THE FASTEST THING in the world for your desktop right now is the dual CPU Xeon W5590 platform. So, as promised on 8 August (although a little later than promised owing to swine flu and typhoon Morakot), here is part two of our review.
This time we focused on trying out the configuration limits of the various components used.
Thermaltake's uber-PSU, the ToughPower 1500 (they have a 2000W version too) would probably load most home power sockets worldwide, and definitely be beyond the grasp of any USA or Japan 100-110V sockets. Of course, this PSU provides sufficient juice even for a hypothetical overclocked dual 4GHz Xeon machine with three dual-GPU cards, if you wanted such a monster.
So, we went to the other extreme, and tried to see how far we could go with the smallest possible PSU. The Seventeam ST850P unit at 850W was a convenient test bed.
The compact PSU managed to handle the dual Xeon with the Asus Matrix GTX285 card without a hitch, even though it didn't even have the full 24 + 8 + 8 pin power connector combination - the last one was missing. This either means that 850W is more than enough, or that the Seventeam unit has darn good +12V rail.
Unlike the overclocked PC systems, the engineering 3-D workstations may favour liquid cooling for another reason: silence. The enormous amount of heat from the dual CPUs, coupled with the other components, such as the huge memory array, GPUs and disks, would require batteries of noisy fast fans. Too much noise can affect the productivity and even health of the highly paid engineers using the systems, so there is a direct economic benefit of having a silent system. And yes, they want it absolutely sealed as no one will have time to tinker with leaks.
So we have an Asetek dual CPU LCLC which behaved well on the thermal tests. On the previous Xeon W5580 Supermicro platform as a test bed before the new CPUs arrived, the Asetek's sealed liquid cooler provided between 8C (idle) and 11C (full load) lower temperature than the Supermicro default heat sink and fan combo.
For curiosity's sake, we also tried twin Coolermaster V8 fans, the best we could muster for standard air cooling on the LGA1366 socket.
The Asetek was just 1 to 2C lower on idle, but managed to keep temperature 4 to 5C lower when running the CPU full load benchmark. You can see here the temperature graph on the W5590 for Asetek LCLC when running the Sandra CPU bench.
Talking of benchmarks, we ran a few more rounds of key tests to see how far we could go.
First, the 3DMark Vantage on the Asus Matrix GTX285, a very fast and very overclockable top-end GPU.
Since these eight cores running at 3.47GHz - yes our samples did run on Turbo - should give quite a bit of physics effects computation power, We ran it with both Nvidia PhysX off (so all on the CPU) and PhysX on the GPU. Both results are great (see below) although there still is a 3Dmark benefit for PhysX on as you would expect.
And talking of 3D, have a look at the CineBench 10 result: good basic CPU run and scaling as expected. But the real challenge - and still unfinished as of now - was getting to 100GFLOPs usable in Linpack double precision matrix multiply benchmark.
This routine, one that often exceeds the CPU TDP, drives everything on the processor and memory system to the hilt, and the results increase depending on the matrix size and data granularity for the allocation.
As our matrix size increased from 10000x10000 to 50000x50000 (the latter requiring well over 30GB RAM) the results crept up to about 99GFLOPs. Note that, for the smaller size, changing the data allocation to 6KB (remember we've got six channels of memory here) did improve things a bit. Now, if we had at least CL7-7-7 - in SPD - DDR3-1333 memory here, we think we'd have over 100GFLOPs as Linpack is sensitive to memory latency.
Oh yes, and all the tests in this review are done with HyperThreading off.
So where do we go from here? Well, before we get hold of the six-core Gulftown in the same socket (due in the next six months or sooner), we expect to see one more possible speed bump for this set-up. Alternatively, a multiplier unlocked version matched with a suitable OC mainboard and lower latency memory would work wonders.
We'll have more tests with this platform soon, and on other mainboards as well. µ
Top performance, balanced processor, memory and I/O, not that much power usage.
Memory latency settings could be adjustable.
Price - $1500 per chip - you get what you pay for though.