Article 70T39 You Can Cool Chips With Lasers?!?!

You Can Cool Chips With Lasers?!?!

by
Jacob Balma
from IEEE Spectrum on (#70T39)
a-laser-shining-on-a-computer-chip-covered-in-snow.png?id=61755630&width=1245&height=700&coordinates=0%2C785%2C0%2C786

Modern high-performance chips are marvels of engineering, containing tens of billions of transistors. The problem is, you can't use them all at once. If you did, you would create hot spots-high temperatures concentrated in tiny areas-with power densities nearing those found at the surface of the sun. This has led to a frustrating paradox known as dark silicon, a term coined by computer architects to describe the growing portion of a chip that must be kept powered down. Up to 80 percent of the transistors on a modern chip must remain dark" at any given moment to keep the chip from sizzling. We are building supercomputers on a sliver of silicon but only using a fraction of their potential. It's like building a skyscraper and being able to use only the first 10 floors.

For years, the industry has battled this thermal limit with bigger fans and more complex liquid cooling systems. But these are fundamentally Band-Aid solutions. Whether using air or liquid, they rely on pulling heat away from the chip's surface. The heat must first conduct through the silicon to the cooling plate, creating a thermal bottleneck that simply cannot be overcome at the power densities of future chips. Hot spots on today's chips produce tens of watts per square millimeter, and they pop up in various places on the chip at different times during computations. Air and liquid cooling struggle to focus their efforts at just the hot spots, when and where they appear-they can only try to cool the whole thing en masse.

We at St. Paul, Minn.-based startup Maxwell Labs are proposing a radical new approach: What if, instead of just moving heat, you could make it disappear? The technology, which we call photonic cooling, is capable of converting heat directly into light-cooling the chip from the inside out. The energy can then be recovered and recycled back into useful electric power. With this approach, instead of cooling the whole chip uniformly, we can target hot spots as they form, with laser precision. Fundamentally, this technique could cool hot spots of thousands of watts per square millimeter, orders of magnitude better than today's chips are cooled.

The Physics of Cooling With Light

Lasers are usually thought of as sources of heat, and for good reason-they are most commonly used for cutting materials or transferring data. But under the right circumstances, laser light can induce cooling. The secret lies in a luminescent process known as fluorescence.

Fluorescence is the phenomenon behind the familiar glow of highlighter markers, coral reefs, and white clothes under black-light illumination. These materials absorb high-energy light-usually in the ultraviolet-and reemit lower energy light, often in the visible spectrum. Because they absorb higher energy than they emit, the difference often results in heating up the material. However, under certain, very niche conditions, the opposite can happen: A material can absorb low-energy photons and emit higher-energy light, cooling down in the process.

a-stack-of-squares-on-top-of-a-layered-cube-with-arrows-pointing-at-the-squares-from-the-top.png?id=61755825&width=980 To cool computer chips with lasers, the team at Maxwell Labs plans to place a grid of photonic cold plates on top of the chip substrate. In their demo setup, a thermal camera detects hot spots coming from the chip. A laser then shines onto the photonic cold plate next to the hot spot, stimulating the photonic process that results in cooling. The photonic cold plate [inset] consists of a coupler that guides light in and out of the plate, the extractor where anti-Stokes fluorescence occurs, the back reflector that prevents light from entering the computer chip, and a sensor that is designed to detect hot spots.GygInfographics.com

The reemission is higher energy because it combines the energy from the incoming photons with phonons, vibrations in the crystal lattice of a material. This phenomenon is called anti-Stokes cooling, and it was first demonstrated in a solid back in 1995 when a team of scientists cooled an ytterbium-doped fluoride glass sample with laser light.

The choice of ytterbium as a dopant was not random: Anti-Stokes cooling works only under carefully engineered conditions. The absorbing material must be structured so that for nearly every absorbed photon a higher-energy photon will be emitted. Otherwise, other mechanisms will kick in, heating rather than cooling the sample. Ions of ytterbium and other such lanthanides have the right structure of electron orbitals to facilitate this process. For a narrow range of laser wavelengths shining on the material, the ions can effectively absorb the incident light and use phonons to trigger emission of higher-energy light. This reemitted, extracted thermal light needs to escape the material quickly enough to not be absorbed again, which would otherwise cause heating.

To date, lab-based approaches have achieved up to 90 watts of cooling power in ytterbium-doped silica glass. As impressive as that is, to achieve the transformative effects on high-performance chips that we anticipate, we need to boost the cooling capacity by many orders of magnitude. Achieving this requires integration of the photonic cooling mechanism onto a thin-film, chip-scale photonic cold plate. Miniaturization not only enables more precise spatial targeting of hot spots due to the tightly focused beam, but is a crucial element for pushing the physics of laser cooling toward high-power and high-efficiency regimes. The thinner layer also makes it less likely that the light will get reabsorbed before escaping the film, avoiding heating. And, by engineering the materials at the scale of the wavelength of light, it allows for increased absorption of the incoming laser beam.

Photonic Cold-Plate Technology

In our lab, we are developing a way to harness photonic cooling to tackle the heat from today's and future CPUs and GPUs. Our photonic cold plate is designed to sense areas of increasing power density (emerging hot spots) and then couple light efficiently into a nearby region that cools the hot spots down to a target temperature.

The photonic cold plate has several components: first the coupler, which couples the incoming laser light into the other components; then, the microrefrigeration region, where the cooling actually happens; next, the back reflector, which prevents light from hitting the CPU or GPU directly; and last a sensor, which detects the hot spots as they form.

The laser shines onto the targeted area from above through the coupler: a kind of lens that focuses the incoming laser light onto a microrefrigeration region. The coupler simultaneously channels the inbound heat-carrying fluorescent light out of the chip. The microrefrigeration region, which we call the extractor, is where the real magic happens: The specially doped thin film undergoes anti-Stokes fluorescence.

To prevent the incoming laser light and fluorescent light from entering the actual chip and heating the electronics, the photonic cold plate incorporates a back reflector.

Crucially, cooling occurs only when, and where, the laser is shining onto the cold plate. By choosing where to shine the laser, we can target hot spots as they appear on the chip. The cold plate includes a thermal sensor that detects hot spots, allowing us to steer the laser toward them.

Designing this whole stack is a complex, interconnected problem with many adjustable parameters, including the exact shape of the coupler, the material and doping level of the extraction region, and the thickness and number of layers in the back reflector. To optimize the cold plate, we are deploying a multiphysics simulation model combined with inverse design tools that let us search the vast set of possible parameters. We are leveraging these tools in the hope of improving cooling power densities by two orders of magnitude, and we are planning larger simulations to achieve bigger improvements still.

Collaborating with our partners at the University of New Mexico in Albuquerque, the University of St. Thomas in St. Paul, Minn., and Sandia National Laboratories in Albuquerque, we are building a demonstration version of photonic cooling at our lab in St. Paul. We are assembling an array of small photonic cold plates, each a square millimeter in size, tiled atop various CPUs. For demonstration purposes, we use an external thermal camera to sense the hot spots coming from the chips. When a hot spot begins to appear, a laser is directed onto the photonic cold plate tile directly atop it, extracting its heat. Our first iteration of the cold plate used ytterbium ion doping, but we are now experimenting with a variety of other dopants that we believe will achieve much higher performance.

In an upcoming integrated implementation of this demo, the photonic cold plates will consist of finer tiles-about 100 by 100 micrometers. Instead of a free-space laser, light from a fiber will be routed to these tiles by an on-chip photonic network. Which tiles are activated by the laser light will depend on where and when hot spots form, as measured by the sensor.

Eventually, we hope to collaborate with CPU and GPU manufacturers to integrate the photonic cold plates within the same package as the chip itself, allowing us to get the crucial extractor layer closer to the hot spots and increase the cooling capacity of the device.

The Laser-Cooled Chip and the Data Center

To understand the impact of our photonic cooling technology on current and future data centers, we have performed an analysis of the thermodynamics of laser cooling combined with and compared to air and liquid cooling approaches. Preliminary results show that even a first-generation laser-cooling setup can dissipate twice the power of purely air and liquid cooling systems. This drastic improvement in cooling capability would allow for several key changes to chip and data-center architectures of the future.

First, laser cooling could eliminate the dark-silicon problem. By sufficiently removing heat from hot spots as they are forming, photonic cooling would permit simultaneous operation of more of the transistors on a chip. That would mean all the functional units on a chip could function in parallel, bringing the full force of modern transistor densities to bear.

Second, laser cooling can allow for much higher clocking frequencies than is currently possible. This cooling technique can maintain the chip's temperature below 50 C everywhere, because it targets hot spots. Current-generation chips typically experience hot spots in the 90-to-120 C range, and this is expected only to get worse. The ability to overcome this bottleneck would allow for higher clocking frequencies on the same chips. This opens up the possibility of improving chip performance without directly increasing transistor densities, giving much needed headroom for Moore's Law to continue to progress.

a-black-box-with-optics-and-a-chip-on-top-and-a-camera-showing-a-colorful-blob-above-it.png?id=61756057&width=980 The demo setup at Maxwell Labs demonstrates how current computer chips can be cooled with lasers. A photonic cold plate is placed on top of the chip. A thermal camera images the hot spots coming from the chip, and a laser is directed at the photonic cold plate directly above the hot spot.Maxwell Labs

Third, this technology makes 3D integration thermally manageable. Because laser-assisted cooling pinpoints the hot spots, it can more readily remove heat from a 3D stack in a way that today's cooling tech can't. Adding a photonic cold plate to each layer in a 3D integrated stack would take care of cooling the whole stack, making 3D chip design much more straightforward.

Fourth, laser cooling is more efficient than air cooling systems. An even more tantalizing result of the removal of heat from hot spots is the ability to keep the chip at a uniform temperature and greatly reduce the overall power consumption of convective cooling systems. Our calculations show that, when combined with air cooling, reductions in overall energy consumption of more than 50 percent for current generation chips are possible, and significantly larger savings would be achieved for future chips.

What's more, laser cooling allows for recovering a much higher fraction of waste energy than is possible with air or liquid cooling. Recirculating hot liquid or air to heat nearby houses or other facilities is possible in certain locations and climates, but the recycling efficiency of these approaches is limited. With photonic cooling, the light emitted via anti-Stokes fluorescence can be recovered by re-collecting the light into fiber-optic cables and then converting it to electricity through thermophotovoltaics, leading to upwards of 60 percent energy recovery.

With this fundamentally new approach to cooling, we can rewrite the rules by which chips and data centers are designed. We believe this could be what enables the continuation of Moore's Law, as well as the power savings at the data-center level that could greenlight the intelligence explosion we're starting to see today.

The Path to Photonic Cooling

While our results are highly promising, several challenges remain before this technology can become a commercial reality. The materials we are currently using for our photonic cold plates meet basic requirements, but continued development of higher efficiency laser-cooling materials will improve system performance and make this an increasingly economically attractive proposition. To date, only a handful of materials have been studied and made pure enough to allow laser cooling. We believe that miniaturization of the photonic cold plate, aided by progress in optical engineering and thin-film materials processing, will have similarly transformative effects on this technology as it has for the transistor, solar cells, and lasers.

We're going to need to codesign the processors, packages, and cooling systems to maximize benefits. This will require close collaboration across the traditionally siloed semiconductor ecosystem. We are working with industry partners to try to facilitate this codesign process.

Transitioning from a lab-based setup to high-volume commercial manufacturing will require us to develop efficient processes and specialized equipment. Industry-wide adoption necessitates new standards for optical interfaces, safety protocols, and performance metrics.

Although there is much to be done, we do not see any fundamental obstacles now to the large-scale adoption of photonic cooling technology. In our current vision, we anticipate the early adoption of the technology in high-performance computing and AI training clusters before 2027, showing an order-of-magnitude improvement in performance per watt of cooling. Then, between 2028 and 2030, we hope to see mainstream data-center deployment, with an accompanied reduction in IT energy consumption of 40 percent while doubling compute capacity. Finally, after 2030 we foresee that ubiquitous deployment, from hyperscale to edge, will enable new computing paradigms limited only by algorithmic efficiency rather than thermal constraints.

For over two decades, the semiconductor industry has grappled with the looming threat of dark silicon. Photonic cooling offers not merely a solution to that challenge but a fundamental reimagining of the relationship between performance, computation, and energy. By converting waste heat directly into useful photons and ultimately back into electricity, this technology transforms thermal management from a necessary evil into a valuable resource.

The future of computing is photonic, efficient, and brilliantly cool.

External Content
Source RSS or Atom Feed
Feed Location http://feeds.feedburner.com/IeeeSpectrum
Feed Title IEEE Spectrum
Feed Link https://spectrum.ieee.org/
Reply 0 comments