Pith. sign in

REVIEW 4 major objections 5 minor 119 references

This paper claims that decomposing INT8-quantized LLM weights into bit-slice matrices exposes ~70% sparsity and repeated column vectors that a co-designed accelerator can exploit, delivering 9.43x speedup and 31.1x energy efficiency over an

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review

2026-08-04 17:53 UTC pith:TWMJCLWO

load-bearing objection Bit-level repetition and prediction ideas are new, but the A100 comparison is overprovisioned and the energy metric is suspect. the 4 major comments →

arxiv 2509.10372 v1 pith:TWMJCLWO submitted 2025-09-12 cs.AR

MCBP: A Memory-Compute Efficient LLM Inference Accelerator Leveraging Bit-Slice-enabled Sparsity and Repetitiveness

Huizheng Wang , Zichuan Wang , Zhiheng Yue , Yousheng Long , Taiquan Wei , Jianxun Yang , Yang Wang , Chao Li
show 3 more authors
Shaojun Wei Yang Hu Shouyi Yin
This is my paper
classification cs.AR
keywords LLM inference acceleratorbit-slice sparsitybit repetitionGEMM computation reductionweight compressionKV cache accesstop-k attention predictionalgorithm-hardware co-design
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Dense value-level storage hides two cheap sources of efficiency in quantized LLM weights: most bits in high-order bit-slices are zero, and many bit-slice column vectors repeat. MCBP claims that decomposing INT8 weights into bit slices, grouping rows into small matrices, and treating identical bit columns as one computation can cut GEMM work, weight traffic, and KV-cache traffic simultaneously, across both prefill and decoding. To keep accuracy, the costly attention top-k prediction is replaced by bit-grained progressive filtering that stops loading Keys once the partial score is hopeless. The paper backs the claim with a 28nm accelerator design and reports 9.43x speedup and 31.1x energy efficiency over A100 on 26 benchmarks, with 35x, 5.2x, and 3.2x energy savings over prior transformer accelerators. A sympathetic reader would take the core claim as: the bit-slice level is where LLM inference redundancy actually lives, and a co-designed architecture can extract it.

Core claim

On the paper's own terms, MCBP's central discovery is that INT8-quantized LLM weight matrices, when decomposed into bit-slice matrices, exhibit about 70% average bit-sparsity and abundant repeated column vectors, while value-level sparsity is only around 7%. The paper establishes that rewriting a grouped bit-slice matrix W as E x I, where E is an enumeration matrix of unique column vectors and I maps each column to its representative, turns redundant GEMM additions into a merge step (I x X) and a reconstruction step (E x X'), cutting additions by up to 12.1x versus value sparsity and 3.8x versus naive bit-serial computing. It further claims BSTC losslessly compresses high-order bit-slices wi

What carries the argument

Grouped bit-slice matrices with enumeration-index factorization. An INT8 weight tile is decomposed into eight bit-slice matrices; m=4 rows are grouped so each column is a 4-bit vector with at most 16 possible types. The pigeonhole principle guarantees repetition because hidden dimension H (4k-12k) vastly exceeds 2^m. Multiplication G x X becomes E x (I x X): I x X merges activations that share a bit-column type in H(1-bs) additions, and E x X' reconstructs outputs in m x 2^{m-1} additions. A CAM-based match unit finds matching columns in one cycle; BSTC's two-state code (0 versus {1,data}) compresses the sparse high-order slices; BGPP's progressive threshold theta = max(A_hat) - alpha*radius

Load-bearing premise

The whole design rests on the empirical observation that bit-slice matrices of INT8-quantized LLM weights are highly sparse (about 70%) with many repeated column vectors, and that attention can be pruned bit-by-bit at a tuned threshold without losing accuracy; if quantized weights had near-random bit distributions or attention were dense, the reported speedups would shrink or vanish.

What would settle it

Compute the average bit-slice sparsity and column-repetition rate for a target model under a specific quantization scheme: BRCR's operation-count formula shows the advantage disappears as bit sparsity approaches 50%, and BSTC's compression ratio drops below 1 when sparsity falls below about 65%. Alternatively, run BGPP on a long-context task where attention is dense (e.g., retrieval across many similar tokens) and measure whether KV-load reduction falls to near zero while generation quality drops.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • BRCR and BSTC are lossless: they remove operations and bytes, so the gains transfer directly to any INT8-quantized model whose bit slices show the claimed ~70% sparsity and repetition.
  • BGPP is lossy: its pruning strength is controlled by alpha_r, and the paper shows accuracy can be held at 0% or 1% loss relative to INT8 by tuning alpha_r per benchmark.
  • Across 26 benchmarks, MCBP's standard/aggressive configurations give average 8.72x/9.43x speedup and 29.2x/31.1x energy efficiency over A100 at the same batch size; on GPU the algorithms alone give 1.03x, so the custom hardware is what realizes the gains.
  • MCBP reports 75.8% average memory-access reduction, with BSTC cutting weight traffic in decoding and BGPP cutting KV traffic, addressing the bottleneck that shifts with prompt length.
  • The scheme extends to INT4: bit sparsity remains about 4x higher than value sparsity, yielding 51% computation and 41% memory-access reduction under PTQ INT4, and MCBP is reported to beat a reproduced INT4 lookup accelerator by 1.5-1.8x while consuming 33-50% less energy.

Where Pith is reading between the lines

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

  • The empirical basis (bit sparsity ~70%, abundant repetition) is measured on five LLMs, not guaranteed by theory; quantization schemes that whiten or balance bit distributions could erase BRCR/BSTC's advantage, so the obvious stress test is measuring bit statistics on more diverse models and quantizers.
  • The group-size sweet spot m=4 balances repetition against the exponential 2^m reconstruction cost; a natural extension is hierarchical or approximate matching that keeps repetition high for larger m, which the fixed m=4 design would miss.
  • BGPP's threshold depends on per-benchmark alpha_r tuning; an accuracy-aware adaptive alpha that self-adjusts during decoding could make the technique parameter-free and extend its KV-traffic savings to longer contexts.
  • The same enumeration-index factorization could in principle apply to activations or KV cache entries, not just static weights; MCBP limits BRCR to weights, so exploiting bit repetition on dynamic values is an open direction.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes MCBP, a bit-slice-level algorithm-hardware co-design for LLM inference targeting GEMM computation, weight access, and KV-cache access. Three mechanisms are introduced: BRCR reduces GEMM work by grouping bit-slice weight vectors and reusing repeated column-vector computations via a CAM-based matcher; BSTC losslessly compresses high-order bit-slice weight matrices using a two-state coding aligned with the BRCR group granularity; BGPP performs progressive bit-level top-k prediction with early termination to reduce KV-cache traffic. The authors implement the design in RTL (TSMC 28nm, with CACTI/Ramulator modeling) and evaluate it on five LLMs and 26 benchmark settings, reporting a 9.43x inference speedup and 31.1x energy-efficiency improvement over an NVIDIA A100, and 35x/5.2x/3.2x energy savings over SpAtten/FACT/SOFA.

Significance. If the headline comparisons are made rigorous, MCBP would be a valuable contribution: the bit-slice repetition and sparsity observations are clearly demonstrated across several LLMs (Figs. 5, 8, 25), the BRCR/BSTC mechanisms are lossless, and the hardware designs (CAM match, lightweight codec, clock-gated BGPP) are concrete and backed by RTL simulation and area/power characterization. The paper also ships a thorough ablation study and a sensitivity analysis across PTQ/QAT/INT4 quantization, which are strengths. However, the contribution is not yet fully supported because the A100 comparison uses an inconsistent processor-scaling calculation, the BGPP threshold is tuned per benchmark on the test set to achieve the reported 0%/1% accuracy loss, and the cross-accelerator GOPS/W comparisons rely on heterogeneous operation definitions. These issues directly affect the claimed speedup and energy numbers, so the evaluation needs substantial revision before the central claims can be accepted.

major comments (4)
  1. [§5.3, Table 4, Fig. 20] The A100 comparison is arithmetically inconsistent as printed. The text says: 'Given the INT8 compute power of A100 is 624 TOPS, we use 148 MCBP processors (total with 622TOPS@INT8)'. Table 4 lists one MCBP processor as 54,463 GOPS ≈ 54.5 TOPS. 148 such processors would provide ≈8,060 TOPS, not 622; their total area/power would be ≈1,409 mm² and ≈354 W, versus the A100's 826 mm² and 400 W. If 'processor' instead means a PE cluster (20 per chip, Table 3), then 148 clusters correspond to 7.4 chips and provide ≈403 TOPS—again not 622. Neither reading supports the stated 622 TOPS, and the comparison is neither iso-area nor iso-power. Moreover, MCBP's GOPS likely counts bit-level additions/shifts rather than INT8 MACs, so normalizing to A100 INT8 TOPS is not established. The 8.72x/9.43x speedup and 31.1x efficiency claims need a corrected scaling and a common metric definition.
  2. [§5.2 and §6 (Fig. 24a)] The 'standard (0% loss)' configuration is obtained by per-benchmark tuning of alpha_r on the test set. Section 5.2 states: 'we adjust the value of alpha_r in 0.1 increments to evaluate the accuracy and overhead for each benchmark. This yields two MCBP configurations: standard (0% loss), aggressive (1% loss)'. Section 6 acknowledges that BGPP introduces alpha_r and that accuracy is affected. Thus the reported latency/KV-reduction numbers are not achieved by a single fixed configuration; they depend on thresholds selected to meet an accuracy target on the evaluation benchmarks. Fig. 24(a) even shows MBPP accuracy already degrading at alpha_r=0.6, so a global alpha_r in the claimed 0.5–0.6 range may not give 0% loss. Please report results for a fixed alpha_r (or a validation-set selection) and show per-benchmark accuracy relative to the INT8 baseline without test-set tuning.
  3. [Table 4, §5.4] The cross-accelerator energy-efficiency comparison is not apples-to-apples. The text states that 'the average energy efficiency is evaluated using the metric from each respective paper'. MCBP's GOPS/W is a bit-level operation count, whereas SpAtten/FACT/SOFA report value-level MAC-based GOPS/W. Different papers also use different workloads, memory systems, and technology normalizations. Comparing these absolute GOPS/W numbers (35x, 5.2x, 3.2x savings) is therefore not a reliable measure of relative efficiency. The authors should normalize all designs to a common operation definition (e.g., INT8 MAC-equivalents) and, ideally, report per-workload energy and latency rather than only aggregate GOPS/W.
  4. [§3.1 (BRCR complexity)] The printed complexity comparison for the value-based sparsity scheme is inconsistent with the stated conclusion. The text says the value-based scheme consumes H×m×k×vs additions, where vs≈0.07 is the value sparsity ratio (fraction of zeros). For H=4000, m=4, k=8, this gives 8960 additions for a k-bit GEMV, which is far less than the claimed 12.1x reduction relative to BRCR's k(H(1−bs)+m·2^{m−1}) ≈ 9856 additions. The claim requires H×m×k×(1−vs) additions. Please correct the formula and re-derive the 12.1x/3.8x ratios.
minor comments (5)
  1. [Fig. 5(g), §3.3] The label 'Theoretically optimal' in Fig. 5(g) is not defined. Clarify what bound this represents (e.g., lower bound on KV traffic given bit-level early termination).
  2. [Table 2] The '0% loss' standard configuration is not reflected exactly in the table: several MCBP(S) entries are 0.1–0.2 absolute points below the corresponding INT8 baseline (e.g., MMLU 34.6 vs 34.7, MBPP 17.1 vs 17.2). State whether 0% loss means rounding to one decimal or a tolerance, and report the actual worst-case delta.
  3. [Fig. 9 / §3.3] The bit ordering for BGPP is described inconsistently: the text says 'fetch the MSB of all Ks' while the figure caption says 'Load 1st bit (K)'. Since weights are stored in sign-magnitude format (§3.2), clarify whether bit 0 is the sign bit and which bit is fetched in round 1.
  4. [§4.1] The text says the workflow operates through 'eight key steps' with numbered markers ❶–❽, but the paragraph only enumerates ❶–❺ and ❻–❽ in the BGPP description. Make the correspondence between markers and steps explicit in the text.
  5. [§7 / abstract] The claim that MCBP is 'the first work that uses bit-level strategies for LLM inference' is too strong given that FuseKNA and Bitwave are bit-level accelerators and are adapted to GEMV in this paper. Soften the novelty claim to 'first to jointly apply bit-level repetition, sparsity, and prediction to both prefill and decoding'.

Circularity Check

1 steps flagged

BGPP's accuracy-preservation claim is fitted per benchmark via alpha_r, making the '0% loss' label true by construction; headline speedups are reported at these tuned operating points.

specific steps
  1. fitted input called prediction [Section 5.2 (Algorithm settings); used in Table 2, Fig. 20 and Abstract/Section 5.3 headline numbers.]
    "we adjust the value of αr in 0.1 increments to evaluate the accuracy and overhead for each benchmark. This yields two MCBP configurations: standard (0% loss), aggressive (1% loss), representing the minimal and maximal performance optimizations, respectively."

    The 0%/1% loss targets are the selection criteria for αr, not outputs of an independent prediction. 'Standard' and 'aggressive' are defined by the loss achieved after tuning αr on the same benchmark, so Table 2's accuracy preservation is an input constraint. Since αr directly sets the BGPP pruning threshold (Eq. 1), the per-benchmark tuned αr determines the KV reduction; the subsequent 8.72×/9.43× speedup and 29.2×/31.1× efficiency gains are measured at these fitted operating points rather than at a single pre-specified threshold, making the accuracy part of the claim circular and the performance claims contingent on in-sample tuning.

full rationale

BRCR and BSTC are self-contained: the bit-slice decomposition, pigeonhole argument, and operation-count formulas (Sec. 3.1, 3.2) derive reductions from measured bit-level sparsity/repetition without circularity. The softmax-gap property is cited to the authors' FACT [72], but it is a standard mathematical fact and not load-bearing. The main circularity is confined to BGPP evaluation: alpha_r is tuned per benchmark to hit 0%/1% loss, so the 'standard (0% loss)' accuracy is true by construction, and the headline speedups are reported at these benchmark-specific operating points. The DSE for group size m on the same workloads is a design choice rather than a prediction, and the A100 comparison contains an internal scaling inconsistency (148 MCBP processors vs Table 4 throughput) that is a validity concern but not circularity. Overall, the central algorithmic derivation is mostly independent, so the score is moderate rather than high.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claims rest on a handful of design parameters (m, alpha_r, radius, compressed bit positions, number of BGPP rounds) and on empirical assumptions about weight distribution and attention sparsity. No new physical entities are posited; the CAM and codec units are hardware components, not invented entities in the scientific sense.

free parameters (5)
  • Group size m (BRCR/BSTC) = 4
    Selected via DSE on the five evaluated LLMs to balance computation reduction and compression rate (Fig. 18); also chosen because it divides common hidden dimensions. This is a tunable design parameter, not derived from first principles.
  • BGPP threshold coefficient alpha_r = 0.5-0.6 per benchmark (standard config 0% loss; aggressive 1% loss)
    Controls the filtering threshold in Eq. (1). The paper adjusts it in 0.1 increments per benchmark to achieve target accuracy (Sec. 5.2), so the reported KV-access reduction is not parameter-free.
  • BGPP radius = 3
    Empirically set default radius in Eq. (1) (Sec. 3.3).
  • BSTC compressed bit-slice set = bits 3-7 compressed; bits 1, 2, 8 uncompressed
    Based on measured sparsity ratios greater than 65% in Fig. 8(c); a data-dependent choice of which bit planes to two-state encode.
  • BGPP number of filtering rounds = unreported
    The text says a 'predetermined number of rounds' but never states the value; the KV traffic reduction depends on it.
axioms (5)
  • standard math Pigeonhole principle: any m-row binary matrix has at most 2^m distinct column vectors, so with H >> 2^m repetitions must occur.
    Used in Sec. 3.1 to justify group-wise merging and the existence of repetitive BS vectors.
  • standard math For large H, all 2^m possible column vectors appear, bounding the reconstruction add cost by m*2^(m-1).
    Assumed in the BRCR complexity formula in Sec. 3.1 ('we can reasonably assume').
  • domain assumption Quantized LLM weights are near-Gaussian, so high-order bit-slices are mostly zero and lower-order slices are roughly uniform, giving high bit sparsity and repetition.
    Basis for BSTC and BRCR; referenced to [35,52] and empirically measured on five LLMs (Figs. 5, 8).
  • domain assumption Attention scores are sparse, and a low-bit progressive estimate with early termination can select the true top-k keys without accuracy loss.
    Basis for BGPP; validated only empirically through the tuned alpha_r experiments (Sec. 5.2, Fig. 24).
  • domain assumption The RTL/Verilator + CACTI + Ramulator + TSMC 28nm synthesis flow accurately represents the performance, power, and area of a real MCBP chip.
    All quantitative claims (speedup, GOPS/W, area) come from this simulation flow; no silicon measurement is provided.

pith-pipeline@v1.3.0-alltime-deepseek · 37993 in / 19056 out tokens · 200340 ms · 2026-08-04T17:53:11.433684+00:00 · methodology

0 comments
read the original abstract

Large language models (LLMs) face significant inference latency due to inefficiencies in GEMM operations, weight access, and KV cache access, especially in real-time scenarios. This highlights the need for a versatile compute-memory efficient accelerator. Unfortunately, existing Transformer accelerators struggle to address both aspects simultaneously, as they focus on value-level processing, missing fine-grained opportunities to optimize computation and memory collaboratively. This paper introduces MCBP, a bit-grained compute-memory efficient algorithm-hardware co-design that leverages bit-slice (BS) enabled repetitiveness and sparsity to accelerate LLM inference. MCBP features three key innovations: 1) BS-repetitiveness-enabled computation reduction (BRCR), which eliminates redundant GEMM computations via leveraging redundancy hidden among BS vectors; 2) BS-sparsity-enabled two-state coding (BSTC), which reduces weight access via exploiting significant sparsity in high-order bit-slice weight; 3) Bit-grained progressive prediction (BGPP), which reduces KV cache access by leveraging early-termination-based bit-grained prediction. These techniques, supported by custom accelerator designs, effectively alleviate the burden in GEMM, weight access, and KV cache access. Extensive experiments on 26 benchmarks show that MCBP achieves 9.43x speed up and 31.1x higher energy efficiency than Nvidia A100 GPU. Compared to SOTA Transformer accelerators, MCBP achieves 35x, 5.2x and 3.2x energy saving than Spatten, FACT and SOFA, respectively.

Figures

Figures reproduced from arXiv: 2509.10372 by Chao Li, Huizheng Wang, Jianxun Yang, Shaojun Wei, Shouyi Yin, Taiquan Wei, Yang Hu, Yang Wang, Yousheng Long, Zhiheng Yue, Zichuan Wang.

Figure 1
Figure 1. Figure 1: (a) Key bottleneck breakdown of end-to-end latency [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Top-𝑘 sparsity prediction for attention acceleration. weight and activation quantization for LLMs with up to 530B. In 2024, Atom [112] implements 8-bit KV cache quantization. Quan￾tization has become a prominent trend for deploying LLMs, sup￾ported by frameworks like TensorRT [66]. Thus, optimizing com￾pute and memory access for integer-quantized LLMs is an increas￾ingly critical topic. 2.2 Attention Spars… view at source ↗
Figure 4
Figure 4. Figure 4: Bit-level sparsity and repetition opportunities. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Challenges and our strategies for applying bit-level computing to computation-memory-efficient LLM inference. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The preparation and execution flow of MCBP. [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Bit-slice-repetitiveness-enabled computation for GEMM (BRCR). [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: Bit-grained progressive top-𝑘 prediction (BGPP). 3.2 BS-Sparsity-enabled two-state Coding (BSTC) While numerous studies [28, 29, 62, 63, 73] have explored coding techniques for sparse weight compression, they largely focus on value-level sparsity, limiting their effectiveness. In contrast, BSTC exploits the key insight that quantized weights exhibit Gaussian￾like distribution [52], thus most non-zero weigh… view at source ↗
Figure 11
Figure 11. Figure 11: Illustration for quantization process in MCBP. [PITH_FULL_IMAGE:figures/full_fig_p007_11.png] view at source ↗
Figure 10
Figure 10. Figure 10: High-level block diagram for MCBP accelerator. [PITH_FULL_IMAGE:figures/full_fig_p007_10.png] view at source ↗
Figure 13
Figure 13. Figure 13: The bit-grained computation dataflow of MCBP. [PITH_FULL_IMAGE:figures/full_fig_p008_13.png] view at source ↗
Figure 16
Figure 16. Figure 16: shows the architecture of BGPP unit. First, 16 bit-serial inner product units compute Q (1*64) × K (64*16), each with a psum max < min > Clipping module Vital K indices A W63(bit) > > > ... > 1 0 1 0 0 0 1 1 0 1 0 ...1 0 1 1 A[0] A[1] A[3] Vital K indices: [0,2,6,…60,62,63] A[63] Radius (3) Clock-gated Progressive Filter Threshold Updating com Bit-serial based IP Unit + <<1 Neg X0 MSB? 0 1 ... x64 X63 W0(… view at source ↗
Figure 17
Figure 17. Figure 17: Normalized computation (prefill stage) and memory access (decoding stage) of LLM inference. 0 1 2 3 4 1 2 3 4 5 6 7 8 9 0 5 10 15 20 25 30 Compression Rate (CR) Group size (m) Comp Reduction (CPR) 0 1 2 3 4 1 2 3 4 5 6 7 8 9 0 5 10 15 20 25 30 Compression Rate (CR) Group size (m) Comp Reduction (CPR) Min comp reduce Max comp reduce Compression rate Optimal Balance Comp Reduction (CPR) Compression Rate (CR… view at source ↗
Figure 18
Figure 18. Figure 18: Design space exploration of the optimal group size [PITH_FULL_IMAGE:figures/full_fig_p010_18.png] view at source ↗
Figure 20
Figure 20. Figure 20: (a) Throughput and (b) energy efficiency gain of [PITH_FULL_IMAGE:figures/full_fig_p011_20.png] view at source ↗
Figure 22
Figure 22. Figure 22: Area/Power of MCBP at TSMC 28nm, 1GHz. breakdown. Software-hardware co-design BRCR, TSBC and BGPP bring 4.24×, 2.98× and 2.44× efficiency gain, respectively. 5.4 Comparison with SOTA Accelerators [PITH_FULL_IMAGE:figures/full_fig_p012_22.png] view at source ↗
Figure 24
Figure 24. Figure 24: (a) Evaluation of MCBP’s optimization impact [PITH_FULL_IMAGE:figures/full_fig_p013_24.png] view at source ↗
Figure 26
Figure 26. Figure 26: Compared with SOTA INT4 accelerator. repetition, while in the decoding stage, it minimizes memory access through bit sparsity and bit-grained early termination. Value sparsity accelerator. Numerous accelerators [16, 21, 22, 24, 28–30, 39, 45, 49, 55, 56, 60, 62, 63, 71, 82, 86, 89, 100, 102, 107, 108, 110] exploit value sparsity to improve NNs performance. EIE [28] utilizes dynamic input and static weight… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

119 extracted references · 19 linked inside Pith

  1. [1]

    Miklós Ajtai. 1994. The complexity of the pigeonhole principle.Combinatorica 14 (1994), 417–433

    1994

  2. [2]

    Jorge Albericio, Alberto Delmás, Patrick Judd, Sayeh Sharify, Gerard O’Leary, Roman Genov, and Andreas Moshovos. 2017. Bit-pragmatic deep neural network computing. InProceedings of the 50th annual IEEE/ACM international symposium on microarchitecture. 382–394

    2017

  3. [3]

    Renzo Andri, Lukas Cavigelli, Davide Rossi, and Luca Benini. 2016. YodaNN: An ultra-low power convolutional neural network accelerator based on binary weights. InProceedings of the IEEE Computer Society Annual Symposium on VLSI (ISVLSI). 236–241

    2016

  4. [4]

    Rohan Anil, Andrew M Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. 2023. Palm 2 technical report.arXiv preprint arXiv:2305.10403(2023)

    Pith/arXiv arXiv 2023

  5. [5]

    Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, et al. 2021. Program synthesis with large language models.arXiv preprint arXiv:2108.07732(2021)

    Pith/arXiv arXiv 2021

  6. [6]

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng X...

    Pith/arXiv arXiv 2023

  7. [7]

    Zhenyu Bai, Pranav Dangi, Huize Li, and Tulika Mitra. 2024. SWAT: Scalable and efficient window attention-based transformers acceleration on FPGAs. In Proceedings of the 61st ACM/IEEE Design Automation Conference. 1–6

    2024

  8. [8]

    Lukas Cavigelli and Luca Benini. 2016. Origami: A 803-GOp/s/W convolu- tional network accelerator.IEEE Transactions on Circuits and Systems for Video Technology27, 11 (2016), 2461–2475

    2016

  9. [9]

    Yu, Qiang Yang, and Xing Xie

    Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, Wei Ye, Yue Zhang, Yi Chang, Philip S. Yu, Qiang Yang, and Xing Xie. 2024. A survey on evaluation of large language models.ACM Transactions on Intelligent Systems and Technology 15, 3 (2024), 1–45

    2024

  10. [10]

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde De Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian...

    Pith/arXiv arXiv 2021

  11. [11]

    Yi Chen, Yongwei Zhao, Yifan Hao, Yuanbo Wen, Yuntao Dai, Xiaqing Li, Yang Liu, Rui Zhang, Mo Zou, Xinkai Song, Xing Hu, Zidong Du, Huaping Chen, Qi Guo, and Tianqi Chen. 2024. Cambricon-C: Efficient 4-Bit Matrix Unit via Primi- tivization. In2024 57th IEEE/ACM International Symposium on Microarchitecture (MICRO). IEEE, 538–550

    2024

  12. [12]

    Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al

  13. [13]

    Mike Conover, Matt Hayes, Ankit Mathur, Jianwei Xie, Jun Wan, Sam Shah, Ali Ghodsi, Patrick Wendell, Matei Zaharia, and Reynold Xin. 2023. Free Dolly: Introducing the world’s first truly open instruction-tuned LLM.Company Blog of Databricks(2023)

    2023

  14. [14]

    Alberto Delmas, Patrick Judd, Sayeh Sharify, and Andreas Moshovos. 2017. Dynamic stripes: Exploiting the dynamic precision requirements of activation values in neural networks.arXiv preprint arXiv:1706.00504(2017)

    Pith/arXiv arXiv 2017

  15. [15]

    Alberto Delmas Lascorz, Patrick Judd, Dylan Malone Stuart, Zissis Poulos, Mostafa Mahmoud, Sayeh Sharify, Milos Nikolic, Kevin Siu, and Andreas Moshovos. 2019. Bit-tactical: A software/hardware approach to exploiting value and bit sparsity in neural networks. InProceedings of the Twenty-Fourth International Conference on Architectural Support for Programm...

    2019

  16. [16]

    Chunhua Deng, Yang Sui, Siyu Liao, Xuehai Qian, and Bo Yuan. 2021. GoSPA: An energy-efficient high-performance globally optimized sparse convolutional neural network accelerator. InProceedings of the ACM/IEEE 48th Annual Inter- national Symposium on Computer Architecture (ISCA). 1110–1123

    2021

  17. [17]

    Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. 2022. GPT3.Int8 (): 8-bit matrix multiplication for Transformers at scale.Advances in Neural Information Processing Systems35 (2022), 30318–30332. 14 MCBP: A Memory-Compute Efficient LLM Inference Accelerator Leveraging Bit-Slice-enabled Sparsity and Repetitiveness MICRO 2025, October 18–22, ...

    2022

  18. [18]

    Claire Cardie Faisal Ladhak, Esin Durmus and Kathleen McKeown. 2020. Wik- iLingua: A new benchmark dataset for multilingual abstractive summarization. InFindings of EMNLP, 2020

    2020

  19. [19]

    Hongxiang Fan, Thomas Chau, Stylianos I Venieris, Royson Lee, Alexandros Kouris, Wayne Luk, Nicholas D Lane, and Mohamed S Abdelfattah. 2022. Adapt- able butterfly accelerator for attention-based NNs via hardware and algorithm co-design. InProceedings of the 55th IEEE/ACM International Symposium on Microarchitecture (MICRO). 599–615

    2022

  20. [20]

    Zichen Fan, Qirui Zhang, Pierre Abillama, Sara Shoouri, Changwoo Lee, David Blaauw, Hun-Seok Kim, and Dennis Sylvester. 2023. Taskfusion: An efficient transfer learning architecture with dual delta sparsity for multi-task natural language processing. InProceedings of the 50th Annual International Symposium on Computer Architecture. 1–14

    2023

  21. [21]

    Chao Fang, Shouliang Guo, Wei Wu, Jun Lin, Zhongfeng Wang, Ming Kai Hsu, and Lingzhi Liu. 2022. An efficient hardware accelerator for sparse Transformer neural networks. In2022 IEEE International Symposium on Circuits and Systems (ISCAS). IEEE, 2670–2674

    2022

  22. [22]

    Chao Fang, Aojun Zhou, and Zhongfeng Wang. 2022. An algorithm-hardware co-optimized framework for accelerating N:M sparse Transformers.IEEE Trans- actions on Very Large Scale Integration (VLSI) Systems30, 11 (2022), 1573–1586

    2022

  23. [23]

    Ashish Gondimalla, Noah Chesnut, Mithuna Thottethodi, and TN Vijaykumar

  24. [24]

    Sumanth Gudaparthi, Sarabjeet Singh, Surya Narayanan, Rajeev Balasubra- monian, and Visvesh Sathe. 2022. CANDLES: Channel-aware novel dataflow- microarchitecture co-design for low energy sparse neural network acceleration. InProceedings of the IEEE International Symposium on high-performance computer architecture (HPCA). 876–891

    2022

  25. [25]

    Tae Jun Ham, Sung Jun Jung, Seonghak Kim, Young H Oh, Yeonhong Park, Yoonho Song, Jung-Hun Park, Sanghee Lee, Kyoung Park, Jae W Lee, and Deog- Kyoon Jeong. 2020. A3: Accelerating attention mechanisms in neural networks with approximation. InProceedings of the IEEE International Symposium on High Performance Computer Architecture (HPCA). 328–341

    2020

  26. [26]

    Tae Jun Ham, Yejin Lee, Seong Hoon Seo, Soosung Kim, Hyunji Choi, Sung Jun Jung, and Jae W Lee. 2021. ELSA: Hardware-software co-design for efficient, lightweight self-attention mechanism in neural networks. InPrceedings of the 48th ACM/IEEE Annual International Symposium on Computer Architecture (ISCA). 692–705

    2021

  27. [27]

    Meng Han, Liang Wang, Limin Xiao, Hao Zhang, Tianhao Cai, Jiale Xu, Yibo Wu, Chenhao Zhang, and Xiangrong Xu. 2024. BitNN: A bit-serial accelerator for k-nearest neighbor search in point clouds. InProceddings of the ACM/IEEE 51st Annual International Symposium on Computer Architecture (ISCA). 1278–1292

    2024

  28. [28]

    Song Han, Xingyu Liu, Huizi Mao, Jing Pu, Ardavan Pedram, Mark A Horowitz, and William J Dally. 2016. EIE: Efficient inference engine on compressed deep neural network.ACM SIGARCH Computer Architecture News44, 3 (2016), 243– 254

    2016

  29. [29]

    Song Han, Huizi Mao, and William J Dally. 2015. Deep compression: Com- pressing deep neural networks with pruning, trained quantization and huffman coding.arXiv preprint arXiv:1510.00149(2015)

    Pith/arXiv arXiv 2015

  30. [30]

    Edward Hanson, Shiyu Li, Hai‘Helen’ Li, and Yiran Chen. 2022. Cascading structured pruning: Enabling high data reuse for sparse DNN accelerators. In Proceedings of the 49th Annual International Symposium on Computer Architecture. 522–535

    2022

  31. [31]

    Kartik Hegde, Jiyong Yu, Rohit Agrawal, Mengjia Yan, Michael Pellauer, and Christopher Fletcher. 2018. UCNN: Exploiting computational reuse in deep neu- ral networks via weight repetition. In2018 ACM/IEEE 45th Annual International Symposium on Computer Architecture (ISCA). IEEE, 674–687

    2018

  32. [32]

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2020. Measuring massive multitask language un- derstanding.arXiv preprint arXiv:2009.03300(2020)

    Pith/arXiv arXiv 2020

  33. [33]

    Seongmin Hong, Seungjae Moon, Junsoo Kim, Sungjae Lee, Minsub Kim, Dong- soo Lee, and Joo-Young Kim. 2022. DFX: A low-latency multi-FPGA appliance for accelerating Transformer-based text generation. InProceedings of the 55th IEEE/ACM International Symposium on Microarchitecture (MICRO). 616–630

    2022

  34. [34]

    Yuxuan Hu, Xiaodong Chen, Cuiping Li, Hong Chen, and Jing Zhang. 2025. QUAD: Quantization and Parameter-Efficient Tuning of LLM with Activation Decomposition.arXiv preprint arXiv:2503.19353(2025)

    Pith/arXiv arXiv 2025

  35. [35]

    Dongseok Im, Gwangtae Park, Zhiyong Li, Junha Ryu, and Hoi-Jun Yoo. 2023. Sibia: Signed bit-slice architecture for dense dnn acceleration with slice-level sparsity exploitation. In2023 IEEE International Symposium on High-Performance Computer Architecture (HPCA). IEEE, 69–80

    2023

  36. [36]

    Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, An- drew Howard, Hartwig Adam, and Dmitry Kalenichenko. 2018. Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Infer- ence. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2704–2713

    2018

  37. [37]

    Patrick Judd, Jorge Albericio, Tayler Hetherington, Tor M Aamodt, and Andreas Moshovos. 2016. Stripes: Bit-serial deep neural network computing. InProceed- ings of the 49th Annual IEEE/ACM International Symposium on Microarchitecture (MICRO). 1–12

    2016

  38. [38]

    Dongyun Kam, Myeongji Yun, Sunwoo Yoo, Seungwoo Hong, Zhengya Zhang, and Youngjoo Lee. 2024. Panacea: Novel DNN Accelerator using Accuracy- Preserving Asymmetric Quantization and Energy-Saving Bit-Slice Sparsity. arXiv preprint arXiv:2412.10059(2024)

    Pith/arXiv arXiv 2024

  39. [39]

    Sanghoon Kang, Donghyeon Han, Juhyoung Lee, Dongseok Im, Sangyeob Kim, Soyeon Kim, Junha Ryu, and Hoi-Jun Yoo. 2021. GANPU: An energy-efficient multi-DNN training processor for GANs with speculative dual-sparsity exploita- tion.IEEE Journal of Solid-State Circuits56, 9 (2021), 2845–2857

    2021

  40. [40]

    Majeed Kazemitabaar, Runlong Ye, Xiaoning Wang, Austin Zachary Henley, Paul Denny, Michelle Craig, and Tovi Grossman. 2024. Codeaid: Evaluating a classroom deployment of an LLM-based programming assistant that balances student and educator needs. InProceedings of the CHI Conference on Human Factors in Computing Systems. 1–20

    2024

  41. [41]

    Yoongu Kim, Weikun Yang, and Onur Mutlu. 2015. Ramulator: A fast and extensible DRAM simulator.IEEE Computer architecture letters15, 1 (2015), 45–49

    2015

  42. [42]

    Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilić, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, Matthias Gallé, et al. 2022. Bloom: A 176B-parameter open-access multilingual language model.arXiv preprint arXiv:2211.05100(2022)

    Pith/arXiv arXiv 2022

  43. [43]

    Jinmook Lee, Changhyeon Kim, Sanghoon Kang, Dongjoo Shin, Sangyeob Kim, and Hoi-Jun Yoo. 2018. UNPU: A 50.6 TOPS/W unified deep neural network accelerator with 1b-to-16b fully-variable weight bit-precision. InProceedings of IEEE International Solid-State Circuits Conference-(ISSCC). 218–220

    2018

  44. [44]

    Brian Leibowitz, Robert Palmer, John Poulton, Yohan Frans, Simon Li, John Wilson, Michael Bucher, Andrew M Fuller, John Eyles, Marko Aleksic, Trey Greer, and Nhat M Nguyen. 2010. A 4.3 GB/s mobile memory interface with power-efficient bandwidth scaling.IEEE Journal of Solid-State Circuits45, 4 (2010), 889–898

    2010

  45. [45]

    Jonathan S Lew, Yunpeng Liu, Wenyi Gong, Negar Goli, R David Evans, and Tor M Aamodt. 2022. Anticipating and eliminating redundant computations in accelerated sparse training. InProceedings of the 49th Annual International Symposium on Computer Architecture. 536–551

    2022

  46. [46]

    Bingbing Li, Santosh Pandey, Haowen Fang, Yanjun Lyv, Ji Li, Jieyang Chen, Mimi Xie, Lipeng Wan, Hang Liu, and Caiwen Ding. 2020. FTRANS: Energy- efficient acceleration of Transformers using FPGA. InProceedings of the ACM/IEEE International Symposium on Low Power Electronics and Design. 175– 180

    2020

  47. [47]

    Gang Li, Weixiang Xu, Zhuoran Song, Naifeng Jing, Jian Cheng, and Xiaoyao Liang. 2022. Ristretto: An atomized processing architecture for sparsity- condensed stream flow in CNN. InProceedings of the 55th IEEE/ACM International Symposium on Microarchitecture (MICRO). 1434–1450

    2022

  48. [48]

    Guoyu Li, Shengyu Ye, Chunyun Chen, Yang Wang, Fan Yang, Ting Cao, Cheng Liu, Mohamed M Sabry, and Mao Yang. 2025. LUT-DLA: Lookup Ta- ble as Efficient Extreme Low-Bit Deep Learning Accelerator.arXiv preprint arXiv:2501.10658(2025)

    Pith/arXiv arXiv 2025

  49. [49]

    Shiyu Li, Edward Hanson, Xuehai Qian, Hai" Helen" Li, and Yiran Chen. 2021. ESCALATE: Boosting the efficiency of sparse CNN accelerator with kernel decomposition. InProceedings of the 54th Annual IEEE/ACM International Sym- posium on Microarchitecture. 992–1004

    2021

  50. [50]

    Zheng Li, Soroush Ghodrati, Amir Yazdanbakhsh, Hadi Esmaeilzadeh, and Mingu Kang. 2022. Accelerating attention through gradient-based learned runtime pruning. InProceedings of the 49th Annual International Symposium on Computer Architecture. 902–915

    2022

  51. [51]

    Bin Lin, Chen Zhang, Tao Peng, Hanyu Zhao, Wencong Xiao, Minmin Sun, Anmin Liu, Zhipeng Zhang, Lanbo Li, Xiafei Qiu, Li Shen, Zhigang Ji, Tao Xie, Yong Li, and Wei Lin. 2024. Infinite-LLM: Efficient LLM service for long context with distattention and distributed kvcache.arXiv preprint arXiv:2401.02669 (2024)

    Pith/arXiv arXiv 2024

  52. [52]

    Fangxin Liu, Ning Yang, Haomin Li, Zongwu Wang, Zhuoran Song, Songwen Pei, and Li Jiang. 2024. SPARK: Scalable and precision-aware acceleration of neural networks via efficient encoding. InProceedings of the IEEE International Symposium on High-Performance Computer Architecture (HPCA). 1029–1042

    2024

  53. [53]

    Jing Liu, Ruihao Gong, Xiuying Wei, Zhiwei Dong, Jianfei Cai, and Bohan Zhuang. 2023. QLLM: Accurate and efficient low-bitwidth quantization for large language models.arXiv preprint arXiv:2310.08041(2023)

    Pith/arXiv arXiv 2023

  54. [54]

    Siqin Liu, Prakash Chand Kuve, and Avinash Karanth. 2024. HSCONN: Hardware-Software Co-Optimization of Self-Attention Neural Networks for Large Language Models. InProceedings of the Great Lakes Symposium on VLSI

    2024

  55. [55]

    Shiwei Liu, Peizhe Li, Jinshan Zhang, Yunzhengmao Wang, Haozhe Zhu, Wen- ning Jiang, Shan Tang, Chixiao Chen, Qi Liu, and Ming Liu. 2023. 16.2 A 28nm 53.8 TOPS/W 8b sparse Transformer accelerator with in-memory butterfly zero skipper for unstructured-pruned NN and CIM-based local-attention-reusable 15 MICRO 2025, October 18–22, 2025, Seoul, Korea Huizheng...

    2023

  56. [56]

    Zhi-Gang Liu, Paul N Whatmough, Yuhao Zhu, and Matthew Mattina. 2022. S2TA: Exploiting structured sparsity for energy-efficient mobile CNN accelera- tion. InProceedings of the IEEE International Symposium on High-Performance Computer Architecture (HPCA). 573–586

    2022

  57. [57]

    Yun-Chen Lo and Ren-Shuo Liu. 2023. Bit-serial cache: Exploiting input bit vector repetition to accelerate bit-serial inference. InProceedings of the 60th ACM/IEEE Design Automation Conference (DAC). 1–6

    2023

  58. [58]

    Hang Lu, Liang Chang, Chenglong Li, Zixuan Zhu, Shengjian Lu, Yanhuan Liu, and Mingzhe Zhang. 2021. Distilling bit-level sparsity parallelism for general purpose deep learning acceleration. InMICRO-54: 54th Annual IEEE/ACM International Symposium on Microarchitecture. 963–976

    2021

  59. [59]

    Liqiang Lu, Yicheng Jin, Hangrui Bi, Zizhang Luo, Peng Li, Tao Wang, and Yun Liang. 2021. Sanger: A co-design framework for enabling sparse attention using reconfigurable architecture. InProceedings of the 54th Annual IEEE/ACM International Symposium on Microarchitecture. 977–991

    2021

  60. [60]

    Mostafa Mahmoud, Isak Edo, Ali Hadi Zadeh, Omar Mohamed Awad, Gennady Pekhimenko, Jorge Albericio, and Andreas Moshovos. 2020. TensorDash: Ex- ploiting sparsity to accelerate deep neural network training. InProceedings of the 53rd Annual IEEE/ACM International Symposium on Microarchitecture (MICRO). 781–795

    2020

  61. [61]

    Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. 2016. Pointer sentinel mixture models. InProceedings of the International Conference on Learning Representations

    2016

  62. [62]

    Bert Moons, Roel Uytterhoeven, Wim Dehaene, and Marian Verhelst. 2017. 14.5 Envision: A 0.26-to-10TOPS/W subword-parallel dynamic-voltage-accuracy- frequency-scalable convolutional neural network processor in 28nm FDSOI. InProceddings of the IEEE International Solid-State Circuits Conference (ISSCC). 246–247

    2017

  63. [63]

    Bert Moons and Marian Verhelst. 2016. An energy-efficient precision-scalable ConvNet processor in 40-nm CMOS.IEEE Journal of solid-state Circuits52, 4 (2016), 903–914

    2016

  64. [64]

    Naveen Muralimanohar, Rajeev Balasubramonian, and Norman P Jouppi. 2009. CACTI 6.0: A tool to model large caches.HP laboratories27 (2009), 28

    2009

  65. [65]

    Daye Nam, Andrew Macvean, Vincent Hellendoorn, Bogdan Vasilescu, and Brad Myers. 2024. Using an LLM to help with code understanding. InProceedings of the IEEE/ACM 46th International Conference on Software Engineering. 1–13

    2024

  66. [66]

    Nvidia. 2023. TensorRT-LLM. https://github.com/NVIDIA/TensorRT-LLM? tab=readme-ov-file

    2023

  67. [67]

    Mike O’Connor, Niladrish Chatterjee, Donghyuk Lee, John Wilson, Aditya Agrawal, Stephen W Keckler, and William J Dally. 2017. Fine-grained DRAM: Energy-efficient DRAM for extreme bandwidth systems. InProceedings of the 50th Annual IEEE/ACM International Symposium on Microarchitecture. 41–54

    2017

  68. [68]

    Angshuman Parashar, Minsoo Rhu, Anurag Mukkara, Antonio Puglielli, Rang- harajan Venkatesan, Brucek Khailany, Joel Emer, Stephen W Keckler, and William J Dally. 2017. SCNN: An accelerator for compressed-sparse convo- lutional neural networks.ACM SIGARCH computer architecture news45, 2 (2017), 27–40

    2017

  69. [69]

    Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer

  70. [70]

    Pratyush Patel, Esha Choukse, Chaojie Zhang, Aashaka Shah, Íñigo Goiri, Saeed Maleki, and Ricardo Bianchini. 2024. Splitwise: Efficient generative LLM infer- ence using phase splitting. In2024 ACM/IEEE 51st Annual International Sympo- sium on Computer Architecture (ISCA). IEEE, 118–132

    2024

  71. [71]

    Eric Qin, Ananda Samajdar, Hyoukjun Kwon, Vineet Nadella, Sudarshan Srini- vasan, Dipankar Das, Bharat Kaul, and Tushar Krishna. 2020. Sigma: A sparse and irregular GEMM accelerator with flexible interconnects for DNN training. In Proceedings of the IEEE International Symposium on High Performance Computer Architecture (HPCA). 58–70

    2020

  72. [72]

    Yubin Qin, Yang Wang, Dazheng Deng, Zhiren Zhao, Xiaolong Yang, Leibo Liu, Shaojun Wei, Yang Hu, and Shouyi Yin. 2023. Fact: FFN-attention co-optimized Transformer architecture with eager correlation prediction. InProceedings of the 50th Annual International Symposium on Computer Architecture. 1–14

    2023

  73. [73]

    Yubin Qin, Yang Wang, Jiachen Wang, Zhiwei Lin, Yushu Zhao, Shaojun Wei, Yang Hu, and Shouyi Yin. 2025. 23.8 An 88.36 TOPS/W Bit-Level-Weight- Compressed Large-Language-Model Accelerator with Cluster-Aligned INT-FP- GEMM and Bi-Dimensional Workflow Reformulation. In2025 IEEE International Solid-State Circuits Conference (ISSCC), Vol. 68. IEEE, 420–422

    2025

  74. [74]

    Yubin Qin, Yang Wang, Zhiren Zhao, Xiaolong Yang, Yang Zhou, Shaojun Wei, Yang Hu, and Shouyi Yin. 2024. MECLA: Memory-compute-efficient LLM accelerator with scaling sub-matrix partition. InProceedings of the 51st ACM/IEEE Annual International Symposium on Computer Architecture (ISCA). 1032–1047

    2024

  75. [75]

    Zheng Qu, Liu Liu, Fengbin Tu, Zhaodong Chen, Yufei Ding, and Yuan Xie. 2022. DOTA: Detect and omit weak attentions for scalable Transformer acceleration. InProceedings of the 27th ACM International Conference on Architectural Support for Programming Languages and Operating Systems. 14–26

    2022

  76. [76]

    Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiao- qing Ellen Tan, Yossi Adi, Jingyu Liu, Romain Sauvestre, Tal Remez, et al. 2023. Code LLaMa: Open foundation models for code.arXiv preprint arXiv:2308.12950 (2023)

    Pith/arXiv arXiv 2023

  77. [77]

    Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2021. Winogrande: An adversarial Winograd schema challenge at scale.Commun. ACM64, 9 (2021), 99–106

    2021

  78. [78]

    Sayeh Sharify, Alberto Delmas Lascorz, Mostafa Mahmoud, Milos Nikolic, Kevin Siu, Dylan Malone Stuart, Zissis Poulos, and Andreas Moshovos. 2019. Laconic deep learning inference acceleration. InProceedings of the 46th International Symposium on Computer Architecture. 304–317

    2019

  79. [79]

    Hardik Sharma, Jongse Park, Naveen Suda, Liangzhen Lai, Benson Chau, Joon Kyung Kim, Vikas Chandra, and Hadi Esmaeilzadeh. 2018. Bit Fusion: Bit-level dynamically composable architecture for accelerating deep neural net- work. InProceedings of the ACM/IEEE 45th Annual International Symposium on Computer Architecture (ISCA). 764–775

    2018

  80. [80]

    Guan Shen, Jieru Zhao, Quan Chen, Jingwen Leng, Chao Li, and Minyi Guo

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.