Found the following on TheFool written by jasonxsmith and worth reading...
It has been a while since I got basic, so here goes. I did all of this from memory, so
please be kind if I messed up a fact or two. I think it is essentially all correct though. This
is a bit long...
Can someone kindly explain the differences between RDRAM and SDRAM to a
newbie?
SDRAM
SDRAM is "synchronous dynamic RAM". It is basically DRAM with some synchronous
registers thrown in to allow it to transfer data faster.
The fastest SDRAM is PC133. It has a 133MHz clock and transfers one bit of data per
pin per clock cycle, or 133Mb/s/pin. With 64 pins dedicated to data, and a bunch of
control and address pins, you get a maximum of about 1064 MB/s throughput. A single
memory access is distributed over a number of different devices, which all return a little
piece of the data.
DDR
DDR is supposed to be the next generation of SDRAM, but it has encountered endless
delays. It was supposed to have captured 10% of the market by now, but we have yet
to see a motherboard or chipset that can handle it.
DDR comes in PC200 or PC266 flavors, or at least it will, presumably, some time.
Whenever. It has the same clock speed as SDRAM, but it transfers two bits per clock
cycle. This is a really good idea! However, it was stolen from Rambus intellectual
property. Bad idea. DDR, like SDRAM, distributes the memory access amongst
different devices.
RDRAM
RDRAM has a 400MHz clock, and it transfers two bits of data per pin per clock cycle
(assuming the fastest RDRAM - PC800). This is 800Mb/s/pin. With 16 data pins and a
few address and control lines, a single channel of RDRAM gives you 1600 MB/s
throughput. A single memory access is handled by a single device, and other devices can
be powered down when they are not being used. Because RDRAM uses far fewer
signals than SDRAM for an equivalent bandwidth, you can use more channels which
allows multiple concurrent memory requests. Having multiple channels increases
bandwidth to 3.2GB/s for 2 channels or 6.4GB/s for 4 channels, and each channel can
handle a separate memory request.
Costs
RDRAM is currently more expensive than SDRAM, but the prices for RDRAM are
dropping quickly and we expect them to be a small part of the cost of a system in the
near future (current differences in Dell systems range from $70 to $100 estimated for an
identically configured system with 128MB DRAM). The price differential is due mainly
to the fact that RDRAM is a new technology with different packaging and testing needs
(higher frequency), and that the die size for RDRAM is currently a bit larger than for
SDRAM. The die size is supposedly going to be 5% to 10% larger than SDRAM very
soon, and the packaging and testing technologies are maturing rapidly.
We know the current price of DDR, but we have no idea how much more expensive
DDR is going to make the computer it goes into. We won't know until someone actually
delivers a DDR system.
Design Philosophies
SDRAM is designed with the old lumped-parameter digital design assumptions. This
makes it difficult to increase the speed beyond the current best 266MHz DDR
(supposedly next-generation SDRAM), which is not yet in production systems. The
DDR specifications do not take trace impedance, matched trace lengths, connector
parasitic effects, and other high-frequency design elements into account. The design they
are using now, in fact, runs counter to these problems. The net effect is that DDR cannot
get much faster than it is now, no matter how fast you make the silicon. The problem is
not the chip, it is the communication between chips.
RDRAM is a radical departure from the SDRAM/DDR design. It still uses DRAM as a
foundation, but it is designed as if the digital signals are microwave signals travelling along
a transmission line. In contrast to the DDR design, which is limited to the current
bandwidth specifications forever, RDRAM can continue to get faster and faster as the
manufacturing processes improve. The current best we know of is 1066Mb/s/pin (or
2132GB/s), and I am quite sure there is more to come in the near future. The PC1066 is
already planned for near term production in small quantities.
At this time, it is questionable whether or not DDR (PC200, PC266) can be made to
work at all in PCs. One thing is for sure - if PowerPoint presentation made stuff work,
DDR would have been in PCs early last year.
Latencies of DRAM
RDRAM, SDRAM, and DDR all have one major limitation: latency. They are all limited
by the latencies inherent in DRAM itself. There is nothing you can do to make this better.
The latencies for RDRAM are slightly better than those for DDR and SDRAM, and
RDRAM latencies are much better at high load factors, but it is a problem in all forms of
DRAM, period. Fortunately, the CPU designers can mitigate most of the effects of
latency with a good design.
Granularity
One of the really cool things about RDRAM is that you can use it easily in embedded
devices. It has a really small form factor and only requires a single chip to provide all the
bandwidth of a channel. The new Playstation II is a good example of this. It uses 2
RDRAM chips (2 Rambus RDRAM channels) for a total of 32MB of DRAM (of
course, the latest chips are much denser than this) and 3.2 GB/s of memory system
bandwidth. That is as much as a P4 is going to have. They would have had to use 8
DDR or 16 SDRAM chips for the same bandwidth, and they would have had much
more memory than they needed.
Sony is using the same core as in their PSII to make a high-end graphics development
machine. They are using something like 32 RDRAM channels for an ungodly 51.2GB/s
bandwidth in a highly parallel system (I might be off by a factor of 2 here - somebody
check my memory). They are using one RDRAM chip per channel. Let's see somebody
do that with DDR or SDRAM! In fact, no one would because the cost of the system
would be prohibitive.
Rambus recently announced a new QRSL technology that allows a bandwidth doubling
from 1.6Gb/s/pin to 3.2Gb/s/pin. Doubling their highest speed device using this
technology, in fact, would get them to 4.2Gb/s/pin, or about 8.4GB/s for a single
channel. This technology will not work for PC main memory, but it will work beautifully
in embedded devices where only a few devices are ganged onto one channel. This would
essentially allow the graphics display in your handheld phone to rival or exceed the best
graphics cards available in today's PCs (it would need one chip for memory).
The Bottom Line
This is kind of getting away from the original question, but I like to remind people of the
numbers every time I get the chance.
Rambus gets about 2% from every sale of an RDRAM chip. They make about 3% to
5% on the sale of every controller chip that interfaces to RDRAM (the cost of chipsets is
a minor concern though). This is a given. The contracts are in place already.
Rambus also claims they have patents which apply to SDRAM and DDR. The royalties
are about 1% for SDRAM and 3% for DDR (since that is considered to be a competing
technology). This is still in dispute, although it looks very much like everyone is going to
have to pay. Rambus appears to be very confident that they can win in court, and have
already moved towards litigation in several instances. Hitachi, Toshiba, and OKI have
agreed to pay the royalties, and Rambus is either suing or in talks with Infineon, though
we are unclear on the progress there.
Rambus stands to make 1.5% to 2.5% of the revenues on 100% of the DRAM sold
worldwide. If the industry grows as expected for the next 5 years, that number will be
somewhere in the $60B to $100B range. Run the numbers, compare that to a $7B
market cap, and rejoice my child, for you have found one of the most undervalued
stocks in the market!
Remember, this is based on DRAM sales alone, and this information is pretty easy to
quantify. Now add to that the other applications where Rambus technology is going to be
necessary, and the picture gets even better. How much better? Nobody knows, except
insiders, and they are not talking. Rambus is working deals now.
Risk
There is a risk that Rambus could lose all its patents when it goes to court. This would
make the company worth virtually nothing, but it would also call into question every
patent issued in the last 17 years. If the patents are not invalidated, there seems to be
virtually no risk that the memory makers will find a way around the patents. Their best
attempt was DDR, and its design isn't really all that good to start with.
Intel is still stating publicly that RDRAM is the way of the future and the major memory 5
years out. There is a risk that this could change. If RDRAM suddenly becomes very
difficult and expensive to make and no one can find a workaround, DDR may find a
second wind. If we get 3% of DDR sales, who cares, right? But DDR is a stop-gap
measure, and if RDRAM does not work as a mainstream solution, another will have to
be found - and quickly. So far, I have seen absolutely no evidence that RDRAM will not
be ready, and cheap, by the time it is needed for value systems. I am getting very
worried about DDR however. :-)
Oh, the FUD!
There is more misinformation on Rambus than there is information. Take a look at the
posts of the past week if you want to know more. Whenever a company has a position
as strong as this one, it scares the bejeebers out of the big, established companies, and
they do everything in their power to diminish the public perception of the newcomer. This
is going to cost them, and they know it. And RDRAM is not a passing fad, but a force to
be reckoned with.
You can tell FUD by:
(1) Anything at all from Tom's Hardware...
(2) Any article that has a quote from anyone at the Semico Marketing...uh, research
firm...
(3) Any article that logically contradicts itself more than twice in one sentence...
(4) Anything that quotes unnamed "inside sources"...
(5) Any article that claims a small company can be sued for abusing its monopolistic
position, usually quoting some poor broken slob from a multi-billion dollar manufacturing
company. You actually have to be a monopoly before you can be reprimanded. And
patents are a government sanctioned monopoly, so who ya gonna sue, the government?
"2% royalties! They are raping us! Now we can only make 10% profits on every sale
instead of 12%! Boo hoo! We're sure the public would rather have EDO than pay 2%
more for a stick of RAM with 16 times the performance! Whine, whine..."
(6) Any article that claims to be about benchmarks and then quickly degenerates into
how evil Rambus and Intel are (and our friend Tom is not the only one to do this).
Politics do not belong in benchmarking articles.
There. That is it. Everything I know, in a nutshell. Good luck. |
