BiTS 2016 general chair Ira Feldman welcomed attendees to the BiTS technical sessions this morning, noting that the goal is to focus on what’s current and coming up next with respect to the burn-in and test of packaged ICs. Complementing eight technical sessions are 47 exhibitors. Feldman also provided a “trip report,” describing the October 2015 BiTS event in Shanghai, which gave him a chance to ride the MagLev train. It costs a buck a minute, he said, but at 267 mph, his trip took only eight minutes. The Shanghai event led to an East Meets West session scheduled for tomorrow at this week’s BiTS event. The next BiTS event is scheduled to begin September 13, 2016, in Suzhou, China. The next U.S. event is scheduled for March 5-8, 2017, once again in Mesa.
Feldman then introduced Dale Ohmart, distinguished member of the technical staff at Texas Instruments, who delivered a keynote address titled “Chip Overtest: Are ICs Tested too Much?”
Ohmart’s starting point is that we generally consider the purpose of test to be proving every part shipped is good, leading to the further proposition that “test must verify all the specs of the part to ensure it’s a good part.” Such thinking leads to test complexity and spiraling test costs, he said, and his address was an attempt to lay the groundwork for a new definition of test.
Total test-related capital spending is about 4% of semiconductor revenue, he said, and is slowly increasing. That’s driven in part by a shift from consumer devices to devices aimed at industrial and automotive applications, which can have a big impact on test cost. Further, while ATE costs have fallen, they seem to have bottomed out.
The challenge to test professionals, he said, is to determine why costs are rising and at what point they become unacceptable. If 4% is acceptable now, what about 10, 20, or even 50%? It’s time to ask, “Is more test the best way to achieve better quality?”
He then asked, “Why test? Only because we can’t build the chips right—there are a few defects in every lot.” He then contended that a chip is finished when released to test—test does not add manufacturing value.”
He then acknowledged that test-related operations like trimming and calibration can add value, as can “sorting for goodness,” which can yield higher prices for the fastest parts or amplifiers with lowest distortion. But the burgeoning IoT devices, he added, are not going to be trimmed or repaired—they are disposable, and testing them does not add any value.
Ohmart pointed out that test can have four consequences: a good part ships, a bad part is rejected, a bad part is shipped, or a good part is rejected (yield loss). Test should maximize the first two and minimize the last two.
Many devices with manufacturing defects (missing bond wires or cracked die) can easily be identified without an expensive tester—they simply won’t function at all. The test results for marginal devices can be influenced by manipulating guardbands, for example, but in fact, he said, many devices returned by customers are not marginal. The test process may have failed to identify failed devices because of a tester failure, handler jam, or tester/handler communication errors. An expensive tester may not prevent such errors, and he added, “Capital decisions are absolutely critical in trying to control test costs.”
Ohmart then turned his attention to multisite test, which some consider a magic bullet for cutting test costs. Indeed, multisite can reduce cost of ownership per test site—if you can effectively use all the sites. But an expensive multisite tester can yield diminishing returns, because time on the tester may not be the most significant contributor to total test time—you have to consider stops between lots, losses in the handler, program changes, and load-board changes.
He also cited techniques such as strip test, but that simply shifts the sort burden to post-test, and consequently, “Final test is not final.”
Turret handlers offer a high-speed alternative, enabling 16,000 pph for small packages and < 200 ms test times. It may be advantageous to test IoT devices that way.
He noted that multisite test may make sense, but evaluation is a must, as the approach entails risk and can drive other cost and quality metrics the wrong way. Further, less test can drive cost and quality metrics the right way, he said.
The conclusion, he said, is, “Find the most bang for the buck!”
