One consequence of the recent advances in AI platforms is the staggering need for more compute power. Major tech players worldwide announced to invest billions in new data centers in order to cover the computational costs for model training and inference. The resulting electricity demand challenges existing energy grids and questions the sustainability of this development.
An alternative on the horizon of energy-efficient compute is the Quantum Computing paradigm. With the potential to solve certain problems exponentially faster, a fault-tolerant quantum computer could compute at a much lower energy cost. In this article we present a classical-vs-quantum use case study based on Google’s Supremacy Experiment and estimate the impact of high-throughput quantum chip testing on the energy footprint.
Goldman Sachs published a report [1] on AI/data centers’ global power surge and the sustainability impact. This figure shows that efficiency gains have decelerated, which in the past outweighed the need for more power. We also see that the power demand of data centers is strongly increasing with an additional acceleration due to AI.
In a recent study the Electric Power Research Institute in Washington concluded that data centers could consume up to 9% of U.S. electricity generation by 2030 – more than double their current consumption. It is expected that more than half of the additional electricity demand will be covered by natural gas, leading to a more than 100% increase (about 215-220 million tons) in data center carbon dioxide emissions by 2030 vs. 2022 [1].
But the problem and limits posed by the energy consumption of traditional CMOS transistors is not new. In 2005 we already reached the maximum speed for operating typical CPUs. Increased heat and power issues prevent us to operate chips above 5 GHz, without the need for significant cooling solutions (Dennard Scaling).
Consequently, the new paradigm called “More than Moore” emerged in the early / mid 2000s. The idea is to enhance the functionality and performance of electronic systems through methods other than just shrinking transistors. Next to approaches based on CMOS technology such as heterogeneous and 3D integration, ASICs, or advanced packaging, there was also an openness to new computational paradigms. Photonic computing, molecular computing, and neuromorphic computing are only some examples. One of the most promising approaches to overcome current physical limitations is quantum computing.
Energy efficient computation by using Quantum Algorithms
The mechanics of quantum computers allow to harness the phenomena of entanglement and superposition, which enable an exponential speedup for computing selected problems. Because exponentials are extremly difficult to understand and we can’t generally model the energy savings, we look at an individual use case. Here, Google’s supremacy experiment from 2019 [2] is used to compare the energy impact of a quantum computer versus a classical supercomputer. Google showed that their Sycamore chip (left) was able to perform a task which is intractable for the Oak Ridge NL Summit Supercomputer (right). Although the problem was of no direct practical use, the experiment showed the potential of the exponential speedup of quantum algorithms
-
- Photograph of the Sycamore processor. (Erik Lucero, Research Scientist and Lead Production Quantum Hardware)
-
- Summit Supercomputer at the Oak Ridge National Laboratory.
In the experiment Google used their 54 superconducting qubit processor Sycamore to sample random quantum circuits. A task which took around 200 seconds, using a total of 1,4 kWh in electrical energy, estimating the full system to require ~25 kW power. For the classical execution the Oak Ridge National Lab Summit Supercomputer was used. In 2019 this was the most powerful computer worldwide, using 13 MW power to compute with 200 petaFLOPS speed [3]. However, it was not feasible to run the algorithm for the largest considered problem size, since Google estimates it would need circa 10.000 years. In numbers this computation would require 41 ExaWatthours (= 41*1015 kWh).
We need to note that the algorithm in question has no practical applications and since then Google’s claims have been challenged by more efficient classical algorithms (for example: ). Nonetheless, the energy savings are representive and highlight the exponential speedup offered by quantum algorithms in certain cases. Using current figures for CO2 emissions per kWh Sycamore would only emit 0,5kg, while running this experiment classically would contribute prohibitive 1,6*1018 kg.
Energy Savings in Quantum Chip Testing with OrangeQS equipment
Quantum Computing also requires a paradigm shift in how to test chips and is especially challenging for high-throughput testing. OrangeQS develops dedicated test equipment solutions to tackle this problem, reducing the time, cost and energy needed to characterize qubits. This year we will launch the first utility-scale test system capable of testing 100+ qubit chips, which serves as benchmark in this analysis.
The following case study compares the energy footprints per tested qubit from a state of the art reference to the OrangeQS industry system as of today, as well as towards a future generation derived from our tech roadmap. The reference describes a typical setup with multiple dilution refrigerators which is in use today, but not optimized for testing purposes.
This table details the improvements for the energy footprint per tested qubit. We estimate a 6x improvement of the OrangeQS industry system 2024 vs. the reference, and a 200x improvement for the future generation system around 2028.
Based on the numbers shown in the table above, we see the main improvements stem from putting more chips in one fridge and testing them significantly faster and in parallel. We enable this via dedicated hardware and switching solutions, as well as our automated calibration software. According to the IBM roadmap the 408-qubit processor Crossbill will be introduced in 2024. When testing 408 qubits the current OrangeQS Industry System can save over 68 MWh in energy, equivalent to 5.360 USD with current electricity costs. Considering that millions of physical qubits will be needed to run useful quantum algorithms, high-throughput testing will be a key bottleneck for economic feasibility.
We conclude that quantum computing is a promising alternative to address the surging electricity needs for data centres. Linked to the energy savings is a significant impact on the CO2 emissions when compared to classical CMOS technology. For now, it remains an open question for which practical use cases the exponential speed up can be implemented, and how broad we can practically achieve this order of magnitude power reduction. Very clear on the other hand is that we need more efficient test systems to handle larger qubit counts. The first OrangeQS Industry System and subsequent generations will address this need and keep the testing throughput on par.
[1] Goldman Sachs Report https://www.goldmansachs.com/intelligence/pages/gs-research/ai-data-centers-global-power-surge-and-sustainability-impact/report.pdf
[2] Supremacy Google Paper; Arute et al. https://www.nature.com/articles/s41586-019-1666-5
[3] Summit Supercomputer https://www.ornl.gov/news/ornl-launches-summit-supercomputer
[4] Electricity Prices US https://www.eia.gov/electricity/monthly/epm_table_grapher.php?t=epmt_5_6_a
