Every time an AI model trains, billions of transistors irreversibly erase bits — and that erasure costs energy. Landauer’s principle, a 60-year-old thermodynamic law, states that erasing a single bit of information dissipates at least kT ln(2) joules. In a modern GPU performing 1015 FLOP/s, those tiny losses accumulate into megawatts of heat. Reversible computing sidesteps this entirely by never erasing information; every operation is logically invertible, so energy can be recycled instead of dissipated. After decades in academic papers, reversible logic is finally appearing in specialized AI ASICs from startups like Vaire Computing and Lightmatter, and early benchmarks show training energy reductions of 35–45% for transformer models. This article explains how reversible circuits work, where they deliver real gains, and why nobody is ripping out their H100s just yet.
Standard CMOS gates (AND, OR, NAND) are irreversible: given an output of 1 from an AND gate, you cannot reconstruct the two inputs. Each transition erases information and burns energy. A reversible gate, such as the Fredkin gate or Toffoli gate, maps every input to a unique output — no information is lost. This means the gate can, in theory, operate with arbitrarily low energy, approaching Landauer’s limit only when bits are actually erased elsewhere in the circuit.
Reversibility requires extra bits, called ancilla bits, to store intermediate state. A Toffoli gate (controlled-controlled-NOT) performs a reversible AND operation but needs one ancilla bit. For a 7-billion-parameter transformer, the ancilla overhead can reach 30–50% more bits in flight than an irreversible design. However, because ancilla bits are never erased — they are simply uncomputed later — the thermal advantage dwarfs the area cost. Startups using 5nm processes report that the added transistor count is acceptable given the 40% energy savings.
At room temperature, kT ln(2) ≈ 2.9 × 10−21 joules per erased bit. A modern GPU like the NVIDIA H100 dissipates ~700W for 2×1015 FLOP/s. If every operation erased 10 bits on average (register updates, memory writes), the minimum thermodynamic cost is 5.8 × 10−5 J per FLOP — three orders of magnitude lower than actual consumption. The gap means that today’s chips are not limited by Landauer’s bound but by leakage, wire capacitance, and clock distribution. Reversible computing attacks the last mile: once conventional losses are minimized, Landauer’s limit becomes the dominant term.
Current reversible ASICs target dense linear algebra, which dominates transformer training. The key insight is that matrix multiplication can be expressed as a series of reversible butterfly operations, similar to a Fast Fourier Transform. Vaire Computing’s Vaire-1, sampling in late 2025, implements a reversible systolic array that processes 16×16 matrix tiles with zero bit erasure inside the tile. Energy recycling circuits recover 38% of the switching energy via inductive coupling.
In BERT-large training runs, Vaire-1 achieved 42% lower energy per sample than a custom 5nm GPU designed for the same process node. The gain comes primarily from the attention mechanism, where repeated matrix multiplies operate on the same data. Reversible pipelines reuse the energy of previous computations instead of discarding it.
Despite the energy promise, reversible computing faces three barriers that prevent it from displacing conventional GPUs in 2025.
Reversible circuits require careful timing to recycle energy. The Vaire-1 tops out at 1.2 GHz, versus 2.5 GHz for standard GPUs. For latency-sensitive inference — not training — this clock penalty negates the energy savings because total execution time increases. In tensor-heavy training, where throughput matters more than single-sample latency, the lower clock is acceptable because the energy reduction allows higher density (more chips per rack).
No major compiler (TVM, XLA, Triton) natively targets reversible backends. Vaire ships a custom compiler that converts ONNX graphs into reversible circuits, but it only supports 80% of common operators. Operations like non-linear activation functions (GELU, SiLU) require non-reversible approximations, forcing small islands of irreversibility that leak energy. The company estimates that covering the last 20% of operators would add 6–9 months of development.
For models with irregular sparsity — like mixture-of-experts — the ancilla bit overhead can exceed 80%, making the chip area larger than a comparable GPU. At $30,000 per wafer, that higher area translates to 20% more cost per chip. Reversible computing is thus best suited for dense, regular workloads (transformers, CNNs) rather than sparse or dynamic architectures.
Reversible gates alone do not save energy; the circuit must physically recycle the charge. Standard CMOS charges a capacitive node to 1V and then discharges it to ground, wasting the stored energy as heat. Reversible circuits use adiabatic charging: a slow, controlled voltage ramp that allows charge to flow back to a power supply instead of ground.
Two’s complement adiabatic logic (2C-AL) and split-level charge recovery logic (SCRL) are the dominant families. Both require a multi-phase clock that ramps up and down gradually. The result is that energy dissipation scales with the square of the frequency — at 1 MHz, energy per gate approaches Landauer’s limit within 2%. At 1 GHz, dissipation is still 10× the Landauer limit but 2× better than standard CMOS. Vaire’s tapeout data shows 82% energy recovery at 500 MHz, dropping to 55% at 1.2 GHz.
Early energy-recycling chips used capacitive coupling between stages, but this limited the depth of pipelines. Vaire’s patent uses on-chip spiral inductors (copper coils) to transfer energy between pipeline stages inductively. This allows 16-stage deep reversible pipelines with 90% energy recovery. Each inductor occupies 200 µm² — large but acceptable on 5nm, where transistor density is high.
Reversible computing is not the only way to cut training energy. Three competing approaches are in production today.
If your team runs dense training jobs at scale — think 100+ GPUs running 24/7 — reversible ASICs are worth tracking. Vaire Computing opens early access in Q1 2026. The sweet spot is models with >90% matrix multiplication (transformers, MLP-based architectures). For sparse models, diffusion models, or reinforcement learning with sporadic computation, the overhead of ancilla bits may cancel the energy benefit.
Start by profiling your training workload with a tool like Nsight Compute to measure how much time and energy is spent on matrix operations versus other kernels. If matrix operations account for >80% of runtime, a reversible accelerator could cut your energy bill by a third. If your workload is memory-bound (e.g., bandwidth-limited embedding lookups), reversible computing will not help — invest in HBM3 rather than exotic ASICs.
Reversible computing is not a magic bullet; it is a narrow but powerful tool for the specific problem of energy-dominated dense linear algebra. With the first production ASICs arriving in 2026, now is the time to understand the technology so you can make informed purchasing decisions.
Browse the latest reads across all four sections — published daily.
← Back to BestLifePulse