Nvidia GPU solder joints at a disadvantage over ATI, scientist says

Posted by Wolfgang Gruener

Los Angeles (CA) – A tiny material issue in Nvidia’s GPUs has cost Nvidia $200 million so
far: The problem
boils down to the solder bump material, in Nvidia’s case high-lead that
was used in all of the firm’s GPUs that were produced until late July ,
and we still do not know how serious the issue really is. According to
our sources, Nvidia has switched to eutectic solder bumps in recent
weeks and there is now a new, apparently independent research report,
that claims that eutectic solder bumps, which are used for example by
AMD’s ATI unit, may live much longer than high-lead versions. Of
course, switching to eutectic isn’t the entire solution, as the
material has a much lower melting point than high-lead.

The research report surfaced late last week on the Electronic Thin Film lab of the Department of Materials Science and Engineering at the University of California at Los Angeles (UCLA), a known authority in material sciences. Author K. N. Tu investigated high-lead (90Pb10Sn) as well as eutectic SnPb (60Pb40Sn) and their exposure to fatigue cracking under thermal stress.

Nvidia’s use of high-lead solder joints and temperatures higher than 70 degrees Celsius are generally believed to be the reasons of the GPU failures reported by the company. The underfill is used to redistribute thermal stress: According to Tu, if “the underfill has [a] low glass transition temperature, it becomes soft at the device operation temperature, and the thermal stress will be taken up mostly by the solder joints.  The solder joint at the corners will fail first.” In each cycle, there is elastic deformation as well as plastic deformation.

The analysis is based on the thought that “when the total plastic energy or work-done to the solder is beyond a threshold value, it will crack due to excessive work hardening and sever plastic deformation.” The critical observation is that the plastic energy produced in the high-lead layer is about 100 times larger than in the eutectic SnPb joint - which leads the scientist to believe that “cycle times needs to fail” are “100 times longer” with the eutectic solder bump. Taking into account that he assumed an “ideal model” for his analysis, he concludes that “it is reasonable to say that it is at the least 10 times longer” in a real life model.

The translation is that eutectic solder joints may have a longer life than the high-lead joints, which would ultimately give such a semiconductor a longer life. We were not able to find out whether Tu’s research was actually commissioned by Nvidia or AMD/ATI, but the timing of the release clearly suggests that the results are aimed at shedding some light on the material issue in Nvidia’s GPUs.   

If Tu’s findings are accurate, then one could conclude that simply switching to eutectic solder joints will cure Nvidia’s headache. However, there are a couple of issues that may prevent a one-to-one transition. First, a transition requires a redesign of the power distribution through a chip, since the melting point of eutectic joints is much longer than in high lead.

Also, according to Tu, “the high-lead solder is known to have better resistance to electromigration, eutectic solder is not, especially when current crowding occurs at the cathode entrance where the electrons enter the solder bumps.” The scientist states that it is unclear whether “electromigration will enhance the failure caused by temperature cycle or power cycle”, but it appears to be clear that a simple switch without further research is simply not possible with a product that is shipped in millions of units.

Overall, Tu believes that the high-lead solder joint “has a built-in weakness” due to the thin layer of eutectic SnPb solder in the joint.  “The eutectic SnPb has a low yield stress and it will deform first and lead to stress concentration and to accumulate high plastic strain energy,” Tu writes. In contrast, the “homogeneous eutectic SnPb solder joint tends to have a uniform composition and much lower plastic deformation so the accumulation of plastic strain energy per unit volume is lower.”

“If we define the failure by an energy criterion, the eutectic SnPb solder joint may have at least ten times longer life-time than the composite solder joint,” Tu wrote. However, he stopped short of actually predicting the actual life time of both joints – which would require “the real microstructure of these two kinds of flip chip solder joints [to] be studied and actual failure rate [to] be measured.”
 
We recently reported that Nvidia, which switched to eutectic solder bumps in late July, has switched back to high-lead solder bumps in an upcoming graphics card series. Now we hear that Nvidia has switched is strategy again and in fact is using eutectic joints in its next generation of GPUs.