March 2011

If Memory Serves


  • Page 1 of 1
    Bookmark and Share

Random-access memory. Can’t live with it, can’t live without it. Whether you’re developing a system around a milliwatt 8-bit micro or a 64-bit superscalar rocket with unearthly onboard cache SRAM, the system won’t boot without the DRAM that holds most of the operating system and application code. Strap it in and prepare for a wild ride.

RAM comes in various flavors—soldered (on the system board), standards-based modules and proprietary modules. RAM chips themselves churn nearly as fast as microcontrollers, microprocessors and chipsets do, with production lifecycles that are much shorter. Yet RAM escapes the blame for single board computer end-of-life (EOL) notices since memory chip vendors do a thorough job of designing pin-compatible functionally equivalent replacement parts. Imagine if the CPU folks did this as a rule, not an exception.

In spite of flashy names and numbers like DDR3 and 800 / 1066 / 1333, memory performance has improved incrementally each year, but has been blown out of the water by more dramatic processor performance increases. It’s no wonder that the nearly free CPU transistors have been used for every core, cache and pipeline trick under the sun to overcompensate for the root cause of throttled system performance.

Single-issue 5-stage pipeline 486 processors have given way to 6-core hyper-threading multi-Gigahertz multi-headed Hydras with on-chip memory and graphics controllers and megabytes of on-chip cache that burst-fill from the DRAM spigot. All this for only a modest thermal design power (TDP) penalty. Many of the old-style integer and floating point benchmarks fit neatly into the cache so that the performance numbers keep goosing new consumer processor sales. 

As usual, the SFF market drafts off higher volume primary markets for standard DDR RAM modules—enterprise and consumer computers and other consumer electronics devices. But the tiniest SBCs and computer-on-modules (COMs) often don’t use SODIMMs. There is a price to pay for soldered onboard RAM in terms of size flexibility, demand forecasts and inventory—not to mention warranty and repair responsibility. But for many of these alternative products, the cost is well worth it in terms of inherent ruggedness and reliability over time in harsh environments.

Extremes of temperature, humidity, corrosive chemicals, dust, shock and vibration can wreak havoc on unsuspecting SODIMM modules that were not designed to handle them. For decades, system manufacturers and integrators in military, aerospace and transportation markets have tried to tolerate these consumer modules with plenty of epoxy (RTV) or straps or clips or brackets. Imagine the puzzled looks on the faces of RMA technicians who thought they’d seen it all.

Many of these users would be willing to pay more for a rugged-by-design memory module if a suitable standards-based solution existed. Over the years, some SFF SBC manufacturers have created modules using pin-in-socket technology or even surface mount board-to-board connectors. These solutions can stand up to harsh environments almost as well as soldered RAM, and can even be conformally coated if each board is coated separately first with the connectors masked. The downside has been single sourcing of these proprietary modules.

As evidence of the resilience and determination of this very fragmented embedded market, a number of companies have been collaborating on various rugged RAM solutions. The most rugged solutions involve the use of board-to-board connectors, as some OEMs are concerned that card edge gold finger-in-socket approaches are vulnerable to gradual micro-etching and contamination and have not yet been proven definitively over many years in harsh environments.

Board space is so limited on SFFs; every square millimeter is used for features. Any time a connector can be eliminated means greater feature density for OEMs who need to reduce size and weight. RAM interfaces are quite wide in order to support the Gbyte/second bandwidth required—64 bits wide for x86 SBCs and 8 more lines for medical and server applications that are supporting ECC. It’s quite efficient to add four lines to a 200-pin RAM connector in order to support SATA Flash on the same module. At the other extreme, it’s not efficient to run several SATA interfaces up an entire stack of cards through multiple large board-to-board connectors.

A truly rugged memory module specification from an open standards organization will be a major step forward for the high end of our market, where the full costs of “blue screens of death” and data corruption far outweigh the price premiums above consumer grade modules. So it’s time for embedded-focused RAM manufacturers to step up and serve their market by advancing the state of the art for mission-critical apps.