Wednesday, February 29, 2012

Fun with a SR-2

As with many of my shenanigans, this one started with "I found these processors on Ebay'', and became an interesting adventure.
Actually, this one started with "I found this motherboard on Ebay", an Intel S5520UR to be exact. Unfortunately, the motherboard was either dead or lacked the appropriate microcode for the D0 ES CPU's I had, because it got stuck in a boot loop and refused to POST.
Version 1, featuring a derpy Intel board
After more staring at Ebay searching for a better motherboard, I gave in to my dark desires and ordered an EVGA SR-2, the undisputed god of enthusiast boards.
Big box, little box...yes, I'm an EVGA fanboy
The SR-2 is the spiritual successor to such greats as Skulltrail (Intel's OC'able dual-socket 771 board), the Asus L1N64-SLI WS (the only implementation of the AMD 4x4 "Quadfather" platform), the Asus PC-DL (somewhat overclockable dual P4 Xeon board), and the Abit BP6 (OC'able dual Celeron motherboard from the late 90's). With the possible exception of the BP6, all of the above boards had their issues - Skulltrail was unholy expensive and needed hot-running FB-DIMMs, Quadfather was slower than a single Core 2 Quad, and the PC-DL was plagued with overclocking bugs. The SR-2 is, in my opinion and that of many review sites, a perfect combination of implementation and timing. Unlike its predecessor Skulltrail, the SR-2 was released at a time when multithreaded software was already commonplace. The SR-2 also runs on Intel's first well-balanced multiprocessor platform - Nehalem eliminated the FSB bottleneck present in the Core-architecture Xeon platform.
You can read more about the SR-2 here. Guru3D documents an excellent benching run here, and it is (as of this writing) available from EVGA new for $550 or from B-Stock for $400.

Assembly and first boot was uneventful; CPU-Z showed a load speed of 3.06 GHz on the pair of Xeon X5560 processors I had installed, thanks to EVGA's out-of-spec implementation of Turbo Boost which could hold the turbo multi under full load on all four cores.
Boxen are sexier when naked

First attempt at overclocking was at 180 BCLK, 1.375 Vcore, 1.35 Vtt. This resulted in a system that crashed after about 2 minutes of stress-testing under LinX. Fortunately, further inspection revealed that Vdroop was dropping Vcore down to about 1.30V, enough to destabilize the processors. Turning off Vdroop let the system pass several hours of LinX on the "all memory" setting.

Sadly, the overclock caused anywhere from 2 GB to 10 GB of RAM to disappear. Both Windows and the BIOS failed to detect it. This is not unheard of - Core i7 systems have been known to drop channels of memory, often due to a mis-seated heatsink or a bent pin. Reseating the heatsink didn't help, and neither did loosening the timings (I had my 9-9-9-24 DDR3-1333 running at 11-11-11-27 DDR3-1066 and it still vanished). Upping Vtt to dangerous levels (1.55V) didn't help either.

I eventually gave up, chucked the awful OCZ Gold DDR3-1333 I had (it was only $3.50 a GB from Directron), and ordered 24 GB of G.Skill DDR3-1600 from Newegg. This mostly cured the problems - the system still drops a channel once every few reboots, but it was stable when it did detect the memory.

Or was it? After about 3 days of running Folding@Home -bigadv, I got a BSOD. Upping Vtt a little helped, but it still crashes around once a week, usually when ambient temps are high. The SR2 currently folds 24/7 at 1.41875 Vcore, 1.425 Vtt, 1.4V ICH, and 11-11-11-28 timings on the memory. Temperatures are in the mid 30's idle, mid 70's load. The RAM itself is actually good for quite a bit more, but the memory controllers can't handle it.

And finally, the obligatory screenshot:
Why hello there, Taylor Swift
The board is benchmark-stable at higher BCLK's and voltages - in fact, it holds the #3 on the 8-CPU wPrime benchmark at 4357 MHz.