Revisiting Broadwell for OC fun

[mackerel]

(note title has been edited to reflect that I will be looking at Broadwell here)

With my main system fitted with the EK loop, I now have a spare H110i GTX and thought I'd use that for some OC fun. I just got another case which it will actually take that cooler. I was going to move my i5-6600k system into it (already on high end air), but then remembered I also have an i5-5675C I never really pushed due to the small cooler I have on it. The "problem" I had with the 5675C was that the stock core clocks were never high enough to stress the L4 cache which is what I really wanted to do... so this seems an ideal opportunity. The possible drawback is, as I originally did this as a micro-ATX build, would the Asus Z97M-plus board be a potential limit? I also have a Gigabyte GA-Z97M-D3H in another system, although also as micro-ATX I'm guessing that wouldn't necessarily be any better.

I think I talked myself into doing Broadwell as Skylake is so common it just doesn't sound as interesting!

 

[EarthDog]

Broadwell is really the stepchild of Intel's chips.. such a short life before skylake released...anyway, choice is yours. We don't have a horse in this race.

As far the boards, both are 4 phased and heatsinked so you should be fine pushing things until the cpu cooling runs out. The form factor of mATX really isn't a problem, there are some pretty robust ones available. But most run out of cpu cooling before it's a problem.

 

[mackerel]

I'm hoping the H110i will do better than any spare air cooler I have (up to Hyper 212 class), so will see where we get with that I really do like the L4 cache as for a lot of tasks it takes ram performance out of the equation, and want to see what I could do with it with sufficient CPU power kicking it. It is even overclockable but it made zero difference when I tried before.

 

[mackerel]

Case transfer kinda done... CPU before had Noctua D9L and prime95 small fft got it up to 70C at 3.5 GHz (not sure voltage). Swapped the H110i GTX in and doing same it is now 59C with a pair of noctua fans on. Not bad.

Right, the ugly part. As always I didn't do my homework. I knew a 280 could fit up top, but it was said the fans had to go outside due to lack of height. I failed to note, there's no cable holes for the fan wires to go in, unless I want to drill some holes on top which I'm not sure I do. Case is a Corsair 230T in orange and it looks so nice I'm debating replacing a bunch of other cases with them... will think about this tomorrow.

 

[mackerel]

Initial overclock quickly tested. 1.25v at 4.0 GHz was prime95 blend stable for 30 minutes, with max temps hitting 71C (I didn't measure ambient to compare against previous).

I've had a bit of a crazy idea for a test and am doing a clean Win10 install on a salvaged 60GB laptop HD right now... what I have in mind, I don't know if it'll work, at least so far I've not managed to find anyone else to have done it...

 

[mackerel]

I think I made a mistake in using that 60GB laptop drive. It is SLOW. Especially as I have a 2TB WD Green idle in another system I could have pulled. Anyway, Win10 installed, Steam installed, 3DMark installed... let the benching commence!

 

[EarthDog]

We use tiny (60GB) ssds.

The green may be a bit faster...

 

[mackerel]

It's just an install for benching, although it is so slow I'm thinking of reverting to my Win7 SSD install while tinkering, and only switch back to Win10 for DX12 benches

Thinking 4.0 GHz at 1.25v wasn't exciting enough, I jumped to 4.2 GHz at 1.30v. Crashed soon after prime95 started. Dropped back to 4.1 GHz still at 1.30v and its survived a few minutes at temps up to 79C. Using Anandtech's review previously as reference, they got 4.2 at 1.325v set, and didn't stabilise 4.3 at 1.4, so I'm probably not far off the wall here.

 

[EarthDog]

I know it's an intstall for benching... that's exactly how we do it.

 

[mackerel]

I could get another spare SSD but the amount of hardware I have at home is getting a bit silly, and I'll never be able to resist putting the SSD into yet another new build, but I'm forcing myself to hold off until Kaby Lake and Zen.

Anyway, running over lunch, I just had 30 minutes prime95 stable at 4.1 GHz 1.30v. It eventually peaked at 81C. Benching that now which will give me a good baseline on how only clocking CPU affects results. Think I'll give it more volts and try for 4.2, before moving onto GPU OC.

 

[mackerel]

Updated mobo bios, with a changelog of simply "improve system stability". While at it, I recalled I had DDR3 2400 ram in another system so I moved that over, replacing the 1600 in there. If I'm going for it, I might as well really go for it. Survived 5 mins of prime95 blend at 4.2 1.35v so far, hitting 87C. I don't think I'll be trying beyond this. Will let it hit 10 minutes then another round of benches to see if the ram might have helped more than the 100 MHz, and after that, GPU overclock...

Anyone want pics?

I'll be doing screenshots for bench submissions later.

Edit: 4.2 unstable. But was it the ram or the CPU since I did both? I'll go back to 4.1...

 

[mackerel]

I think there were enough hints earlier, but I can now confirm my goal was to submit the best Intel GPU result possible for TimeSpy! I'm not quite there yet, but it has been a learning experience.

To recap, the CPU core overclock was pretty simple and I settled on 4.1 GHz at 1.30v. 4.2 wasn't stable at 1.35v and it was getting rather warm already.

Swapping the ram from 1600 CL8 to 2400 CL??? helped a lot, even with the rest unchanged. As much as the L4 eDRAM helps, it is finite in size and once exceeded you're still hitting system ram.

And finally, it was time to overclock the GPU. The bios was no help. I went in, set multiplier, and was left wondering why I was stuck at 650MHz compared to the 1300 set, and 1100 stock. Turns out the base clock was 50MHz, not 100MHz, so I had to double the multiplier. Ok, set to 26 should give me that 1300. Nope. Booted, and still got 1100. After some searching, and there is little information on iGPU OC it seems, it was voltage limiting itself. The bios reported no info on clocks, multipliers or voltages so I was running blind. I set +0.1v offset and partial success. I got 1300, but it glitched immediately on furmark. I know some of you dislike furmark, but it just works. I actually downloaded Heaven first, but it complained of a missing library, and I can't figure out how to fix that. Anyway, after the glitch it seemed to reset itself to 1100. At that point I installed Intel XTU, and finally I could see the voltage. I could also adjust it in software. I have no idea what's a safe GPU voltage, but assume it is similar to the CPU. Even with +0.2v offset it was not stable at 1300, and the temps with the CPU idle were already into the 80's. The iGPU can push out power it seems. I backed off to 1250 and adjusting in 0.025v steps, found I had to run +0.175 to run glitch free.

Where am I at? For now I'm at 714 on TimeSpy, up from 625 at CPU 3500, GPU 1100. Firestrike currently 1901 up from 1672.

These aren't my final results, as I still have two things to try: L4 cache clocks and ram clocks. I think the ram would help for sure, but how much can you push DDR3? The L4 cache is quick to do as I've previously tested it at 2200 up from stock 1800 and seemed stable, just not providing any benefit.

 

[mackerel]

I'm calling it a day now. The TimeSpy result has been submitted, and I got Firestrike up to 1951 with a lucky run at 1300 CPU core. TimeSpy either crashed or when it did run gave a lower score than at 1250. I think there is still something going on which could be improved, but it is one step too far for me. hwinfo64 was showing a flag tripping, prochot for both CPU and GPU. Temperatures of both were fine, not really going above the 60's. Intel XTU didn't report any thermal, power or current throttling.

Oh, the L4 cache, I set to 2200 compared to 1800 stock and... it did nothing. So back to 1800 it was.

I did give a half baked attempt at OC the ram. Still in XMP mode, I set it up a step to 2600 and relaxed CL from 11 to 12, and no boot. Voltage was already 1.65, didn't fancy playing with the rest so left it back at 2400.

I've also decided to remove the watercooler, and put air back in the box. As the case is more spacious than the last one I can put a better air cooler in there with all lids in place. The water cooler will go back into storage for another day. There's one other case it might fit in, if so I might do that.

I also got to get all the system I stole bits from back running again for a CPU based challenge elsewhere.

 

[mackerel]

Before I took the water off, I remembered I had a test to do with the higher CPU clock and L4 cache. In short, with prime95 4096k FFT benchmark 4c/4t, throughput was 227.6 iter/s with stock 3200 L3 and 1800 L4 cache. Overclocking both to 3400 and 2200 respectively, I got 233.6 iter/s, for about 2.6% increased throughput. Not really significant. To try and see if one or the other was responsible, I tried 3400/1800 also, and that was an intermediate 231.6 iter/s. So, I think we can say both contributed, but the contribution was minimal. I still don't think the CPU is stressing enough here, since I worked out previously that Broadwell was ball park 20% slower IPC than Skylake at this type of task. Further attempt to raise L3 to 3600 lead to a system crash.

I swapped the old 1600 ram and found a nice surprise, the system booted with the ram seeming still at 2400 settings from earlier. This might be worth of further investigation, as the 1600 is low voltage CL8 (1.35V), and the 2400 is 1.65V. Maybe they're not that different if you throw enough voltage at it, with the appropriate timings.

Finally, I was going to put a Hyper 212 on it, but then found I had previously ordered a Deepcool Lucifer V2 which was unused. After checking it would actually fit in the case, I went to fit it. I opened out the Intel mount accessory bag next to me and got on with it, until I realised I had 3 spacers and there should be 4. I checked around in case it rolled away or something, but it was not to be found anywhere. Was it missing? I couldn't continue with 3, until I realised they were near enough the same length as the Noctua D9L ones which I took out of this case at the start. They fit, and I completed using it. The Deepcool ones however didn't fit on the Noctua bracket, as the inner hole was smaller diameter. By chance, I think I can still reuse the Noctua without the spacers, since the screws on those are variable width and still seems to hold the upper part stable without. Noctua also came to the rescue a second time, when it came to screwing the heatsink down onto the mount, as my screwdriver wouldn't fit in the recess due to the heatsink body size. The Noctua came with a long thin screwdriver tool which did fit. After all that, I reassembled the case and tested it at 4.1 GHz 1.3V. It topped out around 81C and the fan was rather noisy at that point. Reverting to the original 3.5 GHz at 1.2v, it is stabilising at 69C, so pretty much the same as the Noctua D9L I took out. It does seem quieter though, thanks to the 140mm fan compared to the 92mm of the Noctua.