Pith. sign in

REVIEW 8 cited by

Scalable MatMul-free Language Modeling

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2406.02528 v7 pith:IF3DBONC submitted 2024-06-04 cs.CL

classification cs.CL
keywords modelsmemoryapproachduringlanguagellmsmatmulmatmul-free
verification ladder T0 review T1 audit T2 compute T3 formal

Signed reviews

No signed human review yet.

0 comments
read the original abstract

Large Language Models (LLMs) have fundamentally altered how we approach scaling in machine learning. However, these models pose substantial computational and memory challenges, primarily due to the reliance on matrix multiplication (MatMul) within their attention and feed-forward (FFN) layers. We demonstrate that MatMul operations can be eliminated from LLMs while maintaining strong performance, even at billion-parameter scales. Our MatMul-free models, tested on models up to 2.7B parameters, are comparable to state-of-the-art pre-trained Transformers, and the performance gap narrows as model size increases. Our approach yields significant memory savings: a GPU-efficient implementation reduces memory consumption by up to 61% during training and over 10x during inference. When adapted for a multi-chip neuromorphic system, the model leverages asynchronous processing to achieve 4x higher throughput with 10x less energy than edge GPUs.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 8 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. All-optical temporal integration mediated by subwavelength heat antennas

    physics.optics 2025-05 conditional novelty 6.0 of 10

    A ring resonator with titanium nano-heaters absorbs control light at standing-wave antinodes, heats up, and shifts a probe wavelength, performing all-optical temporal integration and nonlinear activation in a single device.

  2. PROM: Prioritize Reduction of Multiplications Over Lower Bit-Widths for Efficient CNNs

    cs.CV 2025-05 conditional novelty 6.0 of 10

    PROM quantizes pointwise convolutions to ternary weights and all other layers to 8 bits, converting the dominant CNN operations to additions and yielding order-of-magnitude estimated energy savings at similar accuracy.

  3. Neuromorphic Principles for Efficient Large Language Models on Intel Loihi 2

    cs.NE 2025-02 reject novelty 6.0 of 10

    A 370M MatMul-free LLM is mapped onto Intel Loihi 2 and reported to achieve up to 3x higher generation throughput with about 2x less energy than transformer LLMs on an edge GPU, based on preliminary measurements that ...

  4. RandLoRA: Full-rank parameter-efficient fine-tuning of large models

    cs.CL 2025-02 conditional novelty 6.0 of 10

    RandLoRA achieves full-rank weight updates in parameter-efficient fine-tuning by learning diagonal scalings over fixed random low-rank bases, outperforming LoRA across vision and language tasks.

  5. Learnable Sparsification of Die-to-Die Communication via Spike-Based Encoding

    cs.AR 2025-01 conditional novelty 5.0 of 10

    A hybrid accelerator that confines spiking layers to die-to-die interfaces offers simulated latency and energy gains over all-ANN designs while matching accuracy.

  6. Threshold Neuron: A Brain-inspired Artificial Neuron for Efficient On-device Inference

    cs.LG 2024-12 conditional novelty 5.0 of 10

    The paper presents a multiplication-free threshold neuron that builds unified networks with large claimed hardware savings, but the fairest comparisons are still missing.

  7. Sparse Mutual Information Graph Averaging for Improving Random Indexing Embeddings

    cs.CL 2026-08 conditional novelty 4.0 of 10

    On 272 family-category analogy questions from a fairytales corpus, averaging random-indexing vectors over a top-50 PPMI graph raised accuracy from 19.4% to 30.7% across five seeds, while the same averaging hurt strong...

  8. An Extra RMSNorm is All You Need for Fine Tuning to 1.58 Bits

    cs.LG 2025-05 reject novelty 3.0 of 10

    Fine-tuning a pretrained LLM with extra RMSNorm before each quantized linear layer and a gradual two-phase quantization schedule can reach ternary 1.58-bit weights, according to the paper, but the supporting evidence ...

Pith tools