Personal computing discussed

Moderators: Flying Fox, morphine

 
fleetfeather
Gerbil In Training
Topic Author
Posts: 1
Joined: Wed Aug 02, 2017 1:27 am

Frametimes: which uArch is the fastest?

Wed Aug 02, 2017 1:39 am

Hi all,

First time TR poster, long time reader.

What's the current state of play for those users looking to minimise frame times on high refresh rate displays?

Up until recently, TR's CPU reviews included the 5775C in their round-up, and we could see how the L4 cache was providing benefits to frametimes. However, recent reviews for Intel and AMD CPUs haven't included the 5775C in the round-up, and the standardised test platforms have been updated to include different gaming benchmarks and newer GPUs. All together, these factors make it hard to compare old reviews featuring the 5775C to new ones.

Could anyone weigh-in on how the frametime landscape may have changed since the 5775C, if at all?

Thanks!
 
ultima_trev
Gerbil XP
Posts: 361
Joined: Sat Mar 27, 2010 11:14 am
Contact:

Re: Frametimes: which uArch is the fastest?

Wed Aug 02, 2017 2:27 am

Cache tends to provide for better performance, regardless of architecture. In fact, Broadwell-E is (clock for clock) still the best architecture for games in most scenarios, only beaten in certain instances by Kaby or Skylake-X because of their ability to clock a good bit higher.
Ryzen 7 1800X at 3.9 - Corsair H60i - GA AB350 Gaming - 32GB DDR4 2666 at 14,14,14,34 - Gaming X GTX 1080 - 250GB WD Blue SSD - 2TB Toshiba 7200rpm HDD
 
Airmantharp
Gold subscriber
Emperor Gerbilius I
Posts: 6085
Joined: Fri Oct 15, 2004 10:41 pm

Re: Frametimes: which uArch is the fastest?

Wed Aug 02, 2017 3:17 am

You're hitting on the idea in your OP that both microarchitecture and configuration, that is, clockspeeds, cores, and cache, affect game performance.

Right now it's a tossup, and really depends on clockspeeds, between the top-end Intel CPUs; once the 8700k ships, it'll probably take the overall crown.

[the AMD chips hold their own, but they have neither the IPC nor clockspeeds to takeover the 'fastest' designation; we're also not likely to see another Intel release with a massive separate L4 cache]
Canon 6D|24-105/4L |70-300/4-5.6 IS USM|50/1.4 USM|50/2.5 CM|50/1.8 STM|85/1.8 USM|Samyang/Bower 14/2.8|24/2.8 IS|Sigma 150-600 | C
Canon EOS-M|11-22 IS|22/2|EF-M 18-55
 
DPete27
Grand Gerbil Poohbah
Posts: 3186
Joined: Wed Jan 26, 2011 12:50 pm
Location: Wisconsin, USA

Re: Frametimes: which uArch is the fastest?

Wed Aug 02, 2017 9:12 am

If you're really sensitive to that stuff, it's hard to beat an i7-7700K.
Keep in mind, your GPU will oftentimes have a much more profound effect on frame times than your CPU.
Main: i5-3570K, ASRock Z77 Pro4-M, MSI RX480 8G, 500GB Crucial BX100, 2 TB Samsung EcoGreen F4, 16GB 1600MHz G.Skill @1.25V, EVGA 550-G2, Silverstone PS07B
HTPC: A8-5600K, MSI FM2-A75IA-E53, 4TB Seagate SSHD, 8GB 1866MHz G.Skill, Crosley D-25 Case Mod
 
Chrispy_
Gold subscriber
Maximum Gerbil
Posts: 4106
Joined: Fri Apr 09, 2004 3:49 pm
Location: Europe, most frequently London.

Re: Frametimes: which uArch is the fastest?

Wed Aug 02, 2017 10:40 am

What DPete says: 7700K.

Clock it as fast as you can tolerate.
Get fast DDR4 for it (tight-timing PC4-3200).
Make sure that you're using the optimum configuration (1 DIMM per channel, two-channel operation).

Games like high-bandwidth, low-latency, data crunching on a single "primary" thread. There are plenty of games that are outliers and can make use of large cache sizes, or use parallelism better than most, but realistically you want the best IPC, the best clockspeed and you want that on one thread. The i3-7350K at 5GHz is theoretically ideal, but many games now use more than two threads total, and that starts to hinder the primary worker thread in most games. For this reason, the 7700K has the extra cores to ensure that the primary thread can run as fast as possible. Perhaps there will be a 6-core CoffeeLake K-series CPU that clocks to ~5GHz, but I strongly suspect that the quad-core will still be the sweet spot for games.

If you do that you'll have the current best defense against low minimum framerates on the CPU side of things, but all of that is wasted if you're not using a decent GPU with it.
Congratulations, you've noticed that this year's signature is based on outdated internet memes; CLICK HERE NOW to experience this unforgettable phenomenon. This sentence is just filler and as irrelevant as my signature.
 
synthtel2
Gold subscriber
Gerbil Elite
Posts: 610
Joined: Mon Nov 16, 2015 10:30 am

Re: Frametimes: which uArch is the fastest?

Fri Aug 04, 2017 12:42 am

Chrispy_ wrote:
The i3-7350K at 5GHz is theoretically ideal, but many games now use more than two threads total, and that starts to hinder the primary worker thread in most games.

Just for the record, the G3258 I just ditched was feeling entirely inadequate for gaming these days. If something's using three-ish heavy threads, it's little enough juggling that things probably work decently, but three isn't high.

My G3258 -> 1700 upgrade represents a bit of a downgrade in single-threaded performance (and memory performance until I get that dialled in), and it's still a noticable improvement in pretty much all the real-world stuff I do. The question is just whether it's a marginal improvement or a massive improvement. More games fall into the latter category than you might think.
 
Airmantharp
Gold subscriber
Emperor Gerbilius I
Posts: 6085
Joined: Fri Oct 15, 2004 10:41 pm

Re: Frametimes: which uArch is the fastest?

Fri Aug 04, 2017 12:57 am

That Ryzen at >4C8T is competitive in gaming at all is a testament to the usefulness of more than six hardware cores, given the lower IPC and clock ceilings of AMD's new architecture.

However, we're at the point where few games can run optimally on 2C4T, many at 4C8T, and enough at 6C12T that moving to a 6C12T or larger baseline makes sense so long as the IPC and clockrates are there, and they are for Coffee Lake. They may also be for Ryzen's successor architecture, whenever that gets here.

Far less prevalent are games that will do better with 8C16T or more at lower clockrates.
Canon 6D|24-105/4L |70-300/4-5.6 IS USM|50/1.4 USM|50/2.5 CM|50/1.8 STM|85/1.8 USM|Samyang/Bower 14/2.8|24/2.8 IS|Sigma 150-600 | C
Canon EOS-M|11-22 IS|22/2|EF-M 18-55
 
Chrispy_
Gold subscriber
Maximum Gerbil
Posts: 4106
Joined: Fri Apr 09, 2004 3:49 pm
Location: Europe, most frequently London.

Re: Frametimes: which uArch is the fastest?

Fri Aug 04, 2017 5:58 am

I suspect we'll hit an artificial wall at 8 threads, because that's the configuration the current MS and Sony consoles are running and therefore the configuration that developers of games and engines will be targeting.

At some point, parallelism in games might be good enough that SMT isn't enough to handle the extra threads and so real cores will be needed. However, the resulting clockspeed reduction that comes with increasing core count also offsets the advantage of the extra cores, so at the moment I'm still convinced that the fastest (clock*IPC) 8-thread solution is going to be the answer, whether it's a quad-core with SMT or a true 8-core chip.

Maybe next generation when consoles use even more cores, there will be more games that are designed from the ground up for better parallelism, but at the moment it does seem that 3-4 threads is what the games use, which is why 2T chips like the G3258 feel inadequate and also why there's still a significant improvement moving from 2C/4T to 4C/4T, simply because SMT isn't as good as an extra dedicated core.
Congratulations, you've noticed that this year's signature is based on outdated internet memes; CLICK HERE NOW to experience this unforgettable phenomenon. This sentence is just filler and as irrelevant as my signature.
 
synthtel2
Gold subscriber
Gerbil Elite
Posts: 610
Joined: Mon Nov 16, 2015 10:30 am

Re: Frametimes: which uArch is the fastest?

Fri Aug 04, 2017 6:01 pm

Some games tend towards multithreading by breaking particular tasks off into their own threads (likely as not involving some kind of work pool as an auxillary thing), others have some more granular way of splitting up most of the work (often along these or these lines). This distinction is really not binary and a lot of stuff doesn't fit into it well, but for the sake of explanation I'm going to roll with it for a bit.

In the first case, popular intuition on why things perform as they do is mostly correct. If you've got as many threads as are expected, framerate is probably limited by just one or two of them. The trouble is that the number of expected threads is as likely as not to be six (due to the number of cores consoles leave free for devs). On a 4C8T CPU, this is almost certainly completely fine, because each logical core is still a whole lot faster than a physical 1.6 GHz Jaguar core. On a 4C4T CPU, there's still more than enough raw power to handle the work that needs to be done (for normal framerates), but with some threading architectures and/or load patterns, thread management might start to take a bite out of it. At 2C4T, you've no longer got a big advantage over consoles in raw power unless those two cores are clocked to the moon. At 2C2T, the thread contention can probably more than eat up any advantage in raw power you can get.

As an example of that first kind of threading architecture, see Shadow of Mordor (nothing against it, it's just what I tried to play lately and got data for). It runs three heavy threads full-time (clearly doing different kinds of work than each other) and three more heavy ones in camera movement (those three apparently the same). With little camera movement, the three full-time threads can share a G3258's two hardware threads well enough and everything is pretty great. When I moved the camera 90 degrees, the other three threads stole ~30% each (of 200% total) for a second or so, in which time maybe two or three frames would get rendered, input would get polled maybe two or three times, and input interpolation weirdness would likely as not have moved the camera another 90-180 degrees starting the process over. In other words, it was completely unplayable due to this thread contention. The minspec is an i5-750 (4C4T 2.66 GHz Nehalem) and I had my G3258 clocked to 4.1 for that, so I should have barely been below minspec on raw power, but the thread contention killed it.

The other threading pattern (GCD etc) makes it relatively easy to keep however many cores are handy 80+% busy on something or other, but if not done carefully this architecture can be hell on data access patterns. For instance, in the Naughty Dog presentation they say they're doing about a bajillion fiber switches per frame (I don't remember and am on a really cheesy internet connection at the moment so I don't want to search through the video to find out). Work that should be adjacent is probably getting scheduled across different cores in an interleaved way that's tough for caches to handle in the best of cases, and I'd be very impressed if hardware prefetch can see beyond a fiber switch. I'd bet games like that love a fast, unified, and low-latency L3 cache (or an L4), and I bet that explains a lot of Intel's continued dominance at gaming in titles that would otherwise scale very well to higher core counts.

Another thing that may be particularly important for this is what work is synchronous with framerate. If all the heavy logic is synchronized to framerate and it's a heavy game, it may want a whole lot of power across a lot of threads. If a lot of heavy logic is on a separate tick, framerate is more likely to be able to fly on the strength of one thread while fixed-rate logic that takes many console threads fits in 1.5-2 big cores. Other work may be entirely asynchronous.

It's nearly impossible to test objectively if games don't provide their own tools for it, but personally I'd like to see more focus on the performance of things other than framerate. Games with a slow tick, games that can't stream fast enough to avoid pop-in, networked logic that can't keep up on a slow connection, weirdness that results in high input lag in certain situations.... it's all annoying, CPU power is often the limitation, and often things like that can show up while framerate looks perfectly fine.
 
anotherengineer
Graphmaster Gerbil
Posts: 1473
Joined: Fri Sep 25, 2009 1:53 pm
Location: Northern, ON Canada, Yes I know, Up in the sticks

Re: Frametimes: which uArch is the fastest?

Fri Aug 04, 2017 8:06 pm

Next question should be how many frames do you need? If you are playing on 1080p 60Hz screen, then no point going too overkill due to diminishing returns.
I think with AMDs move to cheaper and available 8 core CPUs to the avg Joe consumer, it might allow or push Devs to start producing more software that can take advantage of it going into the future.

Strictly frame time only
some games are close like crysis
http://techreport.com/review/31724/amd- ... part-one/5

and most other games Intel has the advantage and they also have a clock speed advantage
http://techreport.com/review/31724/amd- ... part-one/8

and for basic FPS vs resolution
https://www.techpowerup.com/reviews/AMD ... 0X/12.html
Life doesn't change after marriage, it changes after children!
 
Topinio
Graphmaster Gerbil
Posts: 1345
Joined: Mon Jan 12, 2015 9:28 am
Location: London

Re: Frametimes: which uArch is the fastest?

Sun Aug 06, 2017 3:19 am

anotherengineer wrote:

Next question could be "what games to choose?"

Per that link, if you have a weak CPU then Civ VI will take a big hit and Doom will run like a champ regardless.

Would be nice to see the same test and charts on an average GPU from each vendor, instead on a GTX 1080.
Desktop: E3-1270 v5, X11SAT-F, 32GB, RX Vega 56, 250GB BX100, 2TB Ultrastar, Xonar DGX, XL2730Z + G2420HDB
HTPC: i5-2500K, DH67GD, 6GB, GT 1030 SC, 250GB BX100, 1.5TB Barracuda, Xonar DX
Laptop: MacBook6,1

Who is online

Users browsing this forum: No registered users and 1 guest