Rambus DRAM Part 2: Performance

The whole point of the first Rambus article was to attack the topic of memory bandwidth, but we failed to mention a very important characteristic of memory that would determine how effective RDRAM would be in terms of real world performance: latency.

It has been known from the start that RDRAM has a higher latency (the time before data transfer can actually occur) than SDRAM and definitely than DDR SDRAM.  However, recently Rambus has made the argument that RDRAM actually has a lower latency than SDRAM.

Since today's applications, games and benchmarks aren't entirely memory bandwidth limited (disk transfer speeds, bus, processor and other such limitations come into play before memory bandwidth becomes the primary limiting factor) latency becomes a much more influential factor in the performance of a particular type of memory. 

So who is right?  Does RDRAM have a higher latency than SDRAM, and are most hardware enthusiasts right about the technology or does it have a lower one and has Rambus been telling us the truth all along? 

The answer is, both.  For the longest time Intel has admitted that RDRAM has a higher latency than SDRAM and that the Apollo Pro 133A with PC133 SDRAM actually features a lower latency than the i820 chipset with PC800 RDRAM.  Sound hard to believe? 

Take a look at the following slide from an Intel presentation:

This is where the latency argument begins to make sense.  Remember the 3.7GB/s of memory bandwidth we calculated in the first article as being the amount of memory bandwidth necessary for the next-generation of PCs?  Well, we're not at that point just yet, in fact, we're only beginning to reach the limits of PC100 SDRAM's 800MB/s in our everyday tasks. 

At the start of 1998, we had just begun to saturate the 66MHz memory bus offered by Intel's LX chipset, which was part of the reason that when the transition to the 100MHz FSB/memory bus occurred, there was relatively no performance difference between the newly released Pentium II 350/100 and the older Pentium II 333/66 other than the added performance provided by the increase in clock speed.  It was also evident that we hadn't really begun to saturate the 66MHz memory bus because the 66MHz FSB Celerons were performing quite comparably to their 100MHz FSB Pentium II counterparts. 

Now look at the performance of the 66MHz FSB Celerons in comparison to the 100MHz Pentium II/IIIs today, with today's benchmarks, there is a much more noticeable performance difference: