Wednesday, February 15, 2006
Myth or Facts
Conventional wisdom sugget that investing in bigger, sophisticated and expensive memory tester means getting better results and catching more errors in proportion with its price.The truth is not always so.
In the process of finding that “perfect” production memory tester, many DRAM makers have been burned by investing too much and getting back too little. “Hundred thousands of dollars and countless hours were spent investing in the more expensive test equipment. At the end, we see no significant results boost,” sighed a manufacturing manager of a major DRAM module maker.
That meant the curve for “price versus performance” stays flat at one point regardless of how much more dollars one pours in or how much more features come equipped in that tester. The simple fact is that certain faults in the memory are just not likely to be weed out with just one-pass testing. That reveals the limitation of a tester no matter how the sellers try to say otherwise. No tester can cover all ground as they are just too many factors lurk around waiting to fail the memory module. Furthermore, the ever-speedier memory and increasingly complex systems just add to the challenge.
Every test engineers can tell you that they always walk on the tight rope trying to do the balancing act. At one end, they like to setup the test to pass all modules to increase the yield. But doing so would risk letting the marginal modules sneak through without proper checked. On the other hand, they can’t tighten the test too much that good modules virtually got over-killed. So, what is the right recipe? It turns out that it all depends on the applications and the type of the chips. Experience test engineer would tell you that server modules should be tested more rigorously as compare to regular PC modules. The same principle applies to untested DRAM (wafer) as opposed to fully tested brand-name memory at chip level.
Server DIMM Modules
Registered modules that are responsible to power the mission critical servers are not allowed to fail during the mission. However, in the imperfect world, we still see a small percentage of server modules fail in the field regardless of how tough the memory being tested before shipment. Many memory module makers have learned the hard way that depending solely on a single tester to test the server memory is simply suicidal. Pretty much every server modules company executes multiple test procedures to ensure their product reliability. One of the steps includes testing the modules on the actual systems. If that is the case, why spend hundred thousands of dollars more on test equipment if your final test is going to be on the motherboards?
Burn in Testers
Some companies like Kingston even go a step beyond by including a burn-in test, which they believe, would force the early-life-failure (ELF) out before they reach the field. ELF normally happens when a module's electronic components cease to function during the first three months of operation.
During the burn-in testing process, server modules are subjected to high-heat, high-stress and high-voltage environments and all memory cells are continuously exercised to simulate the aging process of the modules but at a much faster rate. Marginal errors that would otherwise fail in the filed are now screened out at the factory. Burn-in testing takes a long times. Companies usually cannot afford to do so. The only way is to use multiple sites burn-in tester. If you have need for such burn-in tester, please consult CST, Inc.
Over the last few years, pricing on DRAM memories have been really depressing. Demands for PCs are not as strong as the industry predicted and wanted. That leaves ample of supply in the market that in turn pressure the price downward. Although there are some bright signs lately as DRAM shortage, result of capacity allocation to produce NAND flash memory, might help lift the price. That aside, distributors and module makers have, over the years, learned to find innovative ways to cut cost. One of the solutions is to acquire untested DRAM wafers. As the result, some DRAM memories are only tested in the final product instead of at the chip packaging level. For these types of memories, marginal errors bound to be a few more. Testing procedures need to be more rigorous to screen out the errors. And very often companies do test them on the actual systems to screen out any possible system incompatibility and marginal issues.
Again, if expensive tester can’t pick up extra errors, going for the optional lower cost test equipment might be a better investing decision.
So, what is the ideal tester? Since no tester is perfect, the deciding factors will be depending on your applications. More costly, more features, more test patterns might not necessary translate into better results. Sometimes, a lower cost solution might surprise you big time. The only way to decide which tester works better for your money is to get demo units from all suppliers. Test and evaluate them side-by-side. Some said an expensive mid-range tester with some degree of flexibility in timing adjustment is useful for modules’ characterization. Nevertheless, for volume production, lower cost tester with true clock speed is a better investment.
In very general terms, memory modules can be classified into three categories:
Unbuffered modules with branded chips - have few marginal errors. Both lower cost and expensive testers do the job equally good.
Unbuffered modules with untested chips – have few more marginal errors. Statistics show that both lower cost and expensive testers pick up a few more extra errors but also misses a few.
Registered modules for servers – normally companies pre-test and pre-program with the lower cost tester. Then run them on the motherboards.
Many companies are implementing this "TRI-PASS" Process to ensure the highest level of Quality Products.
Pass 1. Connectivity, Full Functional, and SPD Tests ( Low cost memory tester)
Pass 2. System Test - Motherboard
Pass 3. Verification and QC Test
Copyright © 2006 CST, Inc. All Rights Reserved