----- Original Message ----- From: "Benjamin Scott" <[EMAIL PROTECTED]> To: "Greater NH Linux Users' Group" <[EMAIL PROTECTED]> Sent: Wednesday, April 24, 2002 2:14 AM Subject: RAMBUS (was: AMD vs Intel)
> On Tue, 23 Apr 2002, at 1:02pm, Rich C wrote: > >> There is also the latency issue, which I have yet to have confirmed > >> or denied to my satisfaction. > > > > There IS higher latency with RAMBUS ... > > I stated that poorly. What I should have said was that I have not seen an > analysis of the latency issue that provides sufficient information for me to > draw conclusions that satisfy me. Me either. The only thing I have seen so far is that the new DDR400 RAM theoretically has the same memory bandwidth as dual channel PC800 RDRAM (3.2 GB/sec), and holds a slight edge in memory performance benchmarks: http://www.sharkyextreme.com/hardware/motherboards/article.php/10703_999 491__6 http://www.xbitlabs.com/mainboards/sis645dx/ (one big page) Most "experts" I've read attribute the difference to the higher latency of RAMBUS architecture. > > > However, RAMBUS is serial and capable of much higher speeds, which can > > overcome the latency issue. > > For example, the above statement is dubious. Simply increasing the data > transfer rate does not automatically solve latency concerns. > > Allow me to illistrate with an extreme example: Say you want to hold a > conversation with someone in another star system. Even if you had a data > transfer rate of twenty terabytes per second, it would still take years for > you to greet each other. > Your example is flawed. Increasing the bus speed of the memory is analogous to increasing the speed of light in your example, which would indeed improve things. Only when we get to FSB speeds of tens of gigahertz will we run into the propagation "brick wall" of electromagnetic energy in a copper medium. > > Just because a system has higher latency doesn't necessarily mean it's a > > bad design. > > Likewise, just because a system has higher data transfer rates does not > necessarily mean it is a good design. I am not saying that the higher > latency of RAMBUS kills it, but rather, that the higher data transfer rate > does not automatically make it better. > I did not say that higher data rates were the only thing required to overcome latency issues. Such things as interleaving and data prediction (prefetch) can also be used to overcome latency issues. (I mentioned this in my comment on CPU internal latency.) Even in your [flawed] example, with these improvements, and a high enough data rate, we could pass back and forth all the data our respective civilizations would want to know in one transfer cycle. Don't get me wrong, I'm not saying that RAMBUS by itself is a good idea, or that there aren't better ones out there. However, it is only recently with the advent of 2+ gigahertz processors with large caches that the available memory bandwidth of RAMBUS is being fully utilized. And only recently with the advent of DDR400 is the latency of RAMBUS even becoming an issue. It will be interesting to see which technology forms the basis for the new technologies we will see for faster memory systems. Rich Cloutier President, C*O SYSTEM SUPPORT SERVICES www.sysupport.com ***************************************************************** To unsubscribe from this list, send mail to [EMAIL PROTECTED] with the text 'unsubscribe gnhlug' in the message body. *****************************************************************
