Re: THOUGHT WE SETTLED THIS LONG AGO...
We're all just yakking about this. It will be settled by the engineers who design and produce these things - not by us on this list :-)
The issue of performance has been moot for a while while the big question was whether or not rambus could be made to function at all.
My argument has been that even if it did function, its performance was worse than alternative technologies that cost less. At one time, it seemed that Intel was going to be able to force everyone to use Rambus regardless of price or performance so this didn't matter.
The recent problems and delays have changed all that. Systems supporting virtual channel memory have shipped BEFORE rambus computers - not months after rambus was expected.
Rambus has been dropped for servers and dropped by Nintendo and replaced by DDR. DDR systems will be available very shortly after rambus systems - the time to market lead promised to rambus by Intel is gone. Intel has been so embarassed by the whole rambus program that it is no longer telling the Industry it HAS to use rambus. And since everyone but Intel hated the idea of being locked into a proprietary technology that required royalties to be paid, this means that rambus is going to be severely limited in its ability to gain market share through Intel pressure.
Take a look at the famous Samsung white paper praising rambus at:
usa.samsungsemi.com
Read their article carefully. They compare rambus 800/40 (rambus comes in latency binsplits as well as MHZ binspits) to PCXXX. There are few 800/40's out there, but use that part anyway.
The paper shows PC133 total latency at 75ns and Rambus at 70ns. But, instead of comparing best PC133 (CAS 2) with best Rambus, they have compared the Rambus to CAS 3 PC133. If you use CAS 2, you save 7.5ns. So the PC133 fills the cache line 2.5ns faster than the Rambus 800/40.
Now consider DDR 266, doubling the data rate cuts another 15ns off the time required to fill a cache line leaving total latency at 52.5ns for DDR266 and 70ns for rambus - a difference of 17.5ns. On the 64 byte cache line used by rambus and (I believe) Itaniam, rambus will do better against PCXXX but DDR will increase its lead over Rambus to 110 - 82.5 or 27.5ns
This low latency is particularly important for multitasking machines like servers. I think this is why the server manufacturers have pretty much en mass decided to go with DDR, and Intel will too.
Since DDR SDRAM costs less than rambus, and outperforms it, I doubt that rambus will ever be much of a force in PC memory.
Regards,
Dan |
