High-end mainboard directions need performance targets

A WEEK OR SO ago, we saw that a high end mainboard should be strong enough to handle all the stuff piled up on - or hanging off - it, and the IO features that make sense of it. You'll also have seen readers' comments on the above.

Now, what about the real reason these boards sell for, besides the user's egos - their performance capabilities? And, hopefully, delivering that performance for as long as possible without fail?

If you spend US$ 1000 for, say, quad core Intel QX6850 CPU - I doubt its 45 nanometre Yorkfield Extreme follow-on, or the first AMD Phenom FX, will be any cheaper initially - and another, say, US$ 400 for, say, Asus Striker Extreme or similar board, you most probably don't intend to run these anywhere near the stock 3GHz/FSB 1333 speed. As much higher as possible, but still reliable, day after day, at least during the first one year of operation - after that, either the warranty expires, or you, the eager 'technology enthusiast' go for a fresh replacement to stay at the top of the performance charts.

The first, and most obvious, approach to CPU speed up is, of course, increasing its clock speed. If a 3 GHz QX6850 is set at 130W max TDP at somewhere around 1.3 volts, its top TDP will go to around 200W at, say, 3.6 GHz and 1.45 volts supply. If you go a notch further, and go for a -40 C cryocooling solution at 4.2 GHz and 1.65 volts, we're talking about well above 250W TDP. The freezer will take care of the excess heat, but the mainboard's power delivery should better be able to provide well in excess of 150 amps peak load, consistently.

At this stage, only the best 8-phase or, quite possible, 12-phase power supply systems with top notch MOSFETs, digital VRMs and other related componentry can suffice - no space for compromise. The BIOS voltage settings should have 0.0125 v or finer CPU voltage steppings, at the very least. Not to mention factory-enabled provision for copper liquid cooling block right at the VRM area, besides standard fanned heat pipe-linked copper heat sinks. The board PCB should also handle the higher overall temperatures resulting from such prolonged overclocked operation easily, both from the cooling composite and PCB trace reliability. Keeping the board PCB area around CPU and VRMs cool is especially important if using liquid or freeze cooling, where there is little airflow around the area, usually from secondary fans on top of the heat sinks, if any.

What's the point of speeding up your CPU if the FSB stays slow, choking its four hungry cores - however, speeding up Intel's GTL+ FSB isn't exactly simple, or else we'd have had FSB1600 as standard way before, wouldn't we? Firstly, the chipset's North Bridge has to be designed for high FSB operation - Intel's P35 and X38, as well as Nvidia's Nforce 680i and its C72 & C73 follow-ups later this year, are such examples. The 65 nm-based X38 and C72 / 73 should both allow above FSB2000 reliable long term operation with top-end quad-core Yorkfield.

To handle the heat out of a NB and FSB working upwards of 1.5 volts, a quality soldered heat spreader, combining high thermal conductivity and protection for the die, should be on both North and South Bridge - a substandard thermal paste in between the die and spreader is a no-no at the high end. Nvidians, please do not stinge on this when you do one on the C72 / 73! Knowing the current Nforce, their chipset will surely need at least heavy heat pipe, if not liquid, cooling, to show their true performance. Beyond this, watch for the routing design that minimises the path between the CPU and North Bridge - length, noise and so on. This can significantly affect the FSB clocking capability too.

Either way, at the BIOS level, the board should offer appropriate options for both FSB and North Bridge voltage (hopefully in 0.025 volt or finer steps) and GTL+ bus parameter tuning. The current 65 nm Intel Core 2 CPU parts should achieve reasonably reliable initial (i.e. completing the usual Windows benchmarks like 3DMark and surviving for an hour doing that) speeds of FSB1800 for quad-core dual-die top-end parts, and FSB2150 for dual-core single die parts, again top-end. We're talking about P35 and Nforce 680i chipsets in either case.

On the memory side, I am still a firm believer in synchronous, matched memory to FSB approach, and optimising the memory controller for it: my half-year old reference Striker Extreme NF680i at FSB1667 and in-sync DDR2-833 low latency in sync RAM with CL3-3-3-5 settings still beats any DDR2 or DDR3 memory on the P35 chipset, even dual-channel DDR3-1667 CL8 at FSB1667 - either Sandra or Everest will still produce at least 8% better benchmark results, whether bandwidth or latency, in favour of NF680i here.

If a typical memory burst access it X-1-1-1 in DDR2, or X-1-1-1-1-1-1-1 in DDR3, that X at the beginning is the latency to the first word read or written. The larger the X number on a given DRAM frequency (i.e. larger latency), the slower the first word access is. The higher the DRAM frequency and therefore bandwidth, the faster those -1-1-1's are. Some apps prefer the bandwidth, but many still like lower latency more.

After all, what's the point of extreme memory bandwidth if the CPU FSB can't make use of it, yet you add latency due to async operation? Unless of course, you're feeding twin PCIE x16 graphics cards with huge data streams out of that memory at the same time as CPU accesses it - video processing or GPGPU use are the possible candidates. Otherwise, stick with low-latency, matched bandwidth RAM. RAM voltage and drive settings, again as fine as possible, no less than 0.05 v steps - important here is how the board drives all four DIMM sockets at high speed, and how fast it can go sustaining the 1T command rate (there is a measurable performance impact when dropping to 2T rate). For two modules, I did it easily even at DDR2-1150, but for four modules, sometimes even DDR2-800 may be hard - this depends just as much on the mainboard and chipset as on the memory itself.

The Nforce 680i and, now, Intel P35 BIOSes on most high-end board offer now very detailed settings of DRAM timing parameters, usually no less than 10 of them. Besides the ability to fine tune them and, if failing, give you a BIOS warning rather than just freeze at boot, here it is important to ensure that the BIOS settings made are actually running. I remember one MSI board, the 975X Neo PowerUp, where I set the DRAM to 3-2-2-5 in BIOS, but all utilities (Sandra, Everest...) detected 3-5-6-5 as the real timing.

Talking about memory, I feel that bringing back ECC memory support may not be a bad idea at the very top - it opens the door to many "serious" scientific, techical and engineering accounts who'd like such boards in their machines as a mixed desktop cum workstation platform.

Moving to the I/O, before the X38 and C72 / 73 chipsets bring PCIE v2.0 with double the bandwidth per pin, Nvidia's LinkBoost 25% speedup of PCIE x16 lanes on the current 680i has some minor (and I mean, minor) benefit on the 3DMark06 results, especially if your memory system bandwidth is well above the CPU FSB one, resulting in more spare bandwidth for the GPU. You could achieve pretty much the same results by simply overclocking the PCIE as long as the GPU clocks are separate from the PCIE.

Finally, I don't believe in overclocking anything on the South Bridge, unless there are extra GPU PCIE lanes there - the SATA2 ports are still waiting for the hard disks to catch up, which may take another 2 years at least. As for the USB, I guess offloading its interrupts from the CPU via a bit of Intelligent I/O would be far more important than speeding its bandwidth further.

In summary, the high-end board performance, both peak overclocking and sustainable every-day operation speed, will depend on the components used, where even every capacitor and resistor may matter, on the PCB design and quality, on the BIOS tuning capability - and, of course, on the proper cooling of all this, as discussed in the first part of this analysis series. If these high-end entries are the Ferraris and Lotuses of the mainboard world, then they got to follow the same rules of overall component performance and quality brought together by superb design into the finished product. µ