Low Latency DRAMs Continue to Serve Networking Niches

Low Latency DRAMs (LL DRAMs), Survive to Serve an Important Market Niche:

Last month, I was visiting Fujitsu in Japan. Fujitsu, as you may recall, was the innovator of the Fast Cycle (FC) RAM in the late 1990s, one of several Low Latency (LL) DRAMs that rode the wave of the networking bubble 7-8 years ago. The 'Networking' FC RAMs, with densities up to 576M, offered ECC, fast latencies down to 15-20ns (much faster than traditional DRAMs ~50ns latencies), and in most applications, were a single chip substitute for 12-20 pieces of medium-speed low-density SRAMs. Also, the FC RAM architecture was, and is today in its lower density embodiments, a popular “DRAM” option for cell phones, due mostly to its low-power operation, which derives from its highly-segmented array structure, in the same way that its improved latency benefits from such an architectural choice. But the Networking FC RAM was 288Mb and up, with a density void between it, and the 16M to 64M 'phone' FC RAMs.

However, Fujitsu abandoned the 288M FC RAM market for networking in 2004/5...along with Samsung (Networking DRAM), Toshiba (FC RAM), and Infineon's customized version, the Reduced Latency, or RL DRAM (RL I), and its successor, RL II DRAMs. Infineon eventually sold its RL DRAM business to Micron, who still makes the RL I and RL II DRAMs. The leading edge for RL DRAMs today is a 95nm process, and 576Mb densities, with latencies down to 15ns, which was launched about 18 months ago.

In addition to Micron, which makes the RL DRAM family, NEC, it turns out, makes the Low Latency DRAM that Fujitsu used to make. If one looks in their website, you can see their 288Mb Low Latency DRAM, with 15ns latencies. NEC jumping into the Low Latency DRAM market acted to reduce the impact of the one-a-day departure we all witnessed several years back, which left many loyal LL DRAM users high and dry, with only one source, and in need of board redesigns to use the sole surviving RL DRAM.

The Low Latency DRAM was a good idea at the time, and still is, and despite its +/- $35 price tag for 576M, is still way cheaper than the alternative system solution, which uses bundles of semi-fast SRAMs. We gave a memorial essay on the LL DRAM market when the last of the suppliers (exc. Micron) passed into oblivion in 2005-06.

Not that it is extremely profitable for Micron or NEC. Although the gross profit margins might be high compared to DRAMs (What isn't these days?), the market is still tiny and restricted to a few classes of networking applications. The entire design cost has to be recovered with only a few millions, or tens of millions of product lifetime unit shipments. It is truly an ASM = Applications Specific Memory, and one of the best.

Cisco was originally, and seems today, to be the most interested party in staying on top of the RL/FC/LL DRAM business and technical roadmap; personally, I blamed them most for allowing the original roadmap to come apart, with too many vendors, too many distinct products, too little support and too little user understanding of what the Low Latency DRAM maker faced, from a business perspective. (Cypress once told us in an early DMR interview, that they often had to do two or three shrinks of a new network SRAM design before they got their first volume order...the first rev was for ‘user look and see,' the second rev for qualification volumes, and third rev was for production volumes into the system...networking equipment ALWAYS is big boxes with long design cycles, and even longer time in use...making for another headache at the other end of life...Keeping a part in production for 12-15 years is not something chipmakers usually want to do, especially memory makers, though some 64M SDRAMs are now in their fifth shrink since 1998, and have moved from 450nm to 120 or 90nm designs.)

For LL DRAMs, let’s hope this time it proves to be a sustainable market. For sure, what was a good idea in 2001, at 180nm, in 2006 at 95nm is a better idea; and at 60nm and below today — it is even better. What’s not to like, more density, better power performance, and immeasurably more understanding of the market to drive the design roadmaps.

So, Low Latency DRAMs live on, still in some varied forms from a few makers, and are likely to continue to do so, as long as their long-lived host equipment remains in production or use.

And as a final footnote to this decade-long story, though abandoning the networking 256M/288M FC RAM, Fujitsu remained loyal to its 16M and 64M phone FCRAM customers...until just recently. But even that has now come to an end, and they have licensed several Taiwanese companies to build and take FCRAM to market, and are halting their own production of the FC RAM family of parts, thereby leaving themselves as only a ‘DRAM innovator’ and today, a technology licensor.