DDR3: how high can you go?

[mackerel]

I saw the Gigabyte Z97-D3H-CF on sale at a low price, and took the opportunity to replace a H81 chipset mobo and get some faster ram in. The CPU is a i5-4570S which turbos to 3.2 GHz with 4 cores active, and the dual channel dual rank 1600 wasn't cutting it. That ram is many years old (I bought it in Sandy Bridge era) so I looked for cheap 2400 ram to put in there. I settled for Corsair Vengeance Pro 2x8GB CMY16GX3M2A2400C11 as the cheapest option I could find. I didn't need 8GB modules, but as stated in a previous thread dual rank performance is a necessity also, and I wasn't sure if the 4GB modules would satisfy there.

I've been stressing it overnight, and it seems fine. The performance boost from going 1600 to 2400 is around 20% which is nice. I've not really tried OC DDR3 before so it will be interesting to see how much further I could take it. If I can hit 3000+ it would be "practically unlimited" ram. I don't expect to do that, but it'll be fun to have a go.

 

[mackerel]

Boots at 2800 11-16-16-36-421-2T just by increasing memory multiplier and leaving timing on auto. Aida64 gives read 32746, write 42057, copy 34272, latency 47.8. It will bench but stress test is failing.

 

[Woomack]

You can try something like 12-13-14 or 12-14-14 ~1.35V but 2600/2666 will be easier to stabilize and there is a chance you can run it at about 11-13-13.

 

[mackerel]

What at 1.35v? The ram defaults (in XMP) to 1.65v. Also, from my past testing, latency is very much secondary to bandwidth. Actually, I'm probably in the diminishing returns area already anyway, as I'm only seeing about 2.5% performance increase going from 2400 to 2800. So if I push harder it will mostly be for fun.

I haven't got used to the Gigabyte bios yet, having only used Asus and MSI so far. Not sure where everything is...

 

[mackerel]

Raised it to 12-16-16-16 and 15 minutes stable in aida64. Next stop: 3000.

 

[mackerel]

No 3000 multiplier, without hitting the bus... is there much to be got for non-K CPUs?

Back to the ram:
2933 12-16-16 gave lower benchmarks than 2800. Bad multiplier? Or is auto increasing something else in background and slowing things down?
2800 12-14-14 unstable but best benchmark results so far. I didn't think before, but do Haswell CPUs behave much differently than Skylake as far as ram behaviour?

 

[Lochekey]

I don't think you can directly compare the 2 as one users ddr3 and the other uses ddr4.

You messy know this but I will state it anyways. When you start pushing ram like this there is a very real possibility of borking an OS. You may want to do your testing on a backup drive until you find something stable.

- - - Updated - - -

By the way what are you doing that you need high performance from your RAM or is this just for fun?

 

[mackerel]

2800 12-15-15 stable for 10 minutes in aida64, and 20 minutes in prime95 so far. Haven't benched it at all yet. OS is sacrificial as it is a cruncher. Tasks are prime number finding, and prime95 is indicative of performance. Large tasks are substantially ram limited. If there are no errors by bed time I'll try it on real tasks overnight. They're double checked anyway so errors are not fatal).

I don't know if it is my imagination, but I think Windows is taking longer to boot with OC ram. Seems to pause longer between the logo and desktop. Guess I can test that by un-overclocking it and see if it goes away. I was tinkering with driver updates last night so could be something there too.

 

[mackerel]

Been running overnight with no known problems. Thing is, when I looked at the work rate, the median unit time was unchanged from 2400, and the mean was actually 3% slower. There seemed to be more variability. I did have aida64 running in the background so don't know if that might have been sucking a few CPU cycles.

It was then I noticed the PC didn't have a video driver installed. It was showing as standard VGA adapter, even though I had installed the intel integrated driver yesterday. I tried again a few times, and still get standard VGA driver no matter what I do. To get around it I threw in an old nvidia card. Now to see if that helps...

 

[mackerel]

Been doing some testing at 2800 with the nvidia card in. Ram bandwidth benchmarks went up some more, but now I'm failing aida64 stability testing. Could whatever the problem was before be slowing something down to make it not a problem?

I'm going to go back to 2400 XMP for now and see if that has changed performance much. Then at some point I might have another go at stabilising 2800.