Unless it shows that the game doesn't not really use the extra threads but somehow actually has better per-thread performance on the AMD CPUs

[BoilerGamer]

Here is some core load graph

[flip-mode]

Oh my God it's another "wait, FX really isn't so bad" thread. Can we merge with the other one?

[BoilerGamer]

Oh my God it's another "wait, FX really isn't so bad" thread. Can we merge with the other one?

It is actually about Crysis 3 being the first game showing FX-8350 surpassing 3770K. It is about whether this game is poorly coded for hyperthreading(seeing how you don't get as much scaling from i5 to i7 as you do with FX-6 to FX-8) or maybe Intel need more physical cores on mainstream CPUs to keep up.

[JohnC]

Very impressive core utilization. I wish more games would take similar advantage of multithreading... As for AMD's performance in particular - it's not very surprising, AMD worked very closely with Crytek to optimize this game better for their hardware (both CPU and GPU)

[Krogoth]

It looks like Crysis uses three threads at most with the forth thread picking up the what little remains. Cache and memory bandwidth seem to have a greater impact here. Which what allows 3770K and 39xx units to pull ahead by a small amount since clockspeeds are the similar to the lesser SB-IB chips. 8350's larger L2 cache and decent SSE performance helps it out as well.

However, the GPU is still king here and it looks like current generation of GPUs are capable of handling 2Megapxiel gaming w/o AA.

Crysis 3 isn't going to be the system killer that fanboys rave it to be.

[Airmantharp]

Very impressive core utilization. I wish more games would take similar advantage of multithreading... As for AMD's performance in particular - it's not very surprising, AMD worked very closely with Crytek to optimize this game better for their hardware (both CPU and GPU)

Which is a good thing- from a high level perspective, the Piledriver core should actually be faster in fully threaded, homogeneous scenarios that do not allow for efficient hyper-threading, given that hyper-threading generally relies on running two heterogeneous workloads on one physical core. So while we see some use of hyper-threading on the Intel CPUs, what we're probably observing with Crysis 3 is that only a small portion of the total workload is of mixed execution types, and what Crytek has been able to thread out is largely the same type of work.

[BoilerGamer]

Very impressive core utilization. I wish more games would take similar advantage of multithreading... As for AMD's performance in particular - it's not very surprising, AMD worked very closely with Crytek to optimize this game better for their hardware (both CPU and GPU)

Which is a good thing- from a high level perspective, the Piledriver core should actually be faster in fully threaded, homogeneous scenarios that do not allow for efficient hyper-threading, given that hyper-threading generally relies on running two heterogeneous workloads on one physical core. So while we see some use of hyper-threading on the Intel CPUs, what we're probably observing with Crysis 3 is that only a small portion of the total workload is of mixed execution types, and what Crytek has been able to thread out is largely the same type of work.

Imo If more game follow this route(could happen with AMD 8 core Jaguar units in next gen consoles), Intel might have to add more physical cores to their mainstream platforms as HT Quads might be losing in multithreaded games against AMD octas.

[JohnC]

Just read the rest of GameGPU's tests... Looks like my puny GTX680 can't handle it anymore

I guess it's time for a Titan

[flip-mode]

Imo If more game follow this route(could happen with AMD 8 core Jaguar units in next gen consoles), Intel might have to add more physical cores to their mainstream platforms as HT Quads might not be losing in multithreaded games against AMD octas.

All that's necessary is for Intel to merely bump the turbo clockspeed for all cores. But Haswell will launch is just a handful of months, so this is nearly an Ivy Bridge postmortem.

[Chrispy_]

It's nice to see, regardless of an Intel vs AMD discussion, that games are finally taking advantage of more than two or three threads.

I am pleased for AMD too, since Intel has become complacent about increasing thread count, banking on its superior IPC for what feels like far too long

[Ryu Connor]

It's nice to see, regardless of an Intel vs AMD discussion, that games are finally taking advantage of more than two or three threads.

I am pleased for AMD too, since Intel has become complacent about increasing thread count, banking on its superior IPC for what feels like far too long

Intel has an enthusiast answer for high thread counts and based on the benchmarks it does beat the 8350. Consequently that answer doesn't tank on low thread counts either.

[flip-mode]

I am pleased for AMD too, since Intel has become complacent about increasing thread count, banking on its superior IPC for what feels like far too long

Someone will likely strongly disagree with this, but increasing thread count has sharply diminishing returns for desktop users once you hit 4 threads. Beyond 8 thread, it's a near vertical drop. I'm glad Intel (and hopefully AMD has reached this point too) isn't focusing on adding additional thread width that's going to be essentially useless.

[DPete27]

I'm glad Intel isn't focusing on adding additional thread width that's going to be essentially useless.

In fact Intel is trying to make multi-threading more efficient using TSX. If that all pans out, an i7 Haswell CPU for a new build might not be as much of a waste as it is now.

[BoilerGamer]

I'm glad Intel isn't focusing on adding additional thread width that's going to be essentially useless.

In fact Intel is trying to make multi-threading more efficient using TSX. If that all pans out, an i7 Haswell CPU for a new build might not be as much of a waste as it is now.

That would certainly be an encouraging sign, if game like this use more cores(next gen Console's Octacore jaguars will push this direction too), Intel either need to improve HT or increase core count on mainstream platforms to stay ahead of Steamroller. I really want them to do both but if they can get away with milking HT to stay ahead of AMD octas they probably will keep shipping CPU with 1/2 die size of AMD CPUs( Ivy 160mm^2, Vieshera 315mm^2) and enjoy the hefty profit margin edge over AMD.

Am I the only one that think Ivy-E will have a Octacore desktop(Intel bumped Core count last dieshrink Bloomfield -> Gulftown ) ? So Maybe they won't be chasing low TDP on the high end market, those E7 Westmere-EP 10 core monsters need replacements and having a 10 core native Ivy-EP would be the logical answer.

[Bensam123]

Oh my God it's another "wait, FX really isn't so bad" thread. Can we merge with the other one?

Why? If the results are relevant and the methodology is sound, why wouldn't you want to know about results that vastly change the overall performance picture for CPUs? Or at the very least make you look at things in a different light then it's been cast in. It's almost like you're intentionally trying to control the flow of information with a ulterior motive.

I am pleased for AMD too, since Intel has become complacent about increasing thread count, banking on its superior IPC for what feels like far too long

Someone will likely strongly disagree with this, but increasing thread count has sharply diminishing returns for desktop users once you hit 4 threads. Beyond 8 thread, it's a near vertical drop. I'm glad Intel (and hopefully AMD has reached this point too) isn't focusing on adding additional thread width that's going to be essentially useless.

I'm sure the same thing could've been said when quad cores first came out and is a moot point. People can't buy what they don't have access too. Software is generally developed around hardware and if the hardware isn't available, software wont be developed for it. A hexa core CPU doesn't count as having six core CPUs readily available to consumers either.

[Saber Cherry]

Someone will likely strongly disagree with this, but increasing thread count has sharply diminishing returns for desktop users once you hit 4 threads. Beyond 8 thread, it's a near vertical drop. I'm glad Intel (and hopefully AMD has reached this point too) isn't focusing on adding additional thread width that's going to be essentially useless.

I strongly disagree, because your claim is false. Why would you say something like that?

There are always scenarios where more stuff does not yield more quality. For example, a 400 page book is worse than a 100 page book, if the author is terrible; a 500 horsepower engine is worse than a 100 horsepower engine, if you have a tiny gas tank or a light car with slick tires. Throwing 4x as much sugar into a good cookie recipe will not magically yield better cookies. Similarly, a crappy programmer will be unable to gain any advantage from multiple threads. That does not mean there's some magical wall, such that 400 page books, 500 horsepower engines, 50% sugar recipes (like fudge), or 8-core computers fundamentally cannot surpass their smaller siblings.

I've written multithreaded code (for graph smoothing) that showed a ~7.5x speedup on a 4-core hyperthreaded computer, going from 1 thread to 8. That was anomalous (it was floating-point, and mainly I write integer code) and I'm just mentioning it to point out that hyperthreading actually can be useful; otherwise, I've never seen more than a 40% speedup from hyperthreading, and <20% is more typical for my programs. But setting hyperthreading aside, most of my programs scale at about 0.8X the number of cores, up to 80 cores (which is our biggest single node, so the largest I've been able to test). The lines are pretty flat; there's no "vertical drop" after 4. And these programs don't deal with 64k datasets that fit in a cache (which is the perfect setting for scalability) - they deal with hundreds of gigs of data being randomly accessed, which is pretty much a worst-case scenario, since adding cores does not improve the memory bottleneck.

I use other people's programs, too. And the multithreaded ones usually scale like mine do - pretty much linearly, though some start to drop off a bit around 16 cores. If programmers are forced to write scalable multithreaded code (which they are in my field), basically, the incompetent programmers drop out and the remaining people write scalable multithreaded code that is not limited by some imaginary 4-core wall.

[jihadjoe]

Any benchmarks with the old i7-980x?
This looks very interesting! Finally a game with good multicore scaling.

[snakeplissken13]

Link to FX-8350 Crysis tests

http://pclab.pl/art52489-9.html

[Deanjo]

Someone will likely strongly disagree with this, but increasing thread count has sharply diminishing returns for desktop users once you hit 4 threads. Beyond 8 thread, it's a near vertical drop. I'm glad Intel (and hopefully AMD has reached this point too) isn't focusing on adding additional thread width that's going to be essentially useless.

To the hotmail, pintrest, facebag crowd sure, but for a true power user (no, being a gamer does not automatically classify you as a "power-user") those threads are constantly hammered. My financial analysis code makes 32 core systems cry "Mommy!"

[Captain Ned]

To the hotmail, pintrest, facebag crowd sure, but for a true power user (no, being a gamer does not automatically classify you as a "power-user") those threads are constantly hammered. My financial analysis code makes 32 core systems cry "Mommy!"

There is always the exception that proves the rule. Financial analysis, weather forecasting, Big Bang simulations (or development), etc will always want more CPU. Limit flip-mode's statement to the 99.99% of users and he's far more right than wrong.

[vargis14]

I would like to see the sandy and ivy 2600k/3770k cpus benched at 4.3-4.6ghz since that is the speed most are run at...at least match the 8350's 4.0ghz.
Also if tests are done on win 7 .."i know a broken record" core parking should be disabled. When enabled perhaps crysis is like win RAR and ignores HT cores when 3 or more cores are detected.

[Deanjo]

To the hotmail, pintrest, facebag crowd sure, but for a true power user (no, being a gamer does not automatically classify you as a "power-user") those threads are constantly hammered. My financial analysis code makes 32 core systems cry "Mommy!"

There is always the exception that proves the rule. Financial analysis, weather forecasting, Big Bang simulations (or development), etc will always want more CPU. Limit flip-mode's statement to the 99.99% of users and he's far more right than wrong.

99.9% of users could make do with a single core Athlon XP.

[ronch]

Oh my God it's another "wait, FX really isn't so bad" thread. Can we merge with the other one?

You mean, MY thread where I posted a video about this guy who talks about FX-8350 beating several popular Intel chips such as the 3570K, 3770K and 3820?

[flip-mode]

99.9% of users could make do with a single core Athlon XP.

Wow, I disagree completely. All desktop users can benefit from at least two cores. We're not talking about "making do" either. Why are you being difficult? I feel like you either lost the context of the conversation or you're intentionally being difficult.

[Deanjo]

99.9% of users could make do with a single core Athlon XP.

Wow, I disagree completely. All desktop users can benefit from at least two cores. We're not talking about "making do" either. Why are you being difficult? I feel like you either lost the context of the conversation or you're intentionally being difficult.

Considering how the masses are making do with tablets and smartphones now for their computing needs I stand by my statement.