Quote:
|
Originally Posted by stabiz
Oki-doki.
|
Trust me. You never hit 6.4 GB/sec bandwidth for system memory (maybe a file server of some sort might, but not a desktop). What you transfer is a ton of tiny files while your system is in operation. For a large texture transfer such as that found on video cards, higher speed ram is important. For system memory, latency actually is more important simply because each transcation has to wait a latency time before it is able to read and write. So a DDR2 module that has a latency time of say 5 ns, has to wait 5 ns each and every time a transaction is done before it can start completing. A DDR module at 2 ns may not have the throughput of a DDR2 module, but the DDR module is already transfering 2x to 2.5x faster than the DDR2 module.
So for the typical system that transfers millions of tiny files through memory all day long, it doesn't take rocket science to figure out that latency is going to outperform throughput in a scenario like this.
On a completely different subject, your computer never needs 6.4 GB / sec (The speed of stadard DDR2 at 800 MHz, and this is the exact same speed of DDR in dual channel mode at 400 MHz). You computer will never utilize this throughput already - in testing, maybe half of this speed is utilized on a constant basis. System memory is following video card memory, and always has, so maybe they are just making us buy new memory for a new sale without benefit to us. Actually, this may be a negative benefit to us in normal daily transactions.
This makes me question why we had DDR2 shoved down our throats?
Does anyone remember RAMBUS? RAMBUS is like DDR2 in performance, yet back when the heat was on between the competing formats, and all the testing was done, you always had the lower latency DDR outperforming the higher throughput RAMBUS??? Same deal today, but I guess the march of technology must move forward, even if its not in a positive sense.
[...gets off soapbox....]
-S
PS. I just figured it out - this is being pushed because some boards use IGP's with shared memory. This is being pushed on us to be compatible with those, and still make it easy on the consumer to figure out what the proper memory to buy for their system. No other reason I can think of. Video cards need throughput over latency for large textures, so IGP's can perform OK on system RAM if its got the throughput.