October 26, 1998
Whatever happened to Rambus?
By Bob O'Donnell
There's no question that computers of the future will be faster than today's models,
but it isn't always clear where the speed boosts will come from. Most people concentrate
on processor power and point to the ever-increasing megahertz rating of today's
microprocessors as the primary reason. But the reality is, we aren't seeing overall
computer system performance match the speed gains that faster chips imply.
The reason, of course, is that computers are complex systems, and simply improving the
performance of one component -- albeit a pretty important one -- doesn't lead to a
directly proportional improvement in overall performance. Several bottlenecks still exist
in today's PC systems, one of the most important of which is memory. Today's processors
are so fast, in fact, that they often sit idle waiting for more instructions and data to
chew on. The increasing size and speed of Level 2 (L2) caches have certainly helped, but
getting access to data in memory and transferring that data from memory to the L2 cache
(and eventually to the processor) is still a big stalling point.
Numerous companies have developed technologies to speed up the transfer of data from
memory to the processor, but none has received the splash (and industry acceptance) of Rambus Direct R-DRAM. The basic
concept behind R-DRAM is to increase the bandwidth between memory and the processor by
offering a narrow but extremely fast channel that connects the two. The result is that
more data can be transferred between memory and the CPU, thereby helping to alleviate an
important bottleneck in overall system performance.
As a result of this promise, Rambus and its eponymously named technology were widely
hyped as the definitive memory technology of the future, particularly after Intel
announced it would support Rambus in next-generation chip sets and PCs. (Like all other
memory technologies, R-DRAM support needs to be built into the chip set before it can be
used in any PC.) The problem is that, as with many other forward-looking technologies, it
has taken a long time to go from concept to reality. And, in fact, we're still not there
yet. According to Intel's most recent timelines, we can't expect R-DRAM in desktop systems
until sometime in the middle of 1999, and even later than that for portables.
Of course, Intel isn't the only chip set vendor, and just a few weeks ago, AMD
announced that it would add support for R-DRAM into its chip sets for the K7, which are
due around the same time as the Intel offerings.
Even when Rambus does arrive, there are some concerns about how large an impact it will
actually have. First, supporting R-DRAM will be an additional cost for system vendors, and
given the plummeting price of average PC systems, every dollar counts. Also, although no
one can know this for sure, Rambus itself is placing a lot of emphasis on making sure
enough R-DRAM RIMMs are available. It is not difficult to interpret this as a concern as
to whether there really will be enough when the time comes.
Finally, and perhaps most importantly, it's not entirely clear how large the real-world
performance benefits of R-DRAM will be. Again, it's too early to be sure, but although
R-DRAM can offer up to a three-times improvement for applications that require reading
large streams of data from memory, it may have a negligible impact on day-to-day
applications. Part of the reason is that although R-DRAM improves the speed at which data
can be transferred, it doesn't change the initial access time, or latency, required to get
to that data in the first place. And that, in many cases, is where the real problem lies.
In fact, in that regard, R-DRAM is really no different than current synchronous DRAM,
or SDRAM. They offer similar initial access times, according to published specifications.
The reason is that the memory core, which determines how fast you can first get to that
data, hasn't really changed in quite a while.
However, there have been some technological developments in this area as well. NEC's virtual
memory channel technology basically adds a small portion of superfast static RAM (the
type of memory normally used in L2 cache, also called SRAM) to a standard DRAM, thereby
improving the potential access time of standard memory. (NEC was recently sued by Enhanced Memory Systems for patent infringements on this
technology.)
What is nice about virtual channel memory is that it can be used along with any memory
technology, including R-DRAM. The problem with it is that it adds both cost and complexity
to memory design, and it, too, will require widespread chip set support to really have an
impact. Ideally, a memory subsystem that combined the wide bandwidth of R-DRAM and the
potentially fast access time of virtual channel memory (or other similar technologies)
could lead to great improvements in overall system performance.
In the meantime, however, we'll just have to wait.
©
Copyright 1998, by InfoWorld Publishing Corp., a
subsidiary of IDG Communications, Inc. Reprinted from InfoWorld,
155 Bovet Road, San Mateo, CA 94402. Further reproduction is prohibited.
