What actually runs: a measurement study of language model placement and decode speed on the Apple Neural Engine

• URL: http://arxiv.org/abs/2608.22110v1
• Date: Sat, 22 Aug 2026 21:22:24 GMT
• Title: What actually runs: a measurement study of language model placement and decode speed on the Apple Neural Engine
• Abstract summary: We train matched models across size and precision, with quantized checkpoints byte-identical in structure to their fp16 counterparts.<n>We read the ANE's memory-controller byte counters during inference, establishing what actually ran rather than what the compiler intended.
• Score: 0.0
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: We ask what gets a language model onto the Apple Neural Engine (ANE) and what makes it fast there, and we answer with three measurements. We sweep a 64-shape matrix of LLM primitives that varies how a computation is expressed while holding what it computes fixed, recording per-operation device support. We then train matched models across size and precision, with quantized checkpoints byte-identical in structure to their fp16 counterparts, so every deployment measurement is of a real trained artifact. And we read the ANE's memory-controller byte counters during inference, establishing what actually ran rather than what the compiler intended. We support every headline claim with at least two of these three measurement paths. We find that placement is a property of how a computation is expressed, not of what it computes: a fused RMSNorm is fully ANE-eligible while its arithmetically identical decomposition is CPU-only. Weight encoding gates the accelerator: CoreML assigns a 25.85M-parameter conv-heavy fp16 model entirely to the CPU (our counters confirm zero bytes through the engine), while the same graph in int8 or 2-bit returns to ~83% residency and runs 1.8-2.2x faster, and a smaller 22.29M all-attention fp16 model sits at 98.9%. Decode cost is bytes streamed per token, at a constant ~0.77 fraction of nominal encoding width across fp16, int8 and 2-bit. The smallest and fastest models we measured are ternary, and at matched size the operator mix barely moves either axis: every resident 25M ternary model lands within 10.0-10.8 MB and 0.62-0.64 ms/token. The headline pair is half-attention ternary at 25M (10.5 MB, 0.63 ms) and 50M (16.8 MB, 0.86 ms) - 9.8x and 6.1x smaller, 3.0x and 2.2x faster than the conv-heavy fp16 design this work began with. From these measurements we draw a design procedure: choose the encoding first, then spend the byte budget on parameters.

Related papers

• Unfolding the Leech Lattice: Fused Multi-Shell Decoding and VRAM Layouts for 2-Bit LLM Weights [8.550323073861415]
  Leech-lattice vector quantization holds the strongest reported 2-bit quality under its own evaluation protocol.<n>This paper supplies one and measures its serving cost for decode-phase GEMV at batch 1.
  arXiv  Detail & Related papers  (2026-09-02T14:26:06Z)
• Daedalus-150M: A Convolution-Attention Hybrid Designed for CPU Inference [0.0]
  Small language models are usually built like large ones and then squeezed onto a CPU afterwards.<n>We did the opposite: we fixed the target first, one user, one token at a time, 4-bit weights, ordinary CPU, and chose the architecture to suit it.<n>The model scores 47.31 on a five-task benchmark against a bar of 42.20 that was fixed before training began.<n>It beats GPT-2 124M, Pythia-160M, OPT-125M and GPT-neo-125M, all trained on three to six times more data, and exceeds MobileLLM-125M'
  arXiv  Detail & Related papers  (2026-08-20T16:09:43Z)
• VibeVoice-ASR-BitNet Technical Report [86.93881874028209]
  VibeVoice-ASR-BitNet is a compressed variant of VibeVoice-ASR optimized for real-time inference on edge CPUs.<n>We implement a progressive quantization-aware training strategy to preserve accuracy under aggressive compression.<n> VibeVoice-ASR-BitNet is 1.6-2.3x faster than Whisper at comparable model sizes.
  arXiv  Detail & Related papers  (2026-07-23T09:08:04Z)
• A Modern Multimodal Assistant on a 6 GB 2011 GPU: Stage-Validated, All-GPU CUDA Inference for Fermi [0.0]
  This report keeps the hardware and asks what a model that fits can do.<n>We deploy Mini-V-4.6, a modern multimodal assistant pairing a SigLIP2 encoder and window-attention merger.<n>The system answers an image question end-to-end in 1.7s.
  arXiv  Detail & Related papers  (2026-07-16T04:48:44Z)
• Spike-Aware C++ INT8 Inference for Sparse Spiking Language Models on Commodity CPUs [8.419155861590548]
  Spiking language models expose activation sparsity that dense Transformer runtimes do not directly exploit.<n>We implement a C++ CPU inference runtime that treats sparse binary spike states as an execution primitive.<n>Spike-aware execution can improve CPU throughput and memory behavior for sparse language models.
  arXiv  Detail & Related papers  (2026-06-02T02:03:37Z)
• Chronicals: A High-Performance Framework for LLM Fine-Tuning with 3.51x Speedup over Unsloth [0.0]
  We present Chronicals, an open-source training framework achieving 3.5x speedup over Unsloth.<n>We provide complete mathematical foundations: online softmax correctness, FlashAttention IO complexity O(N2 d2 M-1), LoRA+ learning rate gradient approximations.
  arXiv  Detail & Related papers  (2026-01-06T00:00:55Z)
• You Only Cache Once: Decoder-Decoder Architectures for Language Models [132.4064488592704]
  We introduce a decoder-decoder architecture, YOCO, for large language models. YOCO only caches key-value pairs once. The overall model behaves like a decoder-only Transformer, although YOCO only caches once.
  arXiv  Detail & Related papers  (2024-05-08T17:57:39Z)
• QMoE: Practical Sub-1-Bit Compression of Trillion-Parameter Models [64.34635279436054]
  Mixture-of-Experts (MoE) architectures offer a general solution to the high inference costs of large language models (LLMs) via sparse routing. We present a solution to this memory problem, in form of a new compression and execution framework called QMoE.
  arXiv  Detail & Related papers  (2023-10-25T17:24:53Z)
• Binary and Ternary Natural Language Generation [24.295815261826153]
  Ternary and binary neural networks enable multiplication-free computation. They promise multiple orders of magnitude efficiency gains over full-precision networks. However, such networks have proven very difficult to optimize. We show first ternary and binary transformer models on the downstream tasks of summarization and machine translation.
  arXiv  Detail & Related papers  (2023-06-02T18:01:02Z)
• The case for 4-bit precision: k-bit Inference Scaling Laws [75.4335600212427]
  Quantization methods reduce the number of bits required to represent each parameter in a model. The final model size depends on both the number of parameters of the original model and the rate of compression. We run more than 35,000 zero-shot experiments with 16-bit inputs and k-bit parameters to examine which quantization methods improve scaling for 3 to 8-bit precision.
  arXiv  Detail & Related papers  (2022-12-19T18:48:33Z)
• LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale [80.86029795281922]
  We develop a procedure for Int8 matrix multiplication for feed-forward and attention projection layers in transformers. A 175B parameter 16/32-bit checkpoint can be loaded, converted to Int8, and used immediately without performance degradation.
  arXiv  Detail & Related papers  (2022-08-15T17:08:50Z)
• I-BERT: Integer-only BERT Quantization [78.43819756382103]
  We propose I-BERT, a novel quantization scheme for Transformer based models. I-BERT performs an end-to-end integer-only BERT inference without any floating point calculation. We show that for both cases, I-BERT achieves similar (and slightly higher) accuracy as compared to the full-precision baseline.
  arXiv  Detail & Related papers  (2021-01-05T02:42:58Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.