By Nebojsa Novakovic: Sunday 28 March 2004, 17:46
SO THE USUAL CeBIT time came and went away, with its heap of new "attractions". This year, one of the attractions was the new DDR2 memory - in fact, so attractive that Corsair got its expo cargo stolen.
The very first batch of DDR2 memory, to be used in Intel's 915/925 chipsets, provides the same 400 MHz effective burst rate as DDR400, costs more, but consumes less power - of course, 1.8v devices would consume almost half of the power compared to similar 2.5v devices in the same process.
However, at least until low-latency and DDR2-533 or 667 versions become common, is there any benefit experienced by moving from DDR-400 to DDR2-400? I mean, higher price should mean higher performance?
My 400 MHz beats your 400 MHz
Well, in the DDR2 price, higher price does buy you higher latency, therefore, in theory, lower performance! The best DDR2-400 DIMMs available initially can only provide 3-3-3 basic latency figures, compared to 2-2-2 for the top DDR-400 DIMMs. Why is that so? Before looking at the actual DIMMs in actual DDR2 systems, two possibilities come to mind. First, the DDR2 memory uses 4-bit prefetch, i.e. every 4 bits from a memory array that actually runs at 100 MHz are channeled, or multiplexed, to one bit on the external memory data bus, which effectively runs at 400 MHz (200 MHz x 2). In DDR2, with 2-bit prefetch, every 4 bits from a memory array that actually runs faster, at 200 MHz, are multiplexed to one bit on the external 400 MHz memory data bus. So, yes, if current DDR400 DIMMs used 4-bit prefetch instead, their data rate could support a DDR2-800 speed!
And vice versa, a slower internal memory array will definitely add to the both read and write latency. But why is the memory array slower? Well, if using the same process, a lower 1.8 v voltage surely won't help increase the speed - overclockers know jolly well that, to increase speed, voltage needs to go up, not down. Stepping down from 2.5 v in DDR to 1.8 v in DDR2 does reduce power and heat, and improve reliability, however it also reduces the speed (inclusive latency) of the memory array - so 4-bit prefetch was needed to even get DDR2-400 data rates, not to mention future DDR2-533 etc.
Price versus performance
This year, the DDR-400 is still a stable, affordable standard on both Intel and AMD platforms, with quite a choice of various capacity and speed options for both desktop and workstation/server systems. I believe that even 2 GB registered ECC DDR400 DIMMs will be soon available, though not at low latency initially, enabling you to stuff 32 GB of very high speed RAM onto your 64-bit quad-Opteron server in that Tyan S4882 (where is the workstation version?) board - it could make quite a nasty UT2003 multiplayer server, probably.
DDR2-400 is just the initial, low-speed "testbed" run of the new standard, and it can't beat its elder DDR cousin, neither on price nor on performance. Remember, any power savings you may experience with DDR2 are more than offset by monstruous consumption of those Prescott or Nocona CPUs that will read or write to that memory. Yes, the new protocol is more scalable to higher speeds like 533 or 667MHz, but we may have to wait for more RAMs on 90 nm and even 65nm process to give us reasonably low latency on those frequencies. That is where the new platform should hopefully show us its full shine, unless of course Rambus XDR RAM becomes more widely accepted - put aside all the quirks within and around Rambus Inc, the XDR technology just kicks a**.
On the other hand, DDR2 might be just what the doctor has prescribed for getting higher memory bandwidth on power-sensitive notebook platforms, where DDR400 would just consume too much. Let's see how it performs in the upcoming Dothan Pentium-M mobiles. µ