_(cropped).jpg)
In a simulated LLaMA-3.1-8B test on HellaSwag, a commonsense-reasoning benchmark, a new quantization method reached 55.36% accuracy, versus 35.28% for GPTQ under the same memristor-noise protocol. The language model was not run on the fabricated chip.
Memristor crossbars can perform matrix operations where data is stored, reducing the need to shuttle model weights between memory and a processor. A cell’s conductance is the electrical value that represents a weight. Programming variation can leave it different from the intended value, shifting stored weights and lowering a model’s score.
The researchers call their post-training method bit-sliced quantization (BSQ). It represents groups of weights as several binary bit planes plus scaling factors. A bit plane is one layer of binary values in that representation; each plane is mapped to cells at either high or low conductance. The current from each plane is weighted and combined to produce the matrix-operation result. The original model does not need retraining, although the method optimizes the bit representation after training.
With LLaMA-3.1-8B weights quantized to three bits, BSQ reached 55.36% accuracy on HellaSwag under the paper’s simulated noise, compared with 35.28% for GPTQ using the same noise protocol. HellaSwag tests commonsense reasoning, and that 20.08-percentage-point gap applies to this benchmark; results on the smaller handwritten-digit and sentiment models varied by task.
To calibrate the noise model, the team fabricated and characterized a 40-nanometer, 512-by-512 tantalum-oxide (TaOx) memristor crossbar, then used measured cell variation to set simulated perturbations. Those measurements give the simulations a physical reference, but the paper reports no language-model benchmark run on the array. It therefore does not yet demonstrate on-chip LLM accuracy, latency or energy use.
Sources
- Yu et al., “Bit-sliced quantization for robust memristor-based compute-in-memory”, npj Unconventional Computing (2026).
- Zellers et al., “HellaSwag: Can a Machine Really Finish Your Sentence?” (ACL 2019).
- Image: U.S. Department of Energy / NETL memristor device, public domain. Illustrative image of a separate carbon-enabled device, not the TaOx array in this paper.