AMD CPU questions (gaming, Phenom II, fps limited by cache?)

[NieNie]

Originally this was written for the SteamVR post but as it is very CPU specific and requires a response I decided to give it its own post here. TL;DR, is my Phenom II cache limiting my FPS or is there another cause?

I have done a few tests for SteamVR. My system is a Phenom II X6 1090T (3.2GHz, 3.6GHz turbo, no overclock), AMD Fury, and various amounts and speeds of RAM depending on the test.

So the first two tests with 8gb RAM running 9-9-9-24 @ 1333MHz, 6.4. The second test(s) run 4gb and 8gb RAM at the same timing but 1600MHz, 7.0. Finally a test with 8gb RAM with the same timings @ approx 1450MHz and a 10% FSB overclock of the CPU and a score of 6.7.

This presents me with a linear relationship between RAM speeds and VR score where CPU speed makes no difference. As the CPU is running at about 80% and drops its multiplier to approx 2GHz no difference from a CPU OC is to be expected. However 1 stick of 4gb ram vs a matched pair of the same didn't produce any difference which surprised me. Additionally the 20% RAM speed improvement gave a 20% improvement in most game's framerate independent of graphics settings.

The next set of tests, 16gb RAM 9-11-11-25 @ 2000MHz, FSB 250 (25% OC), CPU multipliers dropped to the nearest I could for default clock speeds without going over (approx. 50-100MHz under default). 7.7 VR score. Once again the CPU sat around 2GHz and wasn't fully utilized. The final test was 16gb RAM 6-7-7-15 @ 1600MHz, default FSB and CPU clock. 7.3 VR score. FSB 290 was reached but I don't have sufficient cooling and the rapid rise in CPU temperature in BIOS made me abort the OC.

Firstly this shows that the Phenom II whilst being far below the hard requirement for any games made for a long time can run VR with 0 frames under the 90 FPS level and is fully capable of running every game I have tried. Now my interpretation of the results is that the Phenom II with its lower cache uses RAM as a crutch instead hence the FPS directly related to RAM speeds (throughput more important than latency).

I'd appreciate any input, I don't tend to overclock often and don't have a huge understanding beyond whats required to play games and put hardware together.

 

[pandem0nium]

I'd expect better results if you overclock the cpu-nb as well. It gives a real boost to the phenoms unlike the later cpus.

 

[NieNie]

Awesome I'll give it a go thanks. On the FSB OC i had the nb overclocked but if I can get something of that run's improvement without having to overclock the rest I'll feel more comfortable with leaving it on 24/7.

 

[joeyjojo]

It's hard to follow all the results so to summarise:

So the first two tests with 8gb RAM running 9-9-9-24 @ 1333MHz, 6.4. The second test(s) run 4gb and 8gb RAM at the same timing but 1600MHz, 7.0. Finally a test with 8gb RAM with the same timings @ approx 1450MHz and a 10% FSB overclock of the CPU and a score of 6.7.

This presents me with a linear relationship between RAM speeds and VR score where CPU speed makes no difference.

You've found "VR score" is 1. independent of CPU speed and 2. goes linearly with RAM bandwidth.

The next set of tests, 16gb RAM 9-11-11-25 @ 2000MHz, FSB 250 (25% OC), CPU multipliers dropped to the nearest I could for default clock speeds without going over (approx. 50-100MHz under default). 7.7 VR score. Once again the CPU sat around 2GHz and wasn't fully utilized. The final test was 16gb RAM 6-7-7-15 @ 1600MHz, default FSB and CPU clock. 7.3 VR score. FSB 290 was reached but I don't have sufficient cooling and the rapid rise in CPU temperature in BIOS made me abort the OC.

I can't remember the Phenom II architecture well enough but IIRC increasing the FSB will also increase memory bandwidth, but not cache bandwidth, so I think perhaps the thread title is incorrect.

In any case we can see the VR score is always proportional to memory bandwidth (with maybe a small FSB component).

If I were you I'd go back to all these setups and record the cache speeds and memory bandwidth (I think you can get these numbers quite quickly from memtest 86 but something like sisoft sandra would measure it) and look for a relationship between VR score and those.

 

[pandem0nium]

Awesome I'll give it a go thanks. On the FSB OC i had the nb overclocked but if I can get something of that run's improvement without having to overclock the rest I'll feel more comfortable with leaving it on 24/7.

Great stuff. May be able to get it up to 2900 if you can get 290 bus. I'd leave ht as close to stock as you can.

 

[NieNie]

My guess is that as no games have supported the Phenom II architecture for some time they are expecting my CPU to have a larger amount of cache like newer CPUs all do. I think that it is using my RAM where it should be using the cache and both saturating the bandwidth there and leaving a near constant 50% CPU utilisation.

I don't think my CPU cache speed will have made a difference, I can't see that games would be so poorly designed as to having that limitation and also CPU overclocks made no difference whereas faster RAM timings had a slight improvement.

The relationship between VR score is purely based on RAM. In general games like GTA 5 I've done a little testing, 20% faster RAM speed gave exactly 20% FPS improvement. I haven't tested FSB overclocks or anything else in games as I'm uneasy with current cooling solution and the extreme latency timings produced quite a few errors. I'm assuming my GPU isn't being taxed as it should be capable of around 9 and so the results are purely a case of CPU, RAM and their communication. Also I assume that FSB makes no difference, its just the only way I could get my RAM above 1600MHz.

I've tried to isolate everything in each test as much as possible and only the RAM speeds make any noticeable difference and so this is an attempt to explain and understand that.

It's hard to follow all the results so to summarise:
You've found "VR score" is 1. independent of CPU speed and 2. goes linearly with RAM bandwidth.
I can't remember the Phenom II architecture well enough but IIRC increasing the FSB will also increase memory bandwidth, but not cache bandwidth, so I think perhaps the thread title is incorrect.
In any case we can see the VR score is always proportional to memory bandwidth.

 

[joeyjojo]

My guess is that as no games have supported the Phenom II architecture for some time they are expecting my CPU to have a larger amount of cache like newer CPUs all do. I think that it is using my RAM where it should be using the cache and both saturating the bandwidth there and leaving a near constant 50% CPU utilisation.

I don't think so as games don't normally have much cache or memory dependence. Actually requiring lots of memory bandwidth is pretty rare these days in computing in general. I can only think of a couple of times it shows up (IGP gaming and linear algebra solvers).

I don't think my CPU cache speed will have made a difference, I can't see that games would be so poorly designed as to having that limitation and also CPU overclocks made no difference whereas faster RAM timings had a slight improvement.

When the CPU wants something from the cache but it isn't there it goes to the system memory which is called a cache miss. If you're getting a lot of cache misses as you suspect you are, then cache speed and memory latency (tighter timings) will both help.

The relationship between VR score is purely based on RAM. In general games like GTA 5 I've done a little testing, 20% faster RAM speed gave exactly 20% FPS improvement. I haven't tested FSB overclocks or anything else in games as I'm uneasy with current cooling solution and the extreme latency timings produced quite a few errors. I'm assuming my GPU isn't being taxed as it should be capable of around 9 and so the results are purely a case of CPU, RAM and their communication. Also I assume that FSB makes no difference, its just the only way I could get my RAM above 1600MHz.

Now this is interesting. It sounds to me like running VR is introducing a massive memory bottleneck. It would be very nice if you could run the same games on the same system without VR and see if you get the same behaviour (RAM limited, CPU indifferent). Also see if your system uses more memory with VR than not, and ideally which processes are consuming it.

If you're doing any memory tweaking make sure to run some memtest or linpack first to check stability.

Don't assume the FSB makes no difference. IIRC the hypertransport link is used for memory communication, so overclocking that might increase memory bandiwdth.

 

[NieNie]

Now this is interesting. It sounds to me like running VR is introducing a massive memory bottleneck. It would be very nice if you could run the same games on the same system without VR and see if you get the same behaviour (RAM limited, CPU indifferent). Also see if your system uses more memory with VR than not, and ideally which processes are consuming it.

Don't assume the FSB makes no difference. IIRC the hypertransport link is used for memory communication, so overclocking that might increase memory bandiwdth.

Oh its not just VR, its all games. I couldn't measure VR FPS directly but I could benchmark GTA 5, Dark Souls 3, Witcher 3 etc, every single modern game showed the correlation between ram speeds and bandwidth. Timings showed a lesser effect but not as much as frequencies. Increase the frequency by 20% and the in game FPS is exactly 20% better, rinse and repeat. I also experienced no actual stutters as I tend to find with CPU calculation speeds though I haven't been outmatched CPU wise for over a decade so I'm a little outdated there.

Anyway each test I ran I isolated the component as much as I could, CPU clocks (including HT etc, not including ram frequencies) gave a minimal boost whilst ram overclocks (including fsb, 1600 is the max for the phenom 2, cpu multipliers dropped to keep frequencies the same) gave a pure and dramatic boost. I was using 2400 Mhz ram and kept the timings loose in all overclocks to correlate with the old 1600Mhz ram timings (adjusted for frequency changes etc) so I was only really worried about ram errors during the timing change (1600 @ 5-6-6-15) which I did get, but I ran the test quickly and shut down.

But recently my harddrive died so I had to reinstall windows. My own fault, chronic pain is a hell of a thing. Anyway whilst installing windows I swapped out my old phenom II for a 8370 so that ends that episode. RAM pushed up to 2400 Mhz (factory speed) and the performance instantly jumped. I can run a test at 1600 Mhz with the old timings if you're interested in the puzzle but as a problem it disappeared entirely with the new CPU. I still think its cache, mind so 1600Mhz shouldn't make a difference but I can run it anyway if you'd like.

My reasons for assuming cache - firstly as you said ram should not at all saturate to that level in any game yet the tests very strongly suggest it is the sole bottleneck. The phenom II has been behind the absolute minimum specs for a while now with games all asking for fx series or intel (with far more cache) and some even demanding an i7, the main difference being the cache of an i7 and an fx is higher than that of an i5 (number crunching isn't the limitation if a phenom II x6 can handle it at high fps and top graphics without issue).

Overclocking the processor alone, the fsb without a ram overclock or anything but the ram didn't produce results, usually zero difference occasionally negligable difference but nothing like the 50%+ fps boosts of a ram overclock.

Anyway once again I certainly appreciate the help. And as expected I have an equally quirky problem upon installation of the new CPU. I've made another thread for it at:
https://forums.overclockers.co.uk/showthread.php?p=29528647
In essence the POST is inconsistently consistent. It works, the CPU handles games but it refuses to POST... unless I clear the CMOS and redo identical settings.

 

[Uriel]

It's interesting that on occasion we see what appear to be CPU bottlenecks in cross-platform games also on PS4 / Xbox1.

Those 8 core Jaguars don't have much in the way of single-thread number crunching ability. What they do have is very decent memory bandwidth. Perhaps that will be a limiting factor in a few games of that ilk.