Among us (Levi) works together with semiconductors and another (Aeppli) with X-rays. So, after pondering this issue, we considered using X-rays to nondestructively image chips. Youd have to exceed the resolution found in medical X-ray scanners. Nonetheless it was clear to us that the needed resolution was possible. At that time, what weve been calling the chip scan project was created.
Our first technique, ptychographic X-ray computed tomography, was tested first on some of a 22-nanometer Intel processor constructing an in depth 3D image of the chips interconnects.SLS-USC Chip-Scan team
Many years later, weve managed to get possible to map the complete interconnect structure of even probably the most advanced and complex processors without destroying them. At this time, that process takes greater than a day, but improvements on the next couple of years should enable the mapping of entire chips within hours.
This techniquecalled ptychographic X-ray laminographyrequires usage of a few of the worlds most effective X-ray light sources. But many of these facilities are, conveniently, located near where a lot of the advanced chip design happens. In order access to this system expands, no flaw, failure, or fiendish trick can hide.
After deciding to pursue this process, our first order of business was to determine what state-of-the-art X-ray techniques could do. That has been done at the Paul Scherrer Institute (PSI) in Switzerland, where among us (Aeppli) works. PSI houses the Swiss SOURCE OF LIGHT (SLS) synchrotron, among the 15 brightest resources of coherent X-rays built up to now.
Coherent X-rays change from whats found in a medical or dentist office just as that the highly collimated laser beam from the laser pointer differs from light emitted everywhere from an incandescent bulb. The SLS and similar facilities generate highly coherent beams of X-ray photons by first accelerating electrons almost to the speed of light. Then, magnetic fields deflect those electrons, causing the production of the required X-rays.
To see what we’re able to do with the SLS, our multidisciplinary team bought an Intel Pentium G3260 processor from the local store for approximately US $50 and removed the packaging to expose the silicon. (This CPU was manufactured using 22-nanometer CMOS FinFET technology).
A fly-though of the very best layers of an Intel 22-nanometer processor reconstructed from X-ray scans.SLS-USC Chip-Scan Team
Like all such chips, the G3260s transistors are constructed with silicon, but its the arrangement of metal interconnects that link them around form circuits. In today’s processor, interconnects are designed in a lot more than 15 layers, which from above appear to be a map of a citys street grid. The low layers, nearer to the silicon, have incredibly fine features, spaced just nanometers apart in todays innovative chips. As you ascend the interconnect layers, the features become sparser and bigger, and soon you reach the very best, where electrical contact pads connect the chip to its package.
We began our examination by eliminating a 10-micrometer-wide cylinder from the G3260. We’d to take this destructive step since it greatly simplified things. Ten micrometers is not even half the penetration depth of the SLSs photons, so with something this small wed have the ability to detect enough photons passing through the pillar to find out that which was inside.
We placed the sample on a mechanical stage to rotate it about its cylindrical axis and fired a coherent beam of X-rays through the medial side. Because the sample rotated, we illuminated it with a pattern of overlapping 2-m-wide spots.
At each illuminated spot, the coherent X-rays diffracted because they passed through the chips tortuous tower of copper interconnects, projecting a pattern onto a detector, that was stored for subsequent processing. The recorded projections contained enough information regarding the material by which the X-rays traveled to look for the structure in three dimensions. This process is named ptychographic X-ray computed tomography (PXCT). Ptychography may be the computational procedure for producing a graphic of something from the interference pattern of light through it.
The underlying principle behind PXCT is not at all hard, resembling the diffraction of light through slits. You may recall from your own introductory physics class that should you shine a coherent laser beam by way of a slit onto a distant plane, the experiment produces whats called a Fraunhofer diffraction pattern. It is a pattern of light and dark bands, or fringes, spaced proportionally to the ratio of the lights wavelength divided by the width of the slit.
If, rather than shining light by way of a slit, you shine it on a set of closely spaced objects, ones so small they are effectively points, you’ll get another pattern. It doesnt matter where in the beam the objects are. Provided that they stay exactly the same distance from one another, it is possible to move them around and youd obtain the same pattern.
Independently, neither of the phenomena enables you to reconstruct the tangle of interconnects in a microchip. But in the event that you combine them, youll begin to see how it might work. Put the couple of objects within the slit. The resulting interference pattern comes from the diffraction because of mix of slit and object, revealing information regarding the width of the slit, the length between your objects, and the relative position of the objects and the slit. In the event that you move both points slightly, the interference pattern shifts. And its own that shift which allows one to calculate wherever the objects are within the slit.
Any real sample could be treated as a couple of pointlike objects, which bring about complex X-ray scattering patterns. Such patterns may be used to infer how those pointlike objects are arranged in two dimensions. And the principle may be used to map things out in three dimensions by rotating the sample within the beam, an activity called tomographic reconstruction.
You have to make certain youre setup to get enough data to map the structure at the mandatory resolution. Resolution depends upon the X-ray wavelength, how big is the detector, and some other parameters. For the initial measurements with the SLS, that used 0.21-nm-wavelength X-rays, the detector needed to be placed about 7 meters from the sample to attain our target resolution of 13 nm.
In March 2017, we demonstrated the usage of PXCT for nondestructive imaging of integrated circuits by publishing some very pretty 3D images of copper interconnects in the Intel Pentium G3260 processor. Those images reveal the three-dimensional character and complexity of electrical interconnects in this CMOS integrated circuit. However they also captured interesting details like the imperfections in the metal connections between your layers and the roughness between your copper and the silica dielectric around it.
Out of this proof-of-principle demonstration alone, it had been clear that the technique had potential in failure analysis, design validation, and quality control. So we used PXCT to probe similarly sized cylinders cut from chips constructed with others technologies. The facts in the resulting 3D reconstructions were like fingerprints which were unique to the ICs and in addition revealed much concerning the manufacturing processes used to fabricate the chips.
We were encouraged by our early success. But we knew we’re able to do better, because they build a new kind of X-ray microscope and discovering more effective methods to improve image reconstruction using chip design and manufacturing information. We called the brand new technique PyXL, shorthand for ptychographic X-ray laminography.
The very first thing to cope with was how exactly to scan a complete 10-millimeter-wide chip whenever we had an X-ray penetration depth of only around 30 m. We solved this issue by first tilting the chip at an angle in accordance with the beam. Next, we rotated the sample concerning the axis perpendicular to the plane of the chip. Simultaneously we also moved it sideways, raster fashion. This allowed us to scan all elements of the chip with the beam.
At each moment in this technique, the X-rays passing through the chip are scattered by the materials in the IC, developing a diffraction pattern. Much like PXCT, diffraction patterns from overlapping illumination spots contain redundant information regarding what the X-rays have passed through. Imaging algorithms then infer a structure this is the most in keeping with all measured diffraction patterns. From these we are able to reconstruct the inside of the complete chip in 3D.
Obviously, there’s plenty to be worried about when creating a new sort of microscope. It will need to have a well balanced mechanical design, including precise motion stages and position measurement. Also it must record at length the way the beam illuminates each i’m all over this the chip and the ensuing diffraction patterns. Finding practical answers to these along with other issues required the efforts of a team of 14 engineers and physicists. The geometry of PyXL also required developing new algorithms to interpret the info collected. It had been effort, but by late 2018 we’d successfully probed 16-nm ICs, publishing the outcomes in October 2019.
Todays cutting-edge processors might have interconnects less than 30 nm apart, and our technique can, at the very least in principle, produce images of structures smaller than 2 nm.
In these experiments, we could actually use PyXL to peel away each layer of interconnects virtually to reveal the circuits they form. Being an early test, we inserted a little flaw in to the design apply for the interconnect layer closest to the silicon. Whenever we compared this version of the layer with the PyXL reconstruction of the chip, the flaw was immediately obvious.
In principle, a couple of days of work is all wed have to use PyXL to acquire meaningful information regarding the integrity of an IC stated in even probably the most advanced facilities. Todays cutting-edge processors might have interconnects just tens of nanometers apart, and our technique can, at the very least in principle, produce images of structures smaller than 2 nm.
The brand new version of our X-ray technique, called ptychographic X-ray laminography, can uncover the interconnect structure of entire chips without damaging them, even right down to the tiniest structures [top]. Using that technique, we’re able to easily locate a (deliberate) discrepancy between your design file and that which was manufactured [bottom].
But increased resolution does take longer. Even though hardware weve built can completely scan a location around 1.2 by 1.2 centimeters at the best resolution, doing this will be impractical. Zooming in on a location of interest will be a better usage of time. Inside our initial experiments, a low-resolution (500-nm) scan over a square part of a chip that has been 0.3 mm on a side took 30 hours to obtain. A high-resolution (19-nm) scan of a much smaller part of the chip, just 40 m wide, took 60 hours.
The imaging rate is fundamentally tied to the X-ray flux open to us at SLS. But other facilities boast higher X-ray fluxes, and methods come in the works to improve X-ray source brilliancea mix of the amount of photons produced, the beams area, and how quickly it spreads. For instance, the MAX IV Laboratory in Lund, Sweden, pioneered a method to boost its brilliance by two orders of magnitude. An additional a couple of orders of magnitude can be acquired through new X-ray optics. Combining these improvements should 1 day increase total flux by way of a factor of 10,000.
With this particular higher flux, we have to have the ability to achieve an answer of 2 nm in less time than it now takes to acquire 19-nm resolution. Our bodies may possibly also survey a one-square-centimeter integrated circuitabout how big is an Apple M1 processorat 250-nm resolution in less than 30 hours.
And you can find different ways of boosting imaging speed and resolution, such as for example better stabilizing the probe beam and improving our algorithms to take into account the look rules of ICs and the deformation that may result from an excessive amount of X-ray exposure.
Although we are able to already tell a whole lot about an IC from just the layout of its interconnects, with further improvements we ought to have the ability to discover everything about any of it, like the materials its manufactured from. For the 16-nm-technology node, which includes copper, aluminum, tungsten, and compounds called silicides. We may even have the ability to make local measurements of strain in the silicon lattice, which comes from the multilayer manufacturing processes had a need to make cutting-edge devices.
Identifying materials could become particularly important, given that copper-interconnect technology is approaching its limits. In contemporary CMOS circuits, copper interconnects are vunerable to electromigration, where current can kick copper atoms out of alignment and cause voids in the structure. To counter this, the interconnects are sheathed in a barrier material. But these sheaths could be so thick they leave little room for the copper, making the interconnects too resistive. So alternative materials, such as for example cobalt and ruthenium, are increasingly being explored. As the interconnects involved are so fine, well have to reach sub-10-nm resolution to tell apart them.
Theres reason to believe well make it happen. Applying PXCT and PyXL to the connectome of both hardware and wetware (brains) is among the key arguments researchers all over the world have designed to support the construction of new and upgraded X-ray sources. For the time being, work continues inside our laboratories in California and Switzerland to build up better hardware and software. So someday soon, if youre suspicious of one’s new CPU or interested in a competitors, you can create a fly-through tour through its inner workings to ensure everything is actually in its proper place.
The SLS-USC Chip-Scan Team includes Mirko Holler, Michal Odstrcil, Manuel Guizar-Sicairos, Maxime Lebugle, Elisabeth Mller, Simone Finizio, Gemma Tinti, Christian David, Joshua Zusman, Walter Unglaub, Oliver Bunk, Jrg Raabe, A. F. J. Levi, and Gabriel Aeppli.
This short article appears in the May 2022 print issue because the Naked Chip.