CPTC: a 64-byte-table quantization format claims near-ideal accuracy, built into AWS Trainium 4
An arXiv paper proposes Computed Permutation Trellis Coding, a low-bit tensor format whose small decoder reportedly approaches the information-theoretic ideal and ships in AWS silicon.
In plain words
Quantization shrinks AI model numbers to fewer bits so they take less memory, but simple per-value formats lose accuracy. CPTC instead encodes groups of values jointly using a decoder that needs only two multiplications and a 64-byte lookup table. Its authors report it nearly matches the theoretical best at 4 bits and runs at full speed on AWS's Trainium 4 chip.
What was announced
The paper, an arXiv abstract by the CPTC authors, targets tensors that are written once and read many times, where formats like INT4 and FP4 waste bits against the information-theoretic ideal. Trellis codes close that gap by encoding values jointly, but earlier decoders required either a large lookup table or a hash-based design that limited what they could decode. CPTC's decoder instead multiplies its state by two coefficients and uses the top bits of the products to index a 64-byte table, with rarely used table entries holding far-out tail values.
According to the abstract, one decoder covers every bit width from 1 to 6, reproduces standard integer and floating-point formats exactly, and comes within 0.07 bits of the ideal at 4 bits. It reportedly matches a random-permutation trellis code with a table thousands of times larger and exceeds every QTIP codebook. In LLM weight quantization without fine-tuning, the authors report the lowest KL divergence to the unquantized model at every tested bit width on Llama 3.1 8B and 70B Instruct, beating QTIP's codebooks by up to 27 percent. The decoder has been built into the AWS Trainium 4 accelerator, running at native FP4 and FP8 throughput for a negligible fraction of the systolic array's area, with fused encoding kernels quantizing an 8B model's linear layers in under two minutes on one GPU.
What it does not show
- All claims come from the paper's own abstract; no independent evaluation or peer review is reflected in the supplied evidence.
- The supplied source is an abstract excerpt, so implementation details, full benchmarks, and comparisons beyond those stated are not covered.
- Trainium 4 throughput and area figures rest on cycle-level simulation for encoding, not necessarily measured production hardware runs.
For security and IT teams
- Scope
- The reported Trainium 4 integration and throughput figures are stated by the paper's authors in an abstract and are not independently verified in the supplied evidence.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.