# How Chips Are Using Cryogenic Cooling for Quantum-Class Performance

An episode of [The Hardware Podcast with Fexingo: Chips, Devices, and Electronics Engineering Conversations](https://invisiblenetwork.ai/podcast/1896799982), 27 Jun 2026, 11 min.

Audio: https://audio.fexingo.com/business/the-hardware-podcast/episode-0077.mp3

In this episode of The Hardware Podcast, Lucas and Luna explore the emerging field of cryogenic chip cooling—not for quantum computers, but for classical semiconductor logic. They break down how Intel and others are experimenting with sub-100 Kelvin operation to slash leakage current and boost transistor switching speed. The discussion centers on a 2025 ISSCC paper from Intel Labs showing a 40% frequency improvement at 4 Kelvin, and what it would take to bring liquid-helium-grade cooling into the data center. Is this the next frontier in semiconductor performance, or a dead end for all but the most specialized applications? Lucas and Luna weigh the engineering hurdles and the potential payoff for high-performance computing and AI inferencing. #CryogenicCooling #Intel #Semiconductor #LowTemperatureComputing #LeakageCurrent #TransistorPerformance #ISSCC #HighPerformanceComputing #DataCenter #QuantumComputing #LiquidHelium #ChipDesign #Technology #HardwarePodcast #FexingoBusiness #BusinessPodcast #ChipCooling #AdvancedLogic Keep every episode free: buymeacoffee.com/fexingo

## Transcript

Lucas: So you know how whenever we talk about extreme chip performance, the conversation usually lands on smaller transistors, new materials, better architectures — but there's another lever that's been sitting on the table for decades and it's suddenly getting serious attention: temperature.
Luna: You mean cooling chips way down, like cryogenic cold?
Lucas: Exactly. Not just a fan or a liquid loop — we're talking sub-100 Kelvin. Liquid helium territory. And the idea is that if you drop the temperature enough, you fundamentally change how silicon behaves. Leakage current plummets, electron mobility goes up, and you can potentially clock the chip much higher without burning through your power budget.
Luna: I've heard about this in the context of quantum computers — you need those dilution fridges to keep qubits stable. But for classical logic? That feels like a different kind of engineering challenge.
Lucas: It is a different challenge, but the payoff could be huge. Look, quantum computers need cryogenic temperatures because their qubits are fragile. Classical CMOS doesn't need it to function — it works fine at room temperature. But there's a growing argument that if we're willing to pay the cooling cost, we can get a generation-level performance boost without shrinking the transistor.
Luna: So instead of moving to 2 nanometers, you just freeze your 3 nanometer chip and get the same speed boost? That sounds almost too good to be true.
Lucas: It really does. And there are real trade-offs. But let me give you a concrete data point. At the 2025 International Solid-State Circuits Conference — ISSCC — Intel Labs presented a paper where they took a test chip built on their Intel 4 process — that's 7 nanometer class — and cooled it down to 4 Kelvin. At that temperature, they measured a 40 percent increase in maximum operating frequency compared to room temperature.
Luna: Forty percent? That's enormous. That's better than what you'd typically get from a full process node shrink.
Lucas: Right. And it's not just frequency. They also saw a dramatic reduction in leakage power — like orders of magnitude. At 4 Kelvin, the transistor off-current is basically zero. So the dynamic power for a given frequency actually decreases because you don't have that constant trickle of wasted energy.
Luna: But the cooling itself must consume a huge amount of power. Liquid helium systems aren't exactly efficient.
Lucas: That's the central trade-off. A typical cryocooler that can get down to 4 Kelvin requires multiple stages — it's like a very fancy, very expensive refrigerator. The coefficient of performance is terrible. You might need several kilowatts of input power to remove a few watts of heat at the cold stage. So for most applications, the math doesn't work.
Luna: Unless the chip you're cooling is itself drawing a huge amount of power, right? Then the cooling overhead becomes a smaller fraction.
Lucas: Exactly. That's why the early targets are high-performance computing centers and data centers running AI inference at scale. If a rack of GPUs draws 40 kilowatts, adding a 10 kilowatt cryocooler might be acceptable if it doubles performance per watt. But for a smartphone or a laptop? Not happening.
Luna: So the use case is really about pushing the absolute ceiling on performance for the most demanding workloads — like scientific simulations, maybe some types of AI training.
Lucas: And there's another interesting angle: cryogenic memory. DRAM actually works better at low temperatures too — refresh rates can be slowed way down, and access times improve. So you could imagine a system where both the processor and the memory are in the same cold environment, with superconducting interconnects between them. That would slash latency.
Luna: That's basically a classical computer built with quantum computer infrastructure. The cooling bill must be astronomical.
Lucas: It's not cheap today. But there's a lot of research into more efficient cryocoolers — things like pulse tube refrigerators that have fewer moving parts and lower maintenance. And some groups are looking at using high-temperature superconductors that work at liquid nitrogen temperatures — 77 Kelvin — which is much cheaper than liquid helium. You might not get the full 40 percent frequency boost, but maybe 20 percent with a much lower cooling cost.
Luna: That's a more pragmatic sweet spot. So which companies are actually investing in this?
Lucas: Intel is the most vocal. They have a dedicated cryogenic computing lab. IBM is also active — they've been building cryogenic CMOS controllers for their quantum processors, and they see spillover benefits for classical logic. And there's a startup called CryoLogic out of MIT that's designing standard-cell libraries optimized for 77 Kelvin operation. They claim a 30 percent power reduction at the same frequency compared to room temperature.
Luna: That's interesting. So the idea is to design the chip specifically for cold operation, not just take a room-temperature chip and cool it.
Lucas: Exactly. If you optimize the transistor threshold voltages and the doping profiles for low temperature, you can squeeze out even more benefit. At room temperature, you have to design around leakage — the threshold can't be too low or the transistor never fully turns off. At 77 Kelvin, leakage is so low you can afford to drop the threshold, which gives you faster switching and lower voltage. It's a virtuous cycle.
Luna: So the question becomes: when does the total cost of ownership for a cryo-cooled data center beat air-cooled? And is that crossover point coming in the next decade?
Lucas: Right. And that depends on how much performance improvement you actually get and how efficiently you can deliver the cold. There's a 2024 analysis from the European Laboratory for Particle Physics — CERN — that looked at this for their own compute needs. They concluded that for workloads with high utilization — like 90 percent plus — cryogenic cooling could be cost-competitive if the chip performance gain exceeds 50 percent.
Luna: Fifty percent is a high bar. But the Intel data at 4 Kelvin gave 40 percent. So with design optimization, maybe they can hit 50.
Lucas: And that's just for frequency. If you also factor in the reduced leakage, the power savings could push the effective gain higher. But there's another challenge: reliability. Thermal cycling between room temperature and cryogenic temperatures puts mechanical stress on the chip and the package. Solder joints can crack. Materials expand and contract at different rates.
Luna: So you'd want to keep the system cold once it's started up — avoid those temperature swings.
Lucas: Right. That's what the data center operators would do — bring the whole rack down to operating temperature and keep it there. But then you have the problem of maintenance: if a chip fails, you have to warm up the system to replace it, which takes time and energy. That's a real operational cost.
Luna: It's interesting to think about how this intersects with the push toward more sustainable computing. If you're using more energy to cool, is that a net negative for the environment, even if the chips themselves are more efficient?
Lucas: That's a great point. The environmental math depends on the carbon intensity of the electricity and the overall efficiency gain. If the cooling system adds 30 percent to the total power draw but the chip does 50 percent more work, then the carbon per computation goes down. But if the local grid is coal-heavy, the absolute emissions might still be high.
Luna: Right. So location matters — put these facilities near hydro or nuclear power. Or pair them with on-site renewable generation.
Lucas: And that's actually something we're seeing. Some of the early experimental cryo-computing testbeds are being set up in places like Iceland and the Pacific Northwest where cheap renewable power is available. It's not a coincidence.
Luna: If today's tech conversation gave you something usable — maybe a new angle on chip performance or a better sense of where computing is heading — that's the whole reason we do this show. We deliberately don't run ads on these conversations, and that's a choice we really believe in.
Lucas: Yeah. No sponsors, no interruption. Just the topic, straight through. If that format works for you and you want to support it continuing, there's a straightforward way — buy me a coffee dot com slash fexingo. Totally optional, no pressure, but it makes a difference.
Luna: It really does. And it keeps the show free for everyone. So anyway, back to the cold — Lucas, you mentioned that some of the biggest players are already experimenting. What about AMD and NVIDIA?
Lucas: AMD has published some research on cryogenic SRAM — their memory cells show improved stability at low temperatures. NVIDIA is more quiet publicly, but they have a vested interest since their GPUs are used in HPC and AI — and those workloads are the natural candidates. I'd be surprised if they don't have internal projects.
Luna: So the pieces are coming together: the chip design, the cooling infrastructure, the economic case. What's the one thing that would need to happen for this to go mainstream?
Lucas: A breakthrough in cryocooler efficiency — specifically, a system that can achieve 77 Kelvin with a coefficient of performance better than about 10 percent of Carnot. Right now, commercial coolers are around 5 percent. If someone can double that, the economics shift dramatically.
Luna: That's an engineering challenge, not a physics problem. So it's solvable.
Lucas: Exactly. And there's a lot of motivated research. The Department of Energy in the US has a program called 'Cold Computing' that's funding exactly this kind of work. So I think in five to ten years, we'll start seeing cryo-cooled racks in specialized data centers. Not for everything — but for the jobs that need every last drop of performance.
Luna: And that's really the theme of this whole conversation: we're not done pushing silicon. We're just finding new ways to push it.
Lucas: Right. The end of Moore's Law doesn't mean the end of performance scaling. It means we have to be more creative. And sometimes that creativity means making things really, really cold.

Shows are listed from Apple Podcasts' charts and directory, and Podcast Index where it is used. Every show and episode belongs to its publisher and plays from the publisher's own feed; this site hosts none of them.