Ibexx and Thread,
I'm surprised no one has posted this article on Ramtron. It is full of negative technical talk about Rambus. What do people here think...?
Colorado Firm Claims Real-World SDRAM Speed Record 07/16/98
Newsbytes, Thursday, July 16, 1998 at 14:36
COLORADO SPRINGS, COLORADO, U.S.A., 1998 JUL 16 (NB) -- By Craig
Menefee, Newsbytes. Ramtron International Corp. [NASDAQ:RMTR] says
recent announcements about Rambus Inc. [NASDAQ:RMBS] memory speed
records are impressive but have little to do with the real world.
The firm, through its Enhanced Memory Systems Inc. subsidiary, is
shipping engineering samples of, what it claims is, the fastest
real-world DRAM, a standard-interface chip termed enhanced
synchronous dynamic random access memory (ESDRAM).
Ramtron says the new memory chips have been clocked at 200 megahertz
(MHz) memory burst speeds that make it suitable for use on a 200MHz
bus, faster than any other SDRAM memory device now available. Current
fast systems use a 100MHz data bus with memory chips that can only
clock up to 145MHz or 166MHz, according to Ramtron. That makes
Ramtron's sustainable data handling speeds the fastest now available
in the industry.
Asked why near gigahertz (GHz) data throughput of memory based on
Rambus technology, announced July 15 by Samsung, does not far outrun
the Ramtron chip, spokesperson David Bondurant told Newsbytes it
comes down to latency and real-world conditions.
He explained, "In the press, you almost always see a simplification
of memory speed as being how fast data is clocked out of the chip.
Samsung's announcement only describes data speed -- how much can be
moved out of a pin in memory, how fast, after it has been accessed.
It does not take into account how fast a word can be accessed in
random memory. The term for that is 'latency,' and no one has better
latency speeds than we do."
Bondurant says the latency factor causes Ramtron to advertise its
chips as the fastest DRAM in the "real world," not just in a
specialized laboratory test. He says "bursting" of sequential data
through a pin, used in specifications issued by Rambus, does not
address latency at all. But latency, he says, has a far greater
impact on how fast a system can actually operate.
"We think this kind of gigahertz hype is not the real world," he told
Newsbytes. "I will submit that it is the fastest data rate per pin,
but that's peak data rate. It doesn't say anything about what you can
do in a real system."
He continued, "What we claim is that we're the fastest SDRAM in the
real world, and we are. Other manufacturers including Samsung can
deliver SDRAMS with clock speeds as high as 143MHz, or maybe even
166MHz. What we've announced is a 200MHz version of SDRAM. It has
absolutely the fastest latency in the world."
In technical terms, Bondurant said Ramtron ESDRAM latency figures
come down to 22 nanosecond (ns) row access, 10 ns column access, and
4.3 ns clock access. "Row access" is random access to any major section
of a DRAM, "column" is how long it takes to access any location within
a page, and "clock access" is how long it takes on a clock to find
something in random memory and send it back out.
At 4.3 ns clock access time, the Ramtron device handily exceeds the
five ns figure that would be required to run on some future 200MHz
memory bus.
Bondurant acknowledges that low latency has less positive effect on
high-end systems with large amounts of on-processor or L2 cache. On
such chips, what happens is that, when the processor finds some word
of data in RAM, it reads that word plus the next three sequential
words into L2 cache "in hopes they will be what the processor needs
next."
Much of the time the system works on such chips, reducing the
percentage of reads that come from non-cached, randomly accessed,
latency-delayed memory.
However, says Bondurant, in both notebooks and price-driven
sub-$1,000 machines, one way both costs and power demands are cut is
to shrink or eliminate L2 cache and other high-end features. As that
happens, latency takes a bigger and bigger toll on performance. It
simply takes longer to access what is in memory without a huge L2
cache, no matter what speed the processor advertises.
How often does latency make a difference in the real world? It
depends on factors like total system architecture, says Bondurant,
which is why memory companies advertise the numbers that make them
look the best.
"We quote 22 ns row access, which assumes the system takes its data
from random memory 100 percent of the time," Bondurant said. "The
Rambus equivalent number is 56 ns, almost three times slower. But
Rambus advertises peak chip throughput as if a machine never has to
access memory randomly at all."
Bondurant says a realistic figure for RAM rather than cache access,
absent all the intentional hype, is probably around 75 percent on an
average system. That figure, he says, would make a system based on
Ramtron ESDRAM chips about 30 percent faster than the equivalent
Rambus system.
Bondurant recommends surfing the World Wide Web over to Tom's
Hardware Page at tomshardware.com for a deeper
understanding of actual memory issues. There, Bert McComas, principal
analyst at market research firm InQuest, is running a series of
technical articles on the subject. Ramtron itself has a page on the
World Wide Web at ramtron.com .
Reported by Newsbytes News Network: newsbytes.com .
(19980716/Press & Reader Contact: Lee Brown, Ramtron, 719-481-7213
/WIRES PC, BUSINESS/) |
