It's not actually a brand new product - V8 is basically a "Media Creation PC" platform name combining two 3 GHz or faster ClovertownDP CPUs (8 cores total), 1333 FSB Greencreek or faster chipset on an appropriate mainboard with at least one PCI-E x16 slot for graphics. So, anyone can assemble it from the retail components available, or alternatively buy a Mac Pro - an actual Media Creation system with all this in.
V8 test kit came, right yesterday, with these CPUs plus an Intel S5000XVN Greencreek-based mainboard. Four 1 GB FB-DIMMs of Samsung DDR2-667 variety were in, as well as Coolermaster 850W power supply. I plugged in an Asus X1900XTX card, plus a pre-instaled WinXP 64 and 32 for some quick initial benchmarks.
The mainboard, in the EATX format, is a jewel of its own, although, as you can see from the photos, quite a few 'optional' ICs and connectors were left out. There are six storage connectors, configurable as either 6 x SATA3G or 2xSATA plus 4xSAS (serial SCSI for those funky pro drive arrays). 8 FB-DIMM sockets for up to 32 GB RAM. Only one PCI-E x16 slot is there for your 3-D graphics card, however there are two more PCI-E x4 (for extra 2-D graphics or interconnects) and two 64-bit PCI-X slots. Dual Gigabit, basic sound and plenty of fan connectors (I counted 10 plus the two CPU ones) round up the board.
Now, when I turned on the system first time, boy was it hot! Especially the CPUs, VRMs, North Bridge, South Bridge and of course the memories (the ATI is hot by itself anyway). So, basically, everything was hot - and I kept the airconditioning off right here in Singapore (no, there is no cool weather here, unless you count today's "cold spell down to just 30 C"). So, you can guess it was quite a warming experience, so to say.
I went into the emergency intervention mode, worried that the wooden rack housing the board may catch fire. So, four more fans were mounted - one on the FB-DIMMs, one on NorthBridge (permanent add), one just lying on South Bridge and, finally, one Asus spare vertical fan somehow holding betwen two CPU VRM blocks. And, well, it DID help - suddenly, none of these was finger-frying even after the first 20 minutes of benchmarking, compared to boiling heat right after the boot in the original setup - they were still well warmed though. Here are two photos of the pre and post-intervention setup.
The LGA771 server / workstation chip heatsink mounting is, with two levels of back plates, more complex than the LGA775 desktop one, but I find it far more sturdy and reliable - not to mention, it doesn't end up bending the board like those ridiculous stock cooler fixes on the LGA775.
The BIOS had quite a few admin options (like auto-reset if the OS fails, various error-handling options, EFI console), but, sadly, no performance-tuning features except "high performance memory" matching for FSB1333 - which I used for the test anyway. If Intel seriously wants the high-performance PC crowd to look at V8 as the solution among the so-called "enthusiasts", they will need to provide higher performance LGA771 cooling, unlock the CPU multiplier (3.33 GHz would be very doable with these CPUs, if not even higher) and re-look at the performance tuning options in the BIOS.
Technically, from the point of pure CPU MIPS / MFLOPs, this is the fastest PC ever reviewed here - after all, you got 96 GFLOPs peak in 64-bit precision (around 80 GFLOPs obtainable in Linpack benchmark), and humongous 21 GBytes/s of memory bandwidth. Eight cores at 3 GHz each, processing four ops/cycle each, and with dual FSB1333 - not bad at all! Even Apple put them into their current top-end machine.
However, not all benchmarks show the same outcome. Remember our test few months ago, comparing a well-tuned quad-core PC with eight-core workstation? I hoped this machine, with whatever extra optimisations I could turn in BIOS, would improve on the HP xw8400 2.66 GHz predecessor (which, by the way, will soon be replaced by xw8600, HP's upcoming Penryn / Stoakley-based dual-graphics FSB1600 workstation). Yes, it did improve - but only on the tests that relied on the extra CPU clock speed.
So, in Sandra CPU test under WinXP 64, we got 87348 MIPS and 73098 MFLOPs, the record numbers ever - just like the Sandra multimedia test, with 645550 integer and 463483 FP iterations per second. No AMD unit can come anywhere close here, and I don't think Barcelona will have any real advantage here either.
But, the memory bandwidth numbers of 4074 MB/s integer and 4078 MB/s FP weren't exactly great - it is less than half of a similar AMD dual-socket system right now, and also less than half of my XeonUP FSB1667 system. So, Stoakley platform with Seaburg chipset has got its task cut out - make the best out of that FSB1600 and up the bandwidth numbers: say, twice would be a good achievement.
Cinebench 9.5 in WinXP 64 again scaled reasonably well knowing this app has a big serial - non parallelisable - render setup part. 4.7 times faster than the single CPU. The total score of 2382, again the record one, is not bad. In WinXP 32, I just ran 3DMark06 CPU - with 5649 score, it is the fastest ever too under my hands. That's how far I went with the initial set of benchmarks.
After putting in all these fans and running the system today for benchmarks, it is very stable - not a single crash in the first day. The CPU performance on its own is the very top as expected - it is the fastest desktop ray-trace render machine you can buy today, full stop. Lots of apps in that class, specifically well-threaded 3-D media creation, won't really bother much about the memory latency, but will use these CPUs to the fullest, and will surely welcome up to 32 GB RAM surely - minus the horrific price tag and monthly extra power bill for eight hot 4 GB FB-DIMMs, of course.
You do need to take care of a suitable, well-ventilated casing and matching 24+8+4 EPS power supply - 600 W and above, 700 W if populating all 8 FB-DIMM slots or having drive arrays in there. Each CPU here has 150W TDP, and there are two of them, plus the rest of the goodies. So, get ready for the power bill increase, too.
In summary, Intel has answered AMD's 4x4 taunt here - the initial V8 might be their first foray into dual-socket PCs. While it's far from perfect, it does create a base platform from which Intel should be able to create more refined and tuned entries soon. For instance, something like their X38 or Nvidia 680i chipset, but with dual FSB and four memory channels?
I believe that we'll see far more of this during the year, and that both 'Harpertown' - Penryn successor to Clovertown - and Barcelona will be available this year in high-end dual-socket, dual-graphics PC systems, in more than one configuration each. Both should have better cooling and, of course, overclocking options. And, of course, there are more titles, even in gaming, that can use all those cores well. Now, let's bring on that Daniel Pohl's multithreaded, ray-traced Quake 4, guys! µ
Sign up for INQbot – a weekly roundup of the best from the INQ