DDR400 lives: Twinmos vs Corsair

RAMBUS STILL hangs on, and DDR-II is still far from volume shipments. So, what's DDR doin'? Well, its racing past the 400 MHz mark, and with quite good stability to boot. While 2.5V DDR400 (PC3200) is quite a bit hotter than its upcoming DDRII-400 1.8 v sibling, and only one DIMM per channel is possible for full speed, the advantage is simple: it ships now.

And yes, if you got Nforce 2, or any of those nifty new Intel Granite Bay or SiS of VIA dual-channel DDR DIMMs, well you can have two of these, one per channel, for a stunning 6.4 GB/s bandwidth (same as Intel's Itanium 2 or upcoming Power4-based IBM PowerPC for Apple Macs).

Here I look at two example DDR400 (PC3200) 256 MB DIMMs from Twinmos and Corsair. Both are full sales units, not engineering samples. The Twinmos part is declared with a CL2.5 CAS latency, while Corsair CMX256A-3200C2 is supposed to reach CL2 CAS latency (but not pure 2-2-2; no DDR400 RAM can do it yet). The Twinmos DIMM uses 8 pcs of Twinmos' own 5 ns parts, while Corsair parts are hidden under a thick black heat spreader - supposedly they are 8 pcs of hand-selected 6 ns parts which perform extremely well.

I plugged the DIMMs one after another into a WinXP Pro-based 2.4 GHz P4 (533 MHz FSB) on the Asus P4S533 mainboard (SiS 645DX chipset - a stable board we used in our previous tests). My reference graphics, Nvidia Quadro 4 900XGL, was the choice as usual, but this time that was not important. We selected a couple of tests that are very memory-intensive.

These include memory and multimedia Sandra tests; PCMark2002, and Wstream for more realistic bandwidth figures - and (drum roll please) here are the results:

DDR 400 - PC3200 comparison: Twinmos vs Corsair
Test CPU: Pentium 4 2.4/533

DIMM	Twinmos	Twinmos	Twinmos	Corsair	Corsair	Corsair
Timing	Default	Better	Best	Default	Better	Best
	2.5-4-4-7	2.5-3-3-6	2-3-3-6	2.5-4-4-7	2-3-3-6	2-3-3-5
Sandra2002
Memory int MB/s	2816	2869	2871	2818	2866	2870
Memory FP MB/s	2822	2865	2868	2820	2865	2869
CPU Multimed int	9533	9544	9546	9545	9531	9551
CPU Multimed FP	11613	11615	11627	11615	11612	11620
PCMark2002
CPU	5849	5864	5917	5863	5902	5903
Memory	5609	5728	5733	5587	5742	5749
Wstream (MB/s)
Copy	1405	1459	1478	1405	1476	1478
Scale	1309	1388	1398	1309	1399	1398
Add	1437	1537	1546	1440	1545	1546
Triad	1444	1538	1550	1445	1547	1548

As you can see, we ran both modules under default timing, as well as best timings we could achieve. In the 2-3-3-6 figure, the first digit is CAS latency; the second digit is RAS-to-CAS delay; the third one is RAS precharge, and the fourth one is RAS active - all in 5 ns cycles (200 MHz DDR).

Funny, in the case of Twinmos RAM, much more benchmark oomph is gained by improving the RAS-CAS delay rather than the usually glorified CAS latency.

Also, shortening the RAS Active time by a cycle (the only timing advantage of Corsair over Twinmos in this test) doesn't seem to yield noticeable improvement. Most importantly: there were no crashes even at the best settings.

So, the DDR400 is here to stay - AMD, since your Hammer is late, you may as well update its built-in mem controller for DDR400 spec; you'll need it to compete against Prescott's 667 MHz FSB. Also, how about "Barton" Athlon with 400 MHz FSB?

Back to the DIMMs: In general, barring the usual ~1% benchmark result sway, both modules perform about the same. My preference could go to Corsair for that intimidating-looking heat spreader, useful for hot systems with less-than-perfect airflow. Otherwise, kudos for Twinmos for exceeding the spec and reaching the PC3200 CL2 level - reliably... µ