Intel QX9650 Yorkfield finally comes out

INTEL IS FINALLY unveiling the 45 nm Yorkfield QX9650 Core 2 Extreme part, some seven months after the first salvo of seemingly fully-functional benchmark demo systems flooded the expo halls of Beijing IDF.

Why not come out earlier? Well, there was obviously no competition to rush against this year, so they could take some sweet time to fine tune the process. The very first chip is a good test run for tuning its more important "watershed " successor on the same process, the Nehalem.

And thie ship has seen some serious tuning. The QX9650 part is, by default, a 3GHz FSB1333 - multiplier 9 - part. But few of the "early adopters" shelling out $1K or so for it will run the CPU at anywhere close that speed.

We're not in the business of "suicidal overclocking" but "making use of the inherent die performance headroom to enhance the user experience" or, in layman's terms - reasonable everyday tune-up.

We took an Asus Maximum Extreme high end X38-based mainboard on the newest 401 beta BIOS, with 2 x 1 GB Supertalent DDR3-1600 CL7 DIMMs and a Leadtek Leviathan watercooled 8800Ultra - the CPU was cooled by the standard reference Corsair Nautilus 500 watercooler. This first benchmark run was XP 32-bit only.

The "reasonable tuning" guidelines were to stick within "green" settings in BIOS for the CPU, chipset and memory voltage, ie without going into typical overclockers "yellow" band, not to mention the "dangerous" red band - and still have some headroom on the CPU voltage side within the green band.

The reason for that is that, in 64-bit mode when running stuff like Povray or Linpack, we'll need a few more voltage notches up to achieve the same total stability.

So, running the CPU, chipset and FSB at 1.4 volts, and memory at 1.9 volts, and a couple of "trial and error" sessions while keeping the CPU temp below 40 C in BIOS, we got a nice little setup running at - 4.27GHz and FSB1708. The memory was in sync - DDR3-1708 CL 7-6-6-14, not bad at all!

Now, here are the benchmark results - for curiosity's sake, we did a quick Sandra and PCmark05 comparison with the only other CPU we had run before at the same clock near stock voltages - the Presler dual core Pentium at 4.27GHz as well, exactly two years ago, and at similar power envelope.

Intel achieves between four and eight times higher peak CPU and FPU throughput in two years. Yet only 25 per cent higher memory bandwidth and 15 per cent lower latency. Now you can see why the Nehalem generation is so important for Intel, finally removing that last obstacle that AMD makes fun of over the past three years.

In other benchmarks, 3DMark06 CPU was at a record 7025 score, higher than even the 6970 achieved by the eight- core, dual Harpertown 3GHz box - but of course, the next 3DMark will be finer threaded, so eight-core monsters will fare better then.

Keep in mind, this is, up to now, a very stable configuration. We ran other CPU and system intensive tasks on it over the past few days without a hitch or visible CPU or chipset overheating. As far as we hear from other testers, the performance is same or better, so 4GHz Yorkfield desktops might not be uncommon.

The easy 40 per cent overclock within "green" limits is great, and Phenom will have some huge shoes to fill in here. Our followup instalments in early November will focus more on the 64 bit performance and further tuneups. µ