16 GB on a Skulltrail workstation? And Fast? Yeah, baby...

FB-DIMMS, even in their FBD-800 speed grade, were never really considered performance champions. Yes, with four channels on the Seaburg chipset, these babies do give you 25.6 GB/s theoretical bandwidth, but even the best Sandra scores never went above 9GB/s sustained test results.

On the other hand, use of reliable, high-capacity server modules did let Intel's Skulltrail dual-socket mainboard support some nasty densities fairly early for the desktop. In this case, 16 GB in four sockets would have been a reality - BIOS support quirks aside - far earlier than on any uniprocessor desktops.

But to overclock those? Get real - even Kingston's 1GB heat-piped FB-DIMMs couldn't go above 4-4-3-9 latency settings at FBD-800. And the last parameter, Tras min, was conveniently still preset at 12 by the BIOS no matter what you manually set it to.

Well, maybe all's not that bad - this week we got hold of a quadruplet of Hynix 4 GB FBD-800 modules, with an attractive sexy naming HYMP 151F72CP4D5-S5-C. Each DIMM contains 36 pcs 1Gbit memory dies plus the usual hot Advanced Memory Buffer chip. The modules have a standard defined CL 5-5-5-18 SPD latency, and we didn't expect them to perform much better anyway.

Wrong - without changing any voltages, these were booting at 4-4-3-9 all the way into Windows Vista64. Interestingly, the previous BIOS 1140 didn't recognise 16 GB properly, showing the configuration as 8 or 12 GB depending on the luck. The upgrade to the newest BIOS 1149 solved the problem.

Now we had 16 GB on Windows Vista - however, the memory benchmarks, including Sandra, couldn't complete until I changed the DRAM die voltage to 1.92 volts (the AMB chip voltage stays at 1.5 volts all the while). The 4-4-4-9 is what at the end passed all the bandwidth, latency and stress tests well.

Note that these are the first FB-DIMM modules of this density tested by us to successfully reach this performance level. Even the 2GB modules by Micron and Nanya couldn't go beyond 5-4-5-10 after all attempts. So, memory tuning is possible even on very high capacity FB-DIMMs after all, and Hynix, a company not often associated with overclockable performance memory, is the unexpected beneficiary here.

With an 8cm fan mounted to cool the FB-DIMM modules, the temperature only went up by about 2C from 43C to 45C for the modules, not bad as the ambient test room temperature today was 34C anyway - this is Singapore after all. Can some London weather visit us instead for a change?

Any benefit in the results over the usual 5-5-5 timing? Well, a jump from 8.2 to 8.45GB/s in Sandra 2009 SP1 bandwidth results, and from 102ns to 95ns in latency results - this latency comes close to many high end UP desktops then. I didn't overclock the CPU this time at all for this purpose.

As you can also see, while our Hynix-enhanced Dual quad-core has pretty nice latency, the memory bandwidth test of a Xeon 5500 - Nehalem or Gainestown DP with i5500 Tylersburg DP chipset and six channels of DDR3-1333 registered memory will still be 4.5x faster than a dual-socket Xeon 5400 Skulltrail or equivalent.

The problem is, we may have to wait till sometime early next year for more of these to appear. In the meantime, a dual socket, 8-core setup like Skulltrail or Asus Z7S-WS (the powerful Xeon FSB overclocker workstation mobo - the next test target for these modules), coupled with a quad set of 4 GB DIMMs like this Hynix spread, will give you a taste of a super workstation with performance tuning potential to boot. µ