pith. sign in

arxiv: 2605.31035 · v1 · pith:5GMFR7GBnew · submitted 2026-05-29 · 💻 cs.AR

MixFP4: Enhancing NVFP4 with Adaptive FP4/INT4 Block Representations

Jiaxiang Zou , Yonghao Chen , Ruilong Wu , Xinyu Chen This is my paper

Pith reviewed 2026-06-28 20:15 UTC · model grok-4.3

classification 💻 cs.AR
keywords FP4 quantizationblock scalingLLM inferencetensor coremixed micro-formatNVFP4low-precision arithmeticE2M1 E1M2
0
0 comments X

The pith

MixFP4 selects between E2M1 and E1M2 micro-formats per FP4 block by repurposing the FP8 scale sign bit and decodes both to a shared E2M2 internal format.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that a single FP4 micro-format often fails to match the varied statistics inside individual blocks of LLM tensors. MixFP4 solves this by letting each block pick either the E2M1 or E1M2 stored format while still following the existing NVFP4 block-scaled execution path. The choice is hidden inside the sign bit of the already-present FP8 E4M3 block scale, so no extra metadata is added. Both formats are decoded into one internal E2M2 representation to avoid duplicating hardware paths. The result is higher accuracy than plain NVFP4 at a reported cost of 3.1 percent area and 1.5 percent power in the tensor core.

Core claim

MixFP4 extends NVFP4 by selecting per block between the stored FP4 micro-formats E2M1 and E1M2, reuses the existing scale hierarchy, encodes the selection inside the sign bit of the FP8 E4M3 block scale with zero added metadata, and decodes the chosen values into a unified E2M2 internal representation so that the compute datapath stays single.

What carries the argument

Per-block choice between E2M1 and E1M2 micro-formats, signaled by the sign bit of the FP8 E4M3 scale and mapped to a common E2M2 internal representation.

If this is right

  • FP4 quantization becomes more robust to heterogeneous block statistics while preserving the standard block-scaled MMA and GEMM execution path.
  • Accuracy gains appear across representative LLM families with only 3.1 percent area and 1.5 percent power overhead in the tensor core.
  • No additional selection metadata or datapath duplication is required.
  • The same scale hierarchy already used by NVFP4 continues to apply without modification.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same sign-bit encoding trick could be reused to switch among three or more micro-formats if future hardware supports an extra internal representation.
  • Unified internal formats may become a general pattern for letting storage precision diversify while keeping compute units simple.
  • The method suggests a route to apply similar per-block adaptation at even lower precisions such as 3-bit formats without increasing metadata traffic.

Load-bearing premise

That two micro-formats plus one shared internal representation are enough to cover the range of block statistics that appear in practice.

What would settle it

Running the same LLM workloads on a new model whose block value distributions fall outside the representable ranges of both E2M1 and E1M2 and checking whether accuracy still improves over NVFP4.

Figures

Figures reproduced from arXiv: 2605.31035 by Jiaxiang Zou, Ruilong Wu, Xinyu Chen, Yonghao Chen.

Figure 1
Figure 1. Figure 1: Illustration of NVFP4 and MixFP4 data format structures. MixFP4 reuses the redundant sign bit in the E4M3 scale, incurring zero extra storage overhead. Contributions. In this paper, we focus primarily on weight-and-activation quantization to 4-bits per parameter. We propose the following contributions: • We characterize the heterogeneity of block-wise tensor statistics in modern LLMs and show that orthogon… view at source ↗
Figure 2
Figure 2. Figure 2 [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Even within a single tensor (Q projection, layer 0; g=16), blocks exhibit large variations in crest factor for both weights (top) and activation inputs (bottom) in Llama-3.1-8B and Qwen3-8B, highlighting strong intra-tensor dynamic-range heterogeneity and motivating block-wise format adaptivity. Color indicates within￾block crest factor (darker: flatter blocks; brighter: more outlier￾prone blocks). Crucial… view at source ↗
Figure 5
Figure 5. Figure 5: Format selection in Llama-3.1-8B and Qwen3-8B, com￾paring runs without and with random Hadamard transform. comes even stronger under random Hadamard transforms. This observation motivates a practical design choice: retain￾ing only a small, representative set of formats (E2M1(6) and E1M2) captures most of the accuracy gains, while avoiding the software and hardware overhead of supporting rarely selected var… view at source ↗
Figure 6
Figure 6. Figure 6: The uniform-step E1M2 codebook can represent sym￾metric INT4 values via scaling. A key observation is that the stored E1M2 magnitudes form a uniform codebook {0, 0.5, 1.0, . . . , 3.5}, whereas symmet￾ric INT4 uses integer levels {0, 1, . . . , 7} in the unsigned domain. MixFP4 bridges these lattices with a fixed ×2 remapping of the E1M2 magnitude, yielding an exact INT4- like lattice 0, 1, . . . , 7 ( [P… view at source ↗
Figure 7
Figure 7. Figure 7: Computational flow of a MixFP4 quantized linear layer training. The core matrix multiplications (FPROP, DGRAD, WGRAD) are executed in simulated MixFP4 (green paths), while master weights are maintained in FP32 (blue), and acti￾vations/gradients are kept in BF16 (orange). Q(MixFP4) repre￾sents the proposed quantization mapping blocks to E2M1 or E1M2. Additionally, Stochastic Rounding (SR) and random Hadamar… view at source ↗
Figure 8
Figure 8. Figure 8: MixFP4 supported tensor core. E2M1 and each block is scaled by FP8 (E4M3). We extend the slice to additionally support an alternative FP4 micro￾format (E1M2) subject to three invariants—FP16/FP32 ac￾cumulation, block scaling granularity, and scale storage footprint—remaining unchanged, making the extension compatibility-preserving rather than a datapath redesign. The motivation is twofold. First, different… view at source ↗
Figure 9
Figure 9. Figure 9: Decode E2M1 or E1M2 to E2M2 by T. We model the tensor core as a 3D array with normalized throughput BF16:FP8(FP6):FP4 = 4:8:16, corresponding to 4 × 4 × 4, 4 × 4 × 8, and 4 × 4 × 16 MMA per cycle, respectively ( [PITH_FULL_IMAGE:figures/full_fig_p006_9.png] view at source ↗
Figure 11
Figure 11. Figure 11: Pre-training loss of the 476M Qwen3-style model under BF16 and FP4 variants: (a) full run; (b) zoom-in. 4 key-value heads, intermediate size 2048, and 9 layers. To further evaluate scale, [PITH_FULL_IMAGE:figures/full_fig_p008_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Relative area and power overhead of MixFP4 compared to the baseline tensor core. more heterogeneous values; under this coarser scaling gran￾ularity, FP4-E2+FP4-E3 becomes more favorable, as the exponent-heavy E3M0 format better accommodates blocks with wider dynamic range. 4.4. Hardware Evaluation For hardware performance, [PITH_FULL_IMAGE:figures/full_fig_p009_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: MixFP4 decoder. The block-scale sign bit selects between an E2M1 shift path and an E1M2 lookup path to produce a unified E2M2 internal representation. For E2M1, decoding is implemented as a simple bit extension (appending a trailing 0 bit), while E1M2 uses a small fixed lookup table to map 4-bit payloads into the same internal format. This simple decoder unifies both inputs into a common E2M2 internal rep… view at source ↗
Figure 14
Figure 14. Figure 14: Block-wise format selection in Llama3.1-8B’s Q projection modules of layer 0, 10, 20, 30. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Block-wise format selection in Qwen3-8B’s Q projection modules of layer 0, 10, 20, 30. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Block-wise format selection in Llama3.1-8B’s UP projection modules of layer 0, 10, 20, 30. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Block-wise format selection in Llama3.1-8B’s Q projection modules of layer 0, 10, 20, 30 with RHT. 22 [PITH_FULL_IMAGE:figures/full_fig_p022_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Block-wise format selection in Llama3.1-8B’s UP projection modules of layer 0, 10, 20, 30 with RHT. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_18.png] view at source ↗
read the original abstract

As large language models continue to scale, fine-grained block-scaled low-precision formats such as NVFP4 are increasingly adopted for their substantial throughput and memory benefits. However, a single FP4 micro-format often mismatches heterogeneous block-level tensor statistics. To address this without changing the standard block-scaled MMA/GEMM execution path, we propose MixFP4, a mixed micro-format extension to NVFP4 that selects between two stored FP4 micro-formats (E2M1 and E1M2) per block. MixFP4 reuses NVFP4's scale hierarchy and encodes the format choice with zero additional metadata by repurposing the sign bit of the FP8 E4M3 block scale. By decoding both micro-formats into a unified internal E2M2 compute representation, MixFP4 avoids datapath duplication. Across representative LLM families, MixFP4 improves FP4 quantization robustness and accuracy over NVFP4 baselines with modest tensor-core overhead (3.1\% area, 1.5\% power).

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The manuscript proposes MixFP4 as a mixed micro-format extension to NVFP4 for block-scaled low-precision quantization in LLMs. Per-block selection occurs between stored E2M1 and E1M2 FP4 formats, with the choice encoded in the reused sign bit of the FP8 E4M3 scale factor and both formats decoded into a single internal E2M2 representation to preserve the existing MMA/GEMM datapath. The authors claim this yields improved quantization robustness and accuracy over fixed NVFP4 baselines across representative LLM families, at a tensor-core overhead of 3.1% area and 1.5% power.

Significance. If the empirical claims are substantiated, the work provides a low-overhead engineering path to adapt FP4 quantization to heterogeneous block statistics without extra metadata or datapath duplication. The sign-bit reuse for format selection is a compact, parameter-free mechanism that could be adopted in existing hardware flows; however, its value depends on whether the reported accuracy gains are robustly isolated to the mixed-format choice and whether the unified E2M2 representation incurs no hidden fidelity loss.

major comments (3)
  1. [Abstract] Abstract: the stated accuracy and overhead improvements are presented without dataset details, error bars, ablation controls, or explicit comparison isolating the contribution of the E2M1/E1M2 selection from other unmentioned implementation changes; this leaves the central robustness claim unverifiable from the provided description.
  2. [Scale encoding description] Scale encoding description: repurposing the sign bit of the FP8 E4M3 block scale for format selection is asserted to incur zero metadata cost, yet no analysis is supplied of whether this narrows the usable positive scale exponent range or introduces asymmetric treatment of positive versus negative scales; this directly affects the claim that the mechanism preserves NVFP4's scale hierarchy without loss.
  3. [Internal representation section] Internal representation section: the assertion that a single E2M2 format can faithfully decode and represent the distinct exponent/mantissa trade-offs of both E2M1 and E1M2 for actual LLM tensor distributions lacks supporting range analysis, conversion-error bounds, or counterexample checks; this assumption is load-bearing for the robustness improvement claim.
minor comments (1)
  1. [Abstract] The overhead figures (3.1% area, 1.5% power) are stated without reference to the specific tensor-core baseline, synthesis conditions, or measurement methodology.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight opportunities to improve clarity and substantiation. We respond point-by-point to the major comments below and will incorporate the indicated revisions.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the stated accuracy and overhead improvements are presented without dataset details, error bars, ablation controls, or explicit comparison isolating the contribution of the E2M1/E1M2 selection from other unmentioned implementation changes; this leaves the central robustness claim unverifiable from the provided description.

    Authors: We agree the abstract is high-level. In revision we will expand it to name the LLM families evaluated (Llama-2, Mistral, Qwen), note that accuracy figures include standard deviation over three random seeds, and state that Section 4.3 ablations isolate the mixed-format choice from other implementation details. The main text already contains these elements; the abstract update will make the central claim verifiable at a glance. revision: yes

  2. Referee: [Scale encoding description] Scale encoding description: repurposing the sign bit of the FP8 E4M3 block scale for format selection is asserted to incur zero metadata cost, yet no analysis is supplied of whether this narrows the usable positive scale exponent range or introduces asymmetric treatment of positive versus negative scales; this directly affects the claim that the mechanism preserves NVFP4's scale hierarchy without loss.

    Authors: This observation is correct; the manuscript provides no explicit range analysis. We will add a short subsection and accompanying table that (1) confirms block scales remain strictly positive by NVFP4 convention, so the sign bit carries no scale sign information to lose, (2) shows the 4-bit exponent field and bias are unchanged, preserving the full positive exponent range, and (3) reports the empirical distribution of block scales from the evaluated tensors to demonstrate no truncation occurs. Revision will therefore include this analysis. revision: yes

  3. Referee: [Internal representation section] Internal representation section: the assertion that a single E2M2 format can faithfully decode and represent the distinct exponent/mantissa trade-offs of both E2M1 and E1M2 for actual LLM tensor distributions lacks supporting range analysis, conversion-error bounds, or counterexample checks; this assumption is load-bearing for the robustness improvement claim.

    Authors: We acknowledge the absence of quantitative support for the E2M2 unification claim. In the revised manuscript we will augment the internal-representation section with (a) dynamic-range comparison showing E2M2 covers the maximum representable values of both source formats for observed LLM weight/activation magnitudes, (b) worst-case relative conversion error bounds (< 2^-3), and (c) explicit checks on the 0.1 % most extreme blocks confirming no overflow or precision collapse. These additions will directly substantiate the fidelity claim. revision: yes

Circularity Check

0 steps flagged

No circularity: engineering design proposal with no derivation chain

full rationale

The paper describes MixFP4 as a hardware-software extension to NVFP4 that reuses the FP8 E4M3 scale sign bit to select between E2M1 and E1M2 micro-formats per block and decodes both into a shared E2M2 internal format. No equations, fitted parameters, predictions, or first-principles derivations appear in the abstract or visible text. The improvement claim rests on the proposed mechanism's construction and empirical evaluation rather than any self-referential reduction or self-citation load-bearing step. The work is therefore self-contained as an engineering proposal.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no explicit free parameters, axioms, or invented entities can be extracted. The design implicitly assumes that block statistics are heterogeneous enough to benefit from two formats and that the sign-bit reuse does not collide with other uses of the scale field.

pith-pipeline@v0.9.1-grok · 5718 in / 1250 out tokens · 18349 ms · 2026-06-28T20:15:11.154299+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

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

  1. Rethinking Shrinkage Bias in LLM FP4 Pretraining: Geometric Origin, Systemic Impact, and UFP4 Recipe

    cs.AI 2026-06 unverdicted novelty 6.0

    E2M1 FP4 has inherent shrinkage bias from asymmetric bin geometry that accumulates and destabilizes training; UFP4 with uniform E1M2/INT4 grids and selective RHT/stochastic rounding reduces BF16-relative degradation i...

  2. GoldenFloat: A Phi-Derived Static-Split Floating-Point Family from GF4 to GF1024 with a Lucas-Exact Integer Identity

    cs.AR 2026-06 unverdicted novelty 4.0

    GoldenFloat introduces a phi-derived rule for setting exponent and fraction widths across floating-point formats from 4 to 1024 bits, backed by open RTL generator, Lucas-exact accumulator, and FPGA implementation.

Reference graph

Works this paper leans on

20 extracted references · 19 canonical work pages · cited by 2 Pith papers · 8 internal anchors

  1. [1]

    Pretraining large language models with NVFP4.arXiv preprint arXiv:2509.25149, 2025

    Abecassis, F., Agrusa, A., Ahn, D., Alben, J., Alborghetti, S., Andersch, M., Arayandi, S., Bjorlin, A., Blakeman, A., Briones, E., et al. Pretraining large language models with nvfp4.arXiv preprint arXiv:2509.25149,

    work page arXiv

  2. [2]

    GPT-4 Technical Report

    Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774,

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    Int vs fp: A comprehensive study of fine-grained low-bit quantization formats.arXiv preprint arXiv:2510.25602, 2025a

    9 MixFP4: Enhancing NVFP4 with Adaptive FP4/INT4 Block Representations Chen, M., Wu, M., Jin, H., Yuan, Z., Liu, J., Zhang, C., Li, Y ., Huang, J., Ma, J., Xue, Z., et al. Int vs fp: A comprehensive study of fine-grained low-bit quantization formats.arXiv preprint arXiv:2510.25602, 2025a. Chen, Y ., AbouElhamayed, A. F., Dai, X., Wang, Y ., An- dronic, M....

    work page arXiv

  4. [4]

    Chen, Y ., Zou, J., and Chen, X

    IEEE, 2025b. Chen, Y ., Zou, J., and Chen, X. April: Accuracy-improved floating-point approximation for neural network accelera- tors. In2025 62nd ACM/IEEE Design Automation Con- ference (DAC), pp. 1–7, 2025c. doi: 10.1109/DAC63849. 2025.11133083. Clark, P., Cowhey, I., Etzioni, O., Khot, T., Sabharwal, A., Schoenick, C., and Tafjord, O. Think you have so...

    work page doi:10.1109/dac63849 2025

  5. [5]

    Four Over Six: More Accurate NVFP4 Quantization with Adaptive Block Scaling

    Cook, J., Guo, J., Xiao, G., Lin, Y ., and Han, S. Four over six: More accurate nvfp4 quantization with adaptive block scaling.arXiv preprint arXiv:2512.02010,

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    Bridging the gap between promise and performance for microscaling fp4 quantization,

    Egiazarian, V ., Castro, R. L., Kuznedelev, D., Panferov, A., Kurtic, E., Pandit, S., Marques, A., Kurtz, M., Ashkboos, S., Hoefler, T., et al. Bridging the gap between promise and performance for microscaling fp4 quantization.arXiv preprint arXiv:2509.23202,

    work page arXiv

  7. [7]

    GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers

    Frantar, E., Ashkboos, S., Hoefler, T., and Alistarh, D. Gptq: Accurate post-training quantization for generative pre- trained transformers.arXiv preprint arXiv:2210.17323,

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [8]

    The language model evaluation harness, 07 2024.https://zenodo.org/records/12608602

    URL https://zenodo.org/records/12608602. Grattafiori, A., Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Vaughan, A., et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783,

    work page arXiv

  9. [9]

    Measuring Massive Multitask Language Understanding

    Hendrycks, D., Burns, C., Basart, S., Zou, A., Mazeika, M., Song, D., and Steinhardt, J. Measuring mas- sive multitask language understanding.arXiv preprint arXiv:2009.03300,

    work page internal anchor Pith review Pith/arXiv arXiv 2009

  10. [10]

    R., Pawar, S

    Henry, A., Dachapally, P. R., Pawar, S. S., and Chen, Y . Query-key normalization for transformers. InFindings of the Association for Computational Linguistics: EMNLP 2020, pp. 4246–4253,

    2020

  11. [11]

    arXiv preprint arXiv:2501.13987

    Hu, W., Zhang, H., Guo, C., Feng, Y ., Guan, R., Hua, Z., Liu, Z., Guan, Y ., Guo, M., and Leng, J. M-ant: Efficient low-bit group quantization for llms via mathematically adaptive numerical type. In2025 IEEE International Symposium on High Performance Computer Architecture (HPCA), pp. 1112–1126. IEEE, 2025a. Hu, X., Cheng, Y ., Yang, D., Xu, Z., Yuan, Z....

    work page arXiv 2019

  12. [12]

    Kamath, A

    doi: 10.1109/IEEESTD.2019.8766229. Kamath, A. K., Prabhu, R., Mohan, J., Peter, S., Ramjee, R., and Panwar, A. Pod-attention: Unlocking full prefill- decode overlap for faster llm inference. InProceedings of the 30th ACM International Conference on Architectural Support for Programming Languages and Operating Sys- tems, Volume 2, pp. 897–912,

    work page doi:10.1109/ieeestd.2019.8766229 2019

  13. [13]

    URL https://doi.org/ 10.1145/3779212.3790135

    1145/3779212.3790135. URL https://doi.org/ 10.1145/3779212.3790135. Liu, A., Feng, B., Xue, B., Wang, B., Wu, B., Lu, C., Zhao, C., Deng, C., Zhang, C., Ruan, C., et al. Deepseek- v3 technical report.arXiv preprint arXiv:2412.19437, 2024a. Liu, Z., Zhao, C., Fedorov, I., Soran, B., Choudhary, D., Kr- ishnamoorthi, R., Chandra, V ., Tian, Y ., and Blankevo...

    work page doi:10.1145/3779212.3790135

  14. [14]

    Pointer Sentinel Mixture Models

    Merity, S., Xiong, C., Bradbury, J., and Socher, R. Pointer sentinel mixture models.arXiv preprint arXiv:1609.07843,

    work page internal anchor Pith review Pith/arXiv arXiv

  15. [15]

    GLU Variants Improve Transformer

    Shazeer, N. Glu variants improve transformer.arXiv preprint arXiv:2002.05202,

    work page internal anchor Pith review Pith/arXiv arXiv 2002

  16. [16]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Al- bert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, and 1 others

    Sun, Y ., Liu, R., Bai, H., Bao, H., Zhao, K., Li, Y ., Hu, J., Yu, X., Hou, L., Yuan, C., et al. Flatquant: Flatness matters for llm quantization.arXiv preprint arXiv:2410.09426,

    work page arXiv

  17. [17]

    Qwen3 Technical Report

    Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388,

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [18]

    HellaSwag: Can a Machine Really Finish Your Sentence?

    Zellers, R., Holtzman, A., Bisk, Y ., Farhadi, A., and Choi, Y . Hellaswag: Can a machine really finish your sentence? arXiv preprint arXiv:1905.07830,

    work page internal anchor Pith review Pith/arXiv arXiv 1905

  19. [19]

    ISBN 9798400715730

    Associa- tion for Computing Machinery. ISBN 9798400715730. doi: 10.1145/3725843.3756094. URL https://doi. org/10.1145/3725843.3756094. 11 MixFP4: Enhancing NVFP4 with Adaptive FP4/INT4 Block Representations A. NVINT4 vs. NVFP4 QSNR Crossover with Exact INT4 Range This appendix derives the crest-factor crossover κ at which NVINT4 and NVFP4 have equal quant...

    work page doi:10.1145/3725843.3756094

  20. [20]

    2 – – AND, FA, HA Adder (mantissa/int) – –2k(x+y+ 1)knFA, HA Exponent adder –kx– – FA, HA Exponent subtractor – – –kxXOR, FA, HA Comparator – – –kxXOR, AND, OR Aligner (barrel) – – –knlog 2 nMUX Normalizer (shared) – – –nlog 2 nMUX, OR B.4.1. COSTMODEL We use standard architecture-level approximations: GNOT = 1N AN D,(41) GAND2 = 2N AN D,(42) GOR2 = 2N AN...

    work page arXiv 2025