- ~10,000xclaimed energy reduction versus GPU-based performance on a simple image-generation benchmark
- Up to 1000xenergy efficiency gain claimed by Normal Computing’s CN101 chip on linear algebra and matrix operations in AI training
- Oct 2025when Extropic announced what it calls the world’s first scalable probabilistic computer
Every processor discussed elsewhere in this cluster — GPUs, custom silicon, analog in-memory accelerators — is fundamentally deterministic: given the same inputs, they produce the same outputs, every time. Thermodynamic and probabilistic computing chips reject that premise entirely for a specific class of problems. Rather than fighting against the physical randomness inherent in transistor-level electronics (thermal noise, in particular), these architectures embrace it, using specially designed circuits to generate genuine, programmable randomness in hardware and perform probabilistic computation directly, rather than simulating randomness in software running on deterministic hardware — which is what every conventional chip running a probabilistic algorithm today actually does.
How a probabilistic chip actually works
Rather than relying on deterministic calculations like conventional processors, these architectures use controlled randomness to perform probabilistic computations directly in hardware. Extropic’s approach, described in the paper published in July 2026 following its October 2025 announcement of the first scalable probabilistic computer, uses a grid of thousands of interconnected semiconductor-based components built around conventional CMOS transistors, with specially designed transistor circuits generating programmable random numbers as a first-class hardware primitive rather than a software approximation. That distinction — genuine hardware randomness versus software-simulated randomness — is the entire basis of the energy efficiency claim, because generating true randomness in hardware is dramatically cheaper than the elaborate pseudo-random number generation and iterative sampling a conventional processor needs to approximate the same probabilistic behavior in software.
| Conventional GPU running probabilistic AI | Native probabilistic/thermodynamic chip | |
|---|---|---|
| How randomness is produced | Simulated in software via pseudo-random number generation | Generated directly as a hardware primitive using physical noise |
| Underlying computation model | Deterministic arithmetic approximating probability distributions | Physical process directly encodes and samples probability distributions |
| Energy profile on suitable workloads | Baseline | Claimed up to ~10,000x more efficient on some generative benchmarks |
| Applicable workloads | General-purpose, any algorithm | Specifically generative and probabilistic AI algorithms |
Two related but distinct approaches
Thermodynamic computing and probabilistic computing overlap conceptually but aren’t identical. Thermodynamic computing systems, the subject of active research including Normal Computing’s CN101 chip, are specifically aimed at efficiently solving linear algebra and matrix operations — the computational core of AI training — using physics-based hardware rather than digital simulation, claiming up to 1000x energy efficiency gains on those specific operations. Extropic’s probabilistic computing approach targets a related but distinct goal: running generative AI sampling algorithms directly using hardware-generated randomness. Both are physics-based alternatives to conventional digital computation, and both are explicitly positioned for AI workloads rather than as general-purpose replacements for a CPU or GPU.
- 1. Physical randomness as a resourceSpecially designed CMOS transistor circuits generate genuine programmable randomness rather than simulating it.
- 2. Probability distributions encoded directlyThe hardware’s physical behavior directly represents and samples from probability distributions relevant to the algorithm.
- 3. Generative AI sampling runs nativelyAlgorithms that need to sample from complex distributions — core to generative AI — execute without the overhead of software-simulated randomness.
- 4. Dramatic energy savings on matching workloadsClaimed efficiency gains of up to 10,000x specifically on generative benchmarks suited to this computation style.
How embracing hardware randomness bypasses a major AI energy cost
The efficiency claim is workload-specific — it applies to generative and probabilistic algorithms, not to deterministic arithmetic tasks a conventional chip is already efficient at.
Why this isn’t a general AI accelerator
The obvious question is why every AI workload doesn’t simply move to probabilistic hardware if the energy savings are this dramatic. The answer is scope: the efficiency gain is specific to problems that are naturally probabilistic or generative in structure — sampling from a distribution, generating diverse outputs, exploring uncertain state spaces. A deterministic task like exact matrix inversion, or an inference workload that needs one specific, reproducible answer rather than a sampled one, doesn’t benefit from this architecture the same way, and may not be expressible on it efficiently at all. This makes thermodynamic and probabilistic chips a genuinely promising specialized accelerator category rather than a coming replacement for GPUs across all AI workloads.
Evaluate for generative and sampling-heavy AI workloads Where the win is real
Right for workloads that are naturally probabilistic — image generation, sampling-based inference, and similar tasks where Extropic and Normal Computing’s efficiency claims specifically apply.
Track this as an early, unproven specialized category Necessary caution
These are 2025-2026-era first demonstrations from young companies — genuinely interesting physics, but without the years of production deployment history that conventional accelerators have.
Expect probabilistic chips to replace GPUs for deterministic AI inference Wrong fit
Workloads requiring exact, reproducible outputs don’t benefit from hardware built around embracing physical randomness, and may not map onto this architecture efficiently at all.
Common mistakes
Common mistake
Treating a 10,000x energy efficiency claim as a general statement about AI compute rather than a workload-specific result on a narrow generative benchmark. The gain applies specifically to problems well-suited to hardware-native probabilistic computation, not to AI compute broadly.
Common mistake
Conflating thermodynamic computing (aimed at linear algebra and matrix operations) with probabilistic computing (aimed at generative sampling) as the same technology. They’re related physics-based approaches with meaningfully different target workloads.
What worked
Organizations exploring this space treated it as a research-stage specialized accelerator bet — worth prototyping against genuinely sampling-heavy or generative workloads specifically — rather than assuming near-term general applicability across their AI infrastructure.
Glossary
- Probabilistic computing
- A computing architecture that uses genuine hardware-generated randomness to perform computation directly, rather than simulating probability in software on deterministic hardware.
- Thermodynamic computing
- A physics-based computing approach that uses natural physical processes to efficiently solve computational problems, particularly linear algebra and matrix operations central to AI training.
- Programmable random number generation (hardware)
- Specially designed transistor circuits that produce genuine randomness as a physical hardware primitive, as opposed to pseudo-random numbers computed algorithmically in software.
- Generative sampling
- The process of drawing outputs from a probability distribution, the core operation in generative AI algorithms that probabilistic computing chips are designed to accelerate.
Key takeaways
- Probabilistic computing chips use genuine hardware-generated randomness to run generative AI algorithms directly, rather than simulating probability in software on deterministic hardware.
- Extropic announced the world’s first scalable probabilistic computer in October 2025, claiming roughly 10,000 times less energy than GPU-based performance on a simple image-generation benchmark.
- Normal Computing’s CN101 thermodynamic computing chip targets linear algebra and matrix operations specifically, claiming up to 1000x energy efficiency gains on AI training workloads.
- The efficiency gains are workload-specific, applying to naturally probabilistic and generative algorithms rather than to deterministic AI compute broadly.
- Thermodynamic computing and probabilistic computing are related but distinct approaches, targeting linear algebra operations and generative sampling respectively.
- Workloads requiring exact, reproducible outputs don’t benefit from this architecture the same way generative and sampling-heavy workloads do, and may not map onto it efficiently at all.
- Both technologies remain early-stage, 2025-2026-era first demonstrations from young companies, worth evaluating for suitable workloads but not yet a proven general AI accelerator category.
FAQs
What is a probabilistic computing chip?
It’s a chip that uses genuine hardware-generated randomness, produced by specially designed transistor circuits, to perform probabilistic computations directly, rather than simulating randomness in software running on conventional deterministic hardware.
How much energy can probabilistic computing actually save?
Extropic claims its probabilistic computer can match GPU-based performance on a simple image-generation benchmark while using roughly 10,000 times less energy, though this figure is specific to that benchmark and workload type rather than a general AI compute claim.
What is the difference between thermodynamic computing and probabilistic computing?
Thermodynamic computing, as pursued by Normal Computing’s CN101 chip, targets linear algebra and matrix operations central to AI training with claimed efficiency gains up to 1000x, while probabilistic computing, as pursued by Extropic, targets generative AI sampling algorithms specifically.
Can probabilistic computing chips replace GPUs for all AI workloads?
No. The efficiency gains are specific to naturally probabilistic and generative workloads; tasks requiring exact, reproducible outputs don’t benefit the same way and may not map efficiently onto this architecture at all.
When did Extropic announce its probabilistic computer?
Extropic announced what it calls the world’s first scalable probabilistic computer in October 2025, with a paper describing the system published in July 2026.
How does a probabilistic chip generate randomness differently than a conventional processor?
Conventional processors simulate randomness algorithmically using pseudo-random number generation in software; probabilistic chips use specially designed transistor circuits to generate genuine, programmable randomness as a direct hardware primitive.
Is thermodynamic and probabilistic computing a proven, mature technology?
No, it remains early-stage — both Extropic’s and Normal Computing’s chips are 2025-2026-era first demonstrations from young companies, genuinely promising but without the years of production deployment history that conventional AI accelerators have.
What kind of AI workloads are best suited to probabilistic computing?
Naturally probabilistic and generative tasks — sampling from complex distributions, generating diverse outputs, and exploring uncertain state spaces — rather than deterministic tasks requiring one specific, exactly reproducible answer.
References
- Tom’s Hardware — “World’s first ‘thermodynamic computing chip’ reaches tape out”
- EE Times — “Probabilistic Computing Is Already Here; Here Is How It Works”
- Extropic — “Thermodynamic Computing: From Zero to One”
- The Quantum Insider — “Researchers Propose Thermodynamic Computing Architecture”
Related reading on The Tech Trends: Analog In-Memory Computing, Photonic Computing for AI, HBM Memory Bandwidth, and Inference vs Training Economics.
