Intel Canterwood thrashes opposition, takes Sun to the cleaners

BY NOW we're all aware -- hardwired in our brains - and through Intel's marketing campaign, that the "blazing-fast" Pentium 4 processor was designed with next-generation, high-performance, high-bandwidth, speed-hungry applications like realistic 3-D gaming, high-definition TV, multiple video streams, and broadband Internet in mind. Or, when you remove all the marketspeak, it means that Pentium4 is much more sensitive to memory speed for good performance than the good old Pentium III.

We can see the proof of that over the past two years, as different Pentium 4 revisions with various bus speeds (400 or 533 MHz throughput on a quad-data-rate 100 or 133 MHz FSB clock for 3.2 or 4.2 GBytes/s peak bandwidth respectively) appeared, and with wildly different memory configurations: from high-bandwidth, many-pages-open but also high-latency synchronous-to-FSB dual-channel RDRAM (PC800 and PC1066), single channel PC133, DDR266, DDR333 and DDR400 SDRAM, and lately dual-channel DDR266 and DDR333 SDRAM to match or exceed the P4 FSB bandwidth with DDR for the first time.

Many tests showed that the benchmark performance, not just on low-level tests, but also on many power-hungry apps, were affected quite a bit by the FSB and memory choices, as well as by the performance of memory controllers on the chipsets.

Riding the 800MHz Bus
So, to compete well against the upcoming X86-64 platform from AMD, at least on the 32-bit front, Intel took a major step forward and, instead of going first to 667 MHz FSB, it used the faster-than-expected DDR400 ramp-up to go straight to 800 MHz FSB with dual-channel DDR400. Does this mean it will be the fastest Pentium 4 memory subsystem before Prescott arrives?

Not really -- SiS and Rambus are supposed to be close to completing the SiS659 quad-channel PC1200/PC1333 Rambus chipset for P4 soon, and that thing should give something like 9.6GB/s memory bandwidth with four PC1200 RDRAM channels -- 50% above the 800 MHz FSB 6.4 GB/s bandwidth, hundreds of open pages for "almost-always" RAM page hits - and hopefully a much faster memory controller in the chipset than the existing SiS658 RDRAM chipset.

Intel D875PBZ board
So, expect Intel to move the whole high-end of their Pentium4 line to 800 MHz FSB today - - the fastest part announced is a 3GHz CPU, of course with the holy HyperThreading mantra enabled, and, thanks to Intel's wisdom and foresight, we all got to see the 3GHz D875PBZ config with 512 MB DDR400 RAM (two Kingmax DDR400 CL2.5-3-3-6 5 ns DIMMs) and two Seagate Barracuda 120 MB Serial ATA HDDs.

The "Bonanza" board is a fine piece of work - some 2 cm shaved off at the end after the fifth PCI slot to let us have more cables spread at the casing bottom, the board is based on the i875 Canterwood chipset. The details you can read all over the Net, but in summary -- it is a 800 MHz FSB, speed-matched dual-channel DDR400 (ECC or non-ECC) memory bus to sync with the FSB for maximum speed, PAT (Performance Acceleration Technology without Gelsinger) to lower the chipset latency to the memory further, wirespeed Gigabit Ethernet with the i82547EM chip on a separate DMA-based 266 MB/s link to the north bridge . Hey, this makes it a good visualisation cluster node!

The board, like most new Canterwood boards, has the i82547EM, as well as total of eight USB 2.0 ports. Besides the usual two ATA100 IDE ports, there are also two Serial ATA disk ports (ICH 5R version of the south bridge), supporting either individual disks or RAID 0 (disk striping) enabled. To me, this makes sense in a multimedia streaming workstation, but the OS and apps, as well as critical data, should still be on something more reliable than striped drives - if one drive of the two fails, all data is gone! Maybe for high-end desktops a RAID 10 (mirrored striping) makes more sense. Not only do you gain reliability, but the read speeds improve even further because you can read from four drives at once!

There are no graphics or sound on this high-end Intel board -- after all, after splurging all the money on the top-end CPU, board and memory, you should also get something like NV35 or R350 for graphics, and Creative Audigy2 Platinum or alike for the sound. AGP 8X and 5 PCI slots complete the "Bonanza" picture.

Feed The Need For Speed
So, our team sacrificed their weekend, the usual outings with the opposite sex and other pleasures in life to get this done by the time Intel opens its NDA iron curtain. You can see some of the results in Fudo's article here. And here are some extra ones.

System	3.06GHz	3.0GHz
CPU	Pentium 4	Pentium 4
FSB Speed	533 FSB	800 FSB
	PC1066 dual	DDR 400 dual
Sandra 2003
Memory int MB/s Kingmax	3350	4831
Memory FP MB/s Kingmax	3343	4869
Memory int MB/S Corsair		4857
Memory FP MB/s Corsair		4893
CPU Multimed Int	14096	13824
CPU Multimed FP	22186	22186
CPU Integer	9250	9136
CPU FP	5636	5782
PC Mark 2002
CPU	7521	7456
Memory	7631	Kingmax 8488
		Corsair 8576

Wow, one can say, the memory flies! Especially when I put the Corsair XMS3200C2 pair of DIMMs there with 2-3-2-6 latency settings. This would be the machine for high 3-D highly-textured game frame rates haha, as there is a plenty of memory bandwidth for the AGP8X graphics to stream in gigabytes of textures while the CPU still does its job - well SiS, you can still beat them with quad-PC1200 Rambus on this! Basically, if properly designed, the SiS659 could allow both AGP, South Bridge and full-speed CPU memory access in parallel without major contention - but let's see that miracle happen...

The generic CPU speed in non-memory intensive benchmarks is more or less the same as 3.06 GHz 533 MHz flavour -- the 2% lost on clock down to 3.0 GHz are usually gained back on that extra bit of memory performance even in such apps. But in memory-intensive stuff, it makes a whole load of difference - again, this thing really flies.

Summary
Intel has really made up for the initial shortcomings of Pentium4 design by basically feeding it with code and data at a speed unprecedented for a desktop PC. Basically, this machine has twice the real usable memory throughput of a Sun Blade 1200 dual-CPU 1.2 GHz UltraSPARC III workstation which costs, what, six times as much for a comparable system? Not to mention, this single 3 GHz P4 will probably beat those two SPARCs on most computational jobs too.

Finally, the board is supposed to be Prescott-compatible via a BIOS upgrade -- let's see how that one runs in comparison to the Athlon64 desktop, as both should be out at about the same time. In the meantime, expect a small 3.2 GHz speed bump in a month or two... µ

See Also
Intel Canterwood 3GHz/800 FSB combo is a real road scorcher
Review of an Evesham Canterwood machine