Product Intel D5400XS "SkullTrail" mainboard, est US$ 650 per unit Intel QX9775 quad-core processor, est US$ 1499 per unit

Website www.intel.com

AMD WAS the first to push the idea of dual-socket high-end PCs using optimised flavours of its workstation/server platforms, with the 4x4 QuadFather units late in 2006. An answer to Intel's quad-core rollout then, 4x4 - two times dual-core setup based on an Asus SLI mobo - was never seen much in public, with the limited units available only in selected markets, and even then hard to get. It did achieve stunning memory benchmarks out of its total four DDR2-800 channels, too.

Intel didn't need to rush with this approach - its dual-die quad-core CPU parts performed well, keeping the performance lead till this very day, in fact. The lacklustre memory performance of their early dual-socket chipsets notwithstanding.

The i5400 Seaburg chipset and its dual FSB1600 solved that performance problem, coupled with large 12MB cache 45nm Penryn CPUs and plenty of PCIe v2 lanes for multi-GPU graphics.

Now, Intel could go ahead full steam and create a balanced super-PC platform where it can sell two expensive CPUs per every system sold. We are looking today at the first one, the D5400XS " Skulltrail".

Record breaker

There are a lot of superlatives about Skulltrail: the first to run eight CPU cores at 4GHz, and the first to fit in four PCIe x16, both physical and electrical, graphics - or other - card slots. The first to allocate more than 400W power for the processors - and even that may not be enough, as you will see here. The first to enable standard Pentium/Core2-style LGA775 cooling options on a LGA771 socket Xeon processor.

It's also the first to support both SLI and Crossfire and yes, the first to allow FB-DIMM memory timing changes - but, as we discover, the job there isn't fully done yet. Finally, the first to require a 1000W or stronger power supply as standard - well, it doesn't really need that much power, even when overclocked.

The EATX format board looks quite impressive at the first sight - two CPU sockets surround the Seaburg North Bridge to enable the shortest trace lengthts for the two FSB paths. Same applies to its closeness to the four FB-DIMM 800 memory sockets, one for each of the four channels. Shorter traces should help some overclocking there, even though it doesn't seem to happen with the current BIOSes yet.

The first three PCIe x16 slots are spaced sufficiently apart to allow Nvidia TripleSLI or AMD CrossFire on thick dual-slot cards. The fourth slot is too close though - not if you use slim one-slot waterblocks. You'll probably have to use waterblocks for your multi-GPU stuff here anyway, or else you won't be able fit in any other cards due to no space left.

We like the full use of new EPC 24+8+8 power supply connectors so common on those 800+ W PSUs these days, plus still an extra 4-pin Molex on the far edge of the mainboard for added graphics card juice on the PCIe slots. And, it comes with the usual complement of I/O interfaces - USB, Firewire, SATA2, eight-channel HD software audio and so on - but no PS/2 keyboard or mouse any more.

...but far from perfect
We saw the pros, now the cons: first, look at that North Bridge heat sink. After working with the Asus Z7S WS last month, we felt its copper heat pipe heat sink should be kind of minimum acceptable solution for a hot overclocked dual FSB piece of chippery.

Here we got an old style piece of aluminum, definitely not enough - even wieth a 6,000 rpm server-class fan on it, with the associated noise, the temperature only dropped from 55C to 48C under our tropical extreme non-airconditioned conditions, before overclocking that is. Luckily, Intel provides an easy heat sink replacement here with four screws and a back plate.

To get the SLI and four GPU slots, Skulltrail uses two Nforce 100 PCIe bridges, each attaching to one of the PCIe v2 x16 links of the Seaburg North Bridge, and providing twin PCIe v1 x16 links in return. Looking at the board's final availability now, we'd have rather had Nforce 200 bridges there to support PCIe v2 x16 on all links. It would help with the next-generation cards.

In fact, I'd have rather had NO Nvidia bridges there at all - in my mind, with dual-GPU cards like HD3870X2 or 9800GX2, two PCIe v2 x16 slots are enough for 99% of high end users. That will still give you four GPUs in two slots at full PCIe speed, performance future-proof till 2009 GPU generation at least.

Then, no SLI without Nvidia bridges? Well, if AMD could provide Crossfire license to its archenemy Intel's chipsets, why can't Nvidia do the same then? If they want to be so stubborn and keep it all, why didn't Nvidia use that Xeon bus license Intel gave them, and created a dual-socket, dual-FSB version of, say, Nforce 790i chipset? I'd love that one, shurely.

Finally, the warm South Bridge and uberhot twin Nforce 100's are covered by a broad and wide aluminum heat sink with a skinny 6,000 rpm fan in the centre. Now that thing is noisy - in fact, due to it the system would be noisy whichever way I did it, with or without that server fan on the North Bridge.

The test setup
The D5400XS mobo was run by two unlocked-multiplier QX9775 CPUs, each with 3.2GHz / FSB1600 at 1.25V default setting, with 150W TDP (actual usage far lower, by the way). These are quad-core 12MB cache entries, the most expensive ones among Intel's X86 CPUs right now. Twin full-copper Zalman 9500AT heat sink fans cooled the processors.

Four Micron FBD-800 modules, each a 2 GB unit with 5-5-5-18 latency timings, provided 8GB main memory on all four channels - we added an extra fan on those, and you should do the same in any FBD system.

A Nvidia Quadro FX4600 card and WD 250 GB SATA2 HD completed the hardware. Not to forget Thermaltake Toughpower 1500, a 1500 Watt PSU supplying power to the whole thing. This was the only power supply on hand able to fully cover the maximum load overclocked Skulltrail configuration - dual OC CPUs and four OC GPUs - according to the Intel spec.

All this ran under the 'extreme conditions' 28 C to 32 C room temperature test, under a bit less extreme OS platform: WinXP Pro 64-bit.

Overclocking
As expected from our Penryn experiences, there was no problem with overclocking up to the 4GHz limit. The default-voltage setup worked fine at 3.6GHz / FSB1600 (multiplier 9x), while 1.35volts was needed for 4GHz / FSB 1600, multiplier 10.

The memory could also work at CL5-4-4-10 stable, compared to the defauly 5-5-5-18 setting. Unfortunately, neither the CR4 command rate nor the Trd chipset latency setting could be changed in this beta BIOS rev. Just like with the Asus Z7S-WS, we'd strongly recommend the next BIOS rev includes the basic dozen memory latency parameters, plus command rate and chipset latency. This, together with more powerful cooling will give us "enthusiasts" far more performance headroom even without pushing FSB to the limit.

FSB and memory overclocking didn't go well in this first run, despite all my efforts with noisy high speed fans. The maximum FSB we could get reasonably stable was 4 x 408, or FSB1632, far less than FSB1750 or so on the Asus Z7S-WS board. Even relaxing the memory timings and upping the FSB, NB and memory voltages didn't help. So, we stuck to 4 GHz / FSB1600 as the base benchmark run.

Here are the initial benchmarks - Sandra XII, PovRay and CineBench10, all 64-bit:

As you can see, another record bunch - you'll also notice that, with the FBD timing adjustments, the overall bandwidth and latency come close to top desktop systems too, while keeping the ECC and reliability benefits of server memory. The 93ns latency is a big jump from 108ns latency of Supermicro Seaburg board - in fact, the earlier "alpha" BIOS version gave us an even lower 89ns latency due to seemingly more flexible DRAM timing support.

If Skulltrail had eight FBD sockets, the bandwidth numbers would have been another 20 per cent or so better, as two FB-DIMMs on the same channel can still have reads and writes simultaneously - opposite of normal DIMMs.

As for CineBench 10, we're now close to 30K figure for the first time - a bit more overclocking tuning, and it will be breached.

There's far more to do with this board, guys: first of all, complete cooling change for the CPU and North Bridge portions - we never overclock South Bridge, so we can leave that noisy fan in place for now. At that point, the dual 4.5GHz FSB1800 is the goal to reach - passing Linpack in 64-bit Linux, maybe?

Then, other memory coming in - these Microns are surprisingly inflexible overclocking wise, and we'd expect high speed FBD from Kingston, OCZ and Corsair, some of them hopefully with 3-3-3-6 latency, to close that gap. As you could see from the latency improvement, even 5-5-5-18 to 5-4-4-10 change did have an impact - so it is not the chipset that's the problem here.

Thirdly, multi GPU testing - supporting both SLI and CrossFire makes this a good "apples to apples" comparison reference platform, at least until a follow-on true PCIe v2 board comes along.

Overall
In summary, a watershed product - the first real dual-socket Intel desktop mainboard, and pretty good one at that. It is not the only one now - we showed you a very competitive Asus board, Z7S-WS, targeting the same audience but with a different .

Intel could have put a better default heat sink on the North Bridge in any case: a taller dual-heat pipe of what Asus uses comes as a good idea. The NB and FSB overclocking really needs more boost here.

Also, the CPU power feed section could be a little stronger when it comes to feeding dual 4++ GHz units - the 4.1GHz limit is just too low, knowing that these CPUs come from the same batch that can regularly boot at 4.3GHz on the single-socket board. But OK, we're talking about TWO processors here, not one.

Also, they could have used something like Creative X-Fi XRAM hardware on-board sound rather than software audio, or simply no audio at all like the Asus entry - not to mention keeping the PS/2 keyboard and mouse for awhile still.

Intel might still do another Skulltrail version prior to the Gainestown / Tylersburg DP dual-socket Nehalem unveiling this autumn. After all, the Core 2 platform is going strong - combine new upcoming CPU steppings with better chipsets using, say, native multichannel DDR3 and Intelligent I/O, and there is plenty of extra performance to gain.

In the meantime, D5400XS Skulltrail, and its upcoming Taiwanese brethren, will be the platform of choice for top-end multithreaded enthusiasts. Too bad AMD cancelled the FASN8 - dual Phenom FX would have looked good there, at least memory-wise.

Good Overall performance, quad GPU support, feature set - a true record breaker
Bad Old Nvidia PCIe bridges; overclocking and BIOS need a bit more work
Ugly Chipset cooling