Printing-ink veterans don't take cyberspace journalists too seriously - Roy Greenslade, Guardian Online
WITH THE HIGH-END market firmly in control through current Core i7 Nehalem chippery, Intel could afford to release the 32nm Westmere line slowly, this time starting from the low-end to test the waters and fine-tune the core before releasing the high-end six-core chip later this quarter. Its new Clarkdale processors, the world's first 32nm x86 chips, provide dual-core performance at possibly the lowest ever power usage at such speeds, and of course there is also the added 45nm GPU chip in the package interfacing via QPI to the CPU to access its memory and the rest of the system.
In our two part review, we'll look at the Core i5 661, the flagship CPU and GPU combined package that, in a sense, brings along AMD's promise of Fusion on an Intel platform first. In Part 1 here I turn that new integrated GPU off, and focus on the dual-core Westmere CPU performance first, while in Part 2 I will look at the GPU potential and more benchmarks.
By the way, to use the display output from the integrated GPU, you need the new Intel chipsets, the H55 or H57, the latter having storage acceleration. I used two new motherboards to try out the CPU. Intel's DH55TC is a decent well featured microATX mainboard, but with very limited overclocking capability, while the Asus P7H55-M Pro has far better overclocking potential and matching solid capacitors and so on. Both boards have the usual I/O interface assortment - but not the newest USB and SATA revisions - as well as an PCIe X16 graphics slot if you still want to use an the external graphics card, which as you'll see might still be a good choice in some cases.
I used 2 x 2GB of Kingston DDR3-2000 DIMMs, the old trusty performers that handle both low latency and high bandwidth scenarios well. Interestingly, the memory controller on the new 32nm Core i5 seems not to allow as fast timings as its 45nm brethren. Basically, while DDR3-1600 CL6-6-6 was easy on any of the 45nm parts at 1.6 volts, here it wasn't possible even at 1.64V. At the end, I stuck with CL7-7-7 but at 1.5V, and that worked fine. The importance of high memory bandwidth here is not so much for the CPU, as dual cores would barely saturate the memory, but for GPU performance. For cooling, the new Thermaltake Contac 28 heat-sink fan was used, with direct heat pipe contact to the CPU heat spreader surface.
The CPU is rated at 3.33GHz core clock, 25x the base 133 MHz BCLK, with the GPU at 900MHz. The QPI that links them can be set at a variety of speeds on the Asus board, but any speed above seven Giga transfers per second didn't work, at least at the Auto voltage settings. I tried to see how far we can go while relying on Auto settings, with the iGPU both enabled and disabled. In the first case, the enabled GPU reduced the overclocking potential to about 3.75GHz or 25x the 150MHz BCLK, while the second case, with an external GPU used instead, easily propelled the overclocking potential to well beyond 4GHz.
Here are the benchmarks with the 4GHz CPU at 1.30V with Turbo and HyperThreading enabled, dual channel DDR3-1600 CL7, and the GPU disabled. I used an ATI Radeon HD5870 - yes I know it's an overkill for this system but it was really handy - and the old MGE Magnum PSU with LCD display showing power usage. Interestingly, without the external GPU, the total system power consumption while running Windows 7, including the hard disk and DVD drive, was no more than 62W at any time, or 85W with the HD5870 when the GPU wasn't doing 3D work.
I ran the brand new Sandra 2010c suite, our usual UK-made reference suite. It fully supports the new CPUs, of course. Here are the CPU throughput, cryptography and memory bandwidth bechmarks.
Memory bandwidth bench
The performance is quite impressive. It does overwhelm a 3GHz quad core Athlon II 640, which is a direct competitor price-wise. Then, the AEC acceleration instructions play their part in the enormous score jump. Imagine how this will perform on the six-core "Gulftown" version of WestMere in a few months.
Also, take a look at the CineBench R10 64-bit:
Overall, not bad. Remember we didn't really push this new baby to its limits here yet. Only Auto settings were used, and knowing these are initial steppings, there's plenty of room for improvement speed-wise. Also, the initial CPU benchmarks show a lot of promise. Even if Intel doesn't officially announce higher speed dual-core parts beyond Core i5 670 during 2010, the usual stepping revisions should bring the potential further up.
First Intel 32nm CPU, really low power, decent performance.
Integrated GPU seems to affect overclocking initially, as expected.
There's only a 2-core low-end and a 6-core high-end in 32nm from Intel in 2010, for now.