SHARKY EXTREME
Another area of high-speed RAM demand lately is in the competitive realm of the video accelerator industry. The various firms that design and implement video accelerators have pushed new chip designs and technology at a pace that's been much faster than most of
the other peripherals equipped in current PCs. Sharky Extreme has heard moans
and groans recently from video card manufacturers who place their own
next-generation product shipment delays at the feet of DRAM
manufacturers, stating that the DRAM simply isn't up to
specification.
What unique hurdles does a video accelerator chip
challenge DRAM with in comparison to a standard CPU, and why? Can
RDRAM be applied to video accelerators with the same speed and
capabilities that it will be applied to PCs with?
MICHAEL SPORER
Main memory runs at a speed limited by the interconnect between the
chipset and the RAM. This interconnect has to incorporate at least 2
printed circuit boards, the DIMM socket as well the variability of a
system with minimum or maximum installed RAM (which affects the
capacitive load on the data bus.) Rambus I/O technology addresses these
problems.
In a graphics card implementation the interconnect is very short and
simple (usually without expansion sockets.) So the I/O can be precisely
tuned to the graphics chip and RAM. From the RAM manufacturers
perspective you have a device capable of running at 125MHz (-8 speed)
which due to I/O can only run in a system at 100MHz (PC100). Then the
graphics guys come along and want to run the same level of technology at
166MHz. Granted, the graphics memories are optimized for their
application, but the fundamental core of the DRAM will limit the maximum
performance.
In order to work around the slow graphics memory problem the chip makers
are using 128 bit datapaths. This causes two problems, first the high
pincount is expensive and difficult to implement with short interconnect
distances. Second, the total memory size is rather large due to the
number of memory devices required. It's not easy to implement a low
memory size, high performance graphic card with traditional memory.
Rambus in graphics offers much more bandwidth per pin, so graphics cards
can be made higher performance with fewer pins and fewer memory chips.
The current Rambus specification is designed for main memory I/O
(modules and connectors.) It is conceivable that a special 'graphics'
specification can be developed for higher clock rates. Our RDRAM
products fully meet the 800MHz specification, I might remind you that
our PC100 SDRAM from last year was a popular choice for overclocking and
our new revision SDRAM is almost exclusively PC133 capable.
SHARKY EXTREME
Indeed, the new LG PC133 SDRAM we've been testing the past two months has performed stably at that speed without issue.
Our last question for today concerns progress. What future technologies do you personally look
forward to in the evolution of DRAM? What barriers cap our ability to
progress quickly towards these goals, and when, if ever, will we see
embedded DRAM applied to PCs?
MICHAEL SPORER
Embedded DRAM is the holy grail for semiconductor manufacturers. It has
been the 'next technology' for many years. Embedded DRAM is inevitable
and the longer it remains the 'technology of tomorrow' the more likely
it will actually be tomorrow (rather than the following day.)
Embedded DRAM has been in PC's since 1994. Check out a company called
NeoMagic. They offer a small memory solution that's really high
performance and low power. They were the first, now more and more
companies are offering this technology.
This concludes our interview with Michael Sporer. We'll be revisiting Michael with questions generated by the readers towards the goal of producing a "Part 2" to this initial conversation soon.
Thanks go out to LG Semicon America and Michael Sporer for finding the time to speak with Sharky Extreme.
Craig "Mako" Campanaro
Hardware Editor
