Superpi 32M

Before starting off the actual competition, each team had plenty of time to assemble the hardware. Due to the fact we didn't assemble the complete CPU and ram cooling unit we got our system to boot up first. Sadly enough, the Gigabyte staff wasn't too impressed with my question for an extra bonus point. The reason not to hook up the CPU fan was done intentionally so we could use it as extra cooling for the graphics card during 3DMark01.

We went straight into the bios and noticed the board had F3 bios flashed. Latest bios available was F5D (also the one we used at home) but it seemed stable when testing the Bclocks. Working our way up, it seemed that this CPU responded far better to volt inputs then the Madshrimps sample. 220 Bclock with PCIe over 125Mhz was no problem. The sample motherboard at home even went up to 150Mhz PCIe , but the gain/mhz ratio was not that impressive.


Belgian team with Massman aka the bios wizard in motion, Leeghoofd supervising and thinking this looks good

Then came our second surprise as it was announced that everyone had to flash to F5D. So we kindly obliged but we couldn't get the same speeds running as with the older bios. Small setback, but to keep the playing field level we didn't even bother to flash back to F3 bios. Problem with the F5D was that we lost the SSD once we passed 125Mhz PCIe speeds.

Initially we booted with low ram speeds just to get a feel for the CPU. Rams were tested next. The 4GB Kingston kit was rated at 2000mhz Cas 9 latency. Trying to boot at Cas 8-8-8-24 resulted in a “no post”. Loosening up to 8-9-8-24 got the rig posting again and we did some quick tests to check stability. Once the keyboard and mouse were recognised by the OS, Massman started to tweak the crap out of it. Disabling services, enabling large system cache, changing page files, ... you know the drill. (used weaks at the bottom of this page)

With no subtiming tweaking we managed to pull off better scores then during my pretesting. Comparing times I new by heart and taking careful notes on a paper, we could see what setting had the biggest influence. RTL (Round Trip Latency) values couldn't be changed at all via the bios as it resulted in a no post with out motherboard.

Before anyone posted any scores we saw the estimated Gigabyte values. 3 levels were set:

King

Duke

Villager
To be a King you had to go sub 9 mins. Luckily for everyone, Gigabyte quickly changed the values to more reasonable values. As anecdote : The night before, in the Romanian "party" room, we heard loads of sub 9 min pre-tests and 3Dscores over 82K. But they were all run with very fast rams and very high Bclocks ( 235ish Bclock with associated Uncore speeds ).

When we found a good value for the Copy Waza (PJ still prefers it over the SPItweaker) and tweaked some subtimings we got a nice high 9 min 6secs score. Further subtiming tweaks resulted in a low 9 mins 6 secs. Being comfortable in the lead ( most teams were trailing by a few seconds) we continued to tweak slowly. Some stuff worked, others resulted in a lower score. When the winning 9 mins 3 secs was posted we had about just enough time left to do maximum two more "final suicide" runs. Upping the ram voltage to 1.9Vdimm and tightening the timings to 7-9-7-24 we started the test.

Initial iterations seemed promising, but it crashed half way. So we had to take the plunge and set some stuff looser again (we opted for loosening the TRFC value). Sadly no post after saving and exiting the bios. Being stuck in an infinite loop did make us loose some precious time. When it posted again we had less than 9 mins 30 to do the CW and PI bench. Massman tried as hard, but when we launched it we immediately noticed it was slower then our best run.

So we had to throw in the towel. Good job OBR and teammates.

We did one last run for fun and looking at them iterations, we would have reached a decent 9 mins 4 secs, but still not good enough to take the crown...


Final result table : Czech republic top spot, Belgium 2nd and 3rd for Latvia (Click to enlarge)

SuperPI 32M - The Tweaks used by the Belgian team

Pagefile to slave partition (512-512)

A couple of registry tweaks, most important one being LargeSystemCache

Maxmem to 615

Manual copy-waza with folder of 600+ MB

• d:/ -> d:/folder (two times)
  
• d:/folder cut to c:/
  
• c:/folder copy to d:/
  
• Best ratio of system cache and available memory was 543/547

Disable most services (only event log - RPC - shell hardware detection)

Set theme to olive (although, in my opinion, it doesn't do that much in terms of performance)

Hardware optimisations

Boot at 220MHz BCLK and use setFSB to clock up to 221,9x18. Note that we had to increase the VTT to 1.62V.

Enable only three cores: Core #0 and #1 for SuperPI 32M, Core #2 for applications such as Intel Frequency Tool and CPU-Tweaker

Memory at 2:10 8-9-8-22 TRFC=88 TRRD=5.

Back-to-back Cas Delay=0 and RTL set to auto (53)


The Czech team doing their thing. Note the big spypaper ! These guys were more than prepared for the event


The winning score, sadly no shot of the subtimings. Apparently they booted windows in diagnostic mode which made a world of difference on the prepared OS. In most cases a tweaked normal boot is faster than a boot in diagnostic mode (click to enlarge)