The DDR-II is coming! Not really as menacing as 'The English are coming!' but sure to strike confusion into the minds and panic into the wallets of computing enthusiasts everywhere.
You see this new memory is certainly not far off with chipsets supporting it lining up for release in the coming months, yet surprising few people seem to really know much about it or what benefits it may bring. There are relatively few informed articles on the subject, the only one which really comes to mind being Lost Circuits eight page monster which while full of information is hardly what you could call light reading.
The basic jist of things is that DDR-II can offer double the overall memory bandwidth of DDR memory for the same speed module by effectively using two DRAM cores per device instead of one as used by DDR. Technically there is still only one DRAM core but it is accessed in parellel allowing it to deal with 4 data pre-fetches instead of two. This combined with data buffers running at double data rate means that for each clock cycle up to 4-bits of data can be dealt with instead of the 2-bits regular DDR can handle.
The trade off for this increase in bandwidth comes in the form of increased latencies, the number of clock cycles certain operations take to be carried out.
CAS (Column Address Strobe) latency, the number of cycles it takes for a column in memory to be selected, is expected to rise from 2 cycles as seen on current low latency modules and 3 cycles as used by very high bus speed DDR to most likely a minimum of 4 or 5 cycles for DDR-II. Similarly tRAS, the delay before being able to select a new row in memory, is likely to increase with 8 cycles looking to be the lowest it will be able to go, the same as is currently used on high speed DDR with some low latency DDR being able to run 5 cycles for tRAS.
Other changes for DDR-II include a drop in voltage to 1.8v, resulting in reduced power consumption, On Die Termination and Variable Write Latency, all of which are covered in Lost Circuits article. The end result however is basically that of increased bandwidth at the expense of latency.
Because the speed of the DRAM cores themselves are the same as DDR of half the bandwidth costs should be reduced, however in order to use DDR-II a new motherboard will be required because DDR-II is incompatible with DDR memory, due to the increased pin count, 240-pins instead of 184-pins for DDR, and the lower VDIMM voltage, 1.8v instead of 2.5-2.8v for DDR
So what is DDR-II going to do for our processors?
Well for AMD at the moment its looking like being a case of absolutely nothing. The Athlon 64's integrated memory controller means that all current chips will be totally incompatible with DDR-II and it seems unlikely that AMD will be in any hurry to do anything about that.
The efficiency of the Athlon 64's architecture means that at present it is not in need of the kind of bandwidth that DDR-II is able to offer and the increase in latency would most probably lead to a decrease in performance, especially considering how one of the Athlon 64's great strengths is its ability to complete memory operations in far fewer cycles than Athlon XP and Pentium 4 systems.
For proof that the Athlon 64 does not need the same kind of memory bandwidth as the Pentium 4 you simply have to look at both systems in single and dual channel configurations. The performance gap between an Athlon 64 3400+ using a single low latency PC3200 module and an Athlon FX-51 using two Registered PC3200 modules is surprisingly small and in some cases the 3400+ will actually be quicker since the registered memory modules current Socket 940 chips require incur more latency due to extra buffering, over the unregistered modules of the 3400+. You have a situation where you have double the bandwidth, like DDR-II, increased latency, like DDR-II, and more expensive memory, probably like DDR-II, and performance which is at best slightly quicker, at worst slower.
Now compare this to a Pentium 4 in a similar situation. Take a i865 or i875 with two sticks of memory in it and run a benchmark, now, take one of them out and put it in a different slot, disabling dual channel mode and run the benchmark again, almost without exception the result will be considerably slower. If you apply a divider to run the memory at less than front side bus speed and lower the latencies you also get a fall off in performance.
So for the bandwidth hungry Pentium 4 DDR-II could produce some tangible benefits, especially as processor clock speeds scale towards 4GHz and above. It is going to be interesting to see how the increased latencies affect performance. If the memory accesses are largely sequential reads with minimal switching between columns, such as in video encoding then performance should be significantly boosted thanks to the increased bandwidth, however in situations where the required data is spread through out the modules, such as in server applications, the increased latency is going to come into play and rain on Intels parade.
Because of this I believe it is likely that the path taken by Intels consumer and server market chipsets will diverge with the consumer chipsets taking the route of DDR-II and rapidly scaling clock and front side bus speeds, while Intels server system shall remain largely DDR based. The fact that the forthcoming i915X and i925X chipsets support both DDR and DDR-II means that Intel for now at least are keeping their options open as to which path they are going to take for different product lines. When you take into consideration that these chipsets do not have AGP support and instead use PCI-Express and suddenly even though DDR-II should not cost a great deal more than DDR for the modules themselves, the combination of new memory, motherboard and graphics card could make the cost of transistion between for example a Socket 478 system to a LGA775 (the only CPU interface i915X and i925X chipsets support) system with DDR-II will be considerably steeper than moving to Athlon 64.
Of course its going to be impossible to gauge exactly how these issues have affected things until proper benchmark results for the platform appear, but one thing's for sure, its going to be an interesting couple of months!
Reply With Quote
