Pith. sign in

REVIEW 4 major objections 4 minor 28 references

BRIM: Workload-Balanced Dual-Sided Bit-Serial Sparse Inference Accelerator

T0 review · 4 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read BRIM shows that the workload imbalance capping dual-sided bit-serial accelerators is solvable, reaching over 90% PE utilization and up to 2.37× speedup under iso-area constraints.

desk verdict Solid co-design for a real bottleneck, but accuracy claims are overstated and calibration robustness is untested. read the letter →

arxiv 2607.19431 v1 pith:6G34OA2V submitted 2026-07-20 cs.AR cs.AI

Varun Manjunath , Ruokai Yin , Donghyun Lee , Arkapravo Ghosh , Priyadarshini Panda This is my paper · ORCID
classification cs.ARcs.AI
keywords dual-sidedbit-serialaccelerationbit-levelsparsityworkloadimbalancePEutilizationpost-trainingweightpruningBoothencodinghardware-softwareco-designDNNinferenceaccelerator
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Bit-serial accelerators that skip zero bits on both weights and activations can cut partial products multiplicatively, but in practice they stall: lanes that finish early wait on a slow partner, and utilization sits at 56–64%. BRIM attacks that bottleneck directly. Its offline pruning step, CBP, reshapes weights so that, on average, every lane in a lockstep group faces the same total workload, using calibrated activation statistics to set each weight's target bit-count inversely proportional to the activation's expected bit-count. Its hardware step, pairwise slot donation, lets an idle lane finish its neighbor's remaining partial products. On CNNs, ViTs, and LLMs under iso-area constraints, BRIM reports over 90% PE utilization, up to 2.37× speedup, and up to 1.63× energy-efficiency improvement over prior dual-sided designs, with accuracy mostly held in check by second-order error compensation.

What carries the argument

The product structure of per-lane cost η(w)·η(a), where η is the Booth non-zero term count, is the central object. CBP uses the calibration profile τ to set targets t* = floor(C*/τ) so that t*·τ ≈ constant per group, then realizes targets via a Booth-term-count lookup table with Hessian-based error compensation. Slot donation is a pairwise 2:1 mux that lets one lane execute its partner's next term pair, absorbing residual variance with minimal hardware.

What would settle it

Run the reported W4A8 experiments and compare against the abstract's 'at most 1.2%' accuracy bound: Table 2 already shows ViT-L dropping 1.28 points and OPT-2.7B perplexity rising 1.85 points, which refutes the blanket bound. Separately, evaluate BRIM on a held-out input distribution whose activation term-count statistics differ from the 128-sample calibration set; if PE utilization falls materially below 90%, the calibration-sensitivity premise is confirmed.

Watch

Extended reading notes

Core claim

BRIM's central claim is that the workload-imbalance bottleneck in dual-sided bit-serial sparse inference is not a fixed cost of exploiting both operands' sparsity but a solvable scheduling problem. Because each lane's cycle cost is the product η(w)·η(a) of the two operands' non-zero Booth-term counts, lanes paired in lockstep finish out of sync. BRIM equalizes those products rather than tolerating their spread: CBP replaces each weight offline with the closest value whose Booth-term count matches a target chosen inversely proportional to the profiled activation term count at that position, and a 1.2%-area slot-donation front-end lets a finished lane absorb its neighbor's next pending term pa

Load-bearing premise

CBP sets its balancing targets from activation term-count statistics profiled on 128 calibration samples (Sections 3.1.1, 4.2); if inference-time activation sparsity drifts from that profile, the inverse-proportional targets no longer equalize per-lane costs and the >90% utilization figure cannot be expected to hold.

Editorial extensions

If this is right

  • If BRIM's numbers hold, dual-sided bit-serial accelerators no longer need to sacrifice one operand's sparsity to avoid idle lanes; the 36–44% utilization gap is largely recoverable.
  • CBP is training-free and post-quantization, so it can be dropped into an existing quantized model without retraining or fine-tuning (about 3 hours on a single GPU for a large language model).
  • The 1.2% per-PE area overhead of slot donation means the balancing hardware does not eat the iso-area PE-count advantage that makes bit-serial arrays competitive.
  • The design-space sweep identifies G=16 as a Pareto knee for group size, beyond which CBP must modify over 80% of weights and perplexity degrades sharply.
  • Because the cost is multiplicative, the same architecture stands to benefit if either operand becomes sparser, pointing toward joint mixed-precision allocation (a direction the paper itself flags).

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The balancing targets are only as good as the 128-sample calibration profile; a deployment distribution that shifts from that profile would degrade the >90% utilization claim, since CBP's targets are computed once and weights are fixed.
  • The abstract's 'at most 1.2% accuracy degradation' is narrower than Table 2 shows for W4A8: ViT-L drops 1.28 points and OPT-2.7B's perplexity rises 1.85 points, so the blanket bound should be read as configuration-dependent.
  • Pairwise donation works because CBP shrinks residual imbalance to adjacent-lane variance; a design that used donation without CBP would still be limited, suggesting the two mechanisms are complements rather than substitutes.
  • The same inverse-proportional balancing idea could be applied to other sparse encodings (e.g., higher-radix Booth or canonical signed digit) since only the product-of-term-counts cost model matters.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes BRIM, a hardware–software co-design for dual-sided bit-serial sparse DNN accelerators. BRIM combines Cyclic-Balanced Pruning (CBP), an offline post-training weight reshaping method that uses profiled activation term counts to set per-weight Booth-term targets inversely proportional to expected activation activity, with a Pairwise Slot Donation hardware mechanism that lets an idle lane execute the next pending term pair of its adjacent partner. The authors report that this co-design raises PE utilization above 90%, achieves up to 2.37× speedup and 1.63× energy-efficiency improvement over prior dual-sided designs under iso-area constraints, and claim at most 1.2% accuracy degradation / 1.1 perplexity-point increase. The evaluation spans CNNs, ViTs, and LLMs using RTL synthesis and a cycle-accurate simulator.

Significance. If the claims hold, BRIM addresses a real and previously underappreciated bottleneck: in dual-sided bit-serial execution, per-lane cost is the product of two independent sparsity counts, so even with bit-level skipping, lockstep groups waste cycles waiting for the slowest lane. The CBP idea of reshaping weights offline based on activation statistics is a sensible and practically attractive approach, and the pairwise slot-donation mechanism is a plausible low-overhead complement. The paper includes RTL synthesis, cycle-accurate simulation, iso-area comparisons, ablations, and a design-space exploration, which are appropriate and useful. However, the manuscript contains several load-bearing inconsistencies and a potentially serious hardware-datapath gap that must be resolved before the quantitative claims can be accepted.

major comments (4)
  1. [§3.2.2, Figure 6] The slot-donation datapath as described cannot accumulate donated partial products into the partner’s output. Stage 1 selects L(i+1)_nxt via the 2:1 mux, but the selected pair is then processed in Stage 2 of the donating lane itself and accumulated into that lane’s local Psum Register. Since each lane computes a different weight–activation pair in an output-stationary dataflow, the donated partial product must be added to the partner lane’s psum, not the donating lane’s. No cross-lane psum routing is shown in Figure 6 or described in the text. If the cycle-accurate simulator models donation as functionally correct without such routing, the utilization and speedup results are invalid. Please clarify the accumulator organization or correct the datapath; if cross-lane accumulation is required, reassess the claimed 1.2% area overhead and single-mux critical-path delay.
  2. [Abstract; §5.1, Table 2] The abstract and Section 5.1 state that CBP incurs 'at most 1.2% accuracy degradation' and 'at most 1.1 perplexity points,' but Table 2 contradicts this. ViT-S W4A8 drops from 80.56 to 79.2 (−1.36%), ViT-L W4A8 drops from 84.92 to 83.64 (−1.28%), and OPT-2.7B W4A8 perplexity rises from 20.45 to 22.3 (+1.85). These exceed both stated bounds. The accuracy claims must be corrected and the abstract and text aligned with the actual reported data.
  3. [§3.1.1, §4.2] CBP’s balancing targets rely on an activation term-count profile τ estimated from 128 calibration samples (Section 4.2). The inverse-proportional assignment t*_k = round(C*/τ) is only optimal when the calibration profile matches inference-time activation statistics. The paper provides no sensitivity analysis: no variation of calibration-set size, no domain-shift experiment, and no comparison between calibration-profile τ and actual test-time τ. Since activations are data-dependent and weights are fixed at deployment, systematic miscalibration would degrade the >90% utilization claim, and pair-wise slot donation can absorb only adjacent-lane residual variance, not a global shift. Please add a sensitivity study or provide a formal bound on the mismatch.
  4. [§2.2, Eq. (5)] The quantity defined in Eq. (5) is not actual PE utilization. For G lanes operating in lockstep, true utilization is C̄/Cmax (total useful lane-cycles divided by G·Cmax), whereas Eq. (5) equals (G·C̄/Cmax − 1)/(G − 1). The two differ: for example, a group with costs [4,1,1,1] has true utilization 7/16 = 43.75%, but Eq. (5) gives 25%. All reported utilization numbers (56–64% for Laconic, >90% for BRIM) are in this transformed metric, so the headline 'over 90% PE utilization' claim is not literally utilization. Please either report true utilization or clearly rename the metric (e.g., 'normalized balance') and avoid using 'PE utilization' in the abstract.
minor comments (4)
  1. [Table 1] The 'PEs' column mixes notations such as '32×10' for Stripes/BitL and '32×32' for Laconic/BRIM. The text says all baselines are scaled to match BRIM’s 32×32 footprint; clarify whether the table lists the resulting PE counts and why some are 32×10.
  2. [Figure 7] The text refers to 'purple bar' and 'pink bar,' but the figure appears to use grayscale and may not be distinguishable in print. Use labels or hatched patterns in addition to color.
  3. [Algorithm 1, line 8] The text says t* = round(C*/τ), and Algorithm 1 uses the nearest-integer bracket notation. This is consistent, but the caption or pseudocode could explicitly state that ⌊·⌉ denotes round-to-nearest to avoid confusion.
  4. [General] Minor typographical issues: the ACM reference format shows '2018' in the copyright line while the venue is 'ICCAD ’26', and the page numbers are placeholder. These should be fixed in the camera-ready version.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the central speedup/utilization claims are measured empirically; only minor self-citation of [26] for the utilization formula is present.

full rationale

The main derivation chain is not circular. CBP's target assignment t*_k = floor(C*/tau_k) is explicitly constructed so that t*_k * tau_k ≈ C* (Sec. 3.1.1, "By construction..."), but this is the algorithm's design objective, not a fitted parameter being relabeled as a prediction. The reported results—56-64% baseline utilization, >90% BRIM utilization, up to 2.37x speedup, and energy improvements—are obtained from a cycle-accurate simulator, RTL synthesis at FreePDK 45nm, and CACTI-based SRAM modeling, against external baselines (Stripes, BitL, Laconic, Laconic-Xbar). These are empirical measurements, not consequences of the balancing equation. The only self-citation is ref. [26] for the PE utilization formula (Eq. 5) and the workload-balanced pruning concept; that prior work shares authors with this paper, but it is used as a definition and conceptual precursor, not as an unverified external theorem that forces the conclusion. The paper's contribution—Booth-term-count balancing in dual-sided bit-serial hardware—is evaluated independently. Two non-circularity concerns are worth noting separately: (1) the abstract's "at most 1.2% accuracy degradation" is internally inconsistent with Table 2 (e.g., ViT-S W4A8 drops 80.56→79.2 = 1.36pp; OPT-2.7B W4A8 perplexity rises 20.45→22.3 = 1.85), which is a correctness/consistency issue, not circularity; and (2) the reliance on a 128-sample activation profile is a robustness limitation, not a circularity, because the evaluation uses disjoint test data and the profile is not fitted to the reported utilization numbers.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The design rests on three tuned hyperparameters (δ, G, S) and on the representativeness of calibration statistics; no new physical entities are introduced. The Booth LUT and histogram are implementation details, not fitted parameters.

free parameters (4)
  • Cycle-budget offset set O = {-6, -4, -2, 0}
    Chosen by hand as the candidate offsets for the per-group cycle budget C* = mean + δ; the per-group selection minimizes squared weight distortion, but the candidate range is a tunable hyperparameter (Sec. 3.1.1).
  • Lockstep group size G = 16
    PE capacity is 16 lanes; Sec. 5.4 sweeps G on OPT-6.7B and selects G=16 as the Pareto knee. This is a hardware design parameter that controls how many weights CBP must modify.
  • Calibration set size S = 128
    The activation profile and Hessian are estimated from 128 samples (Sec. 4.2); larger or smaller sets would change the profile quality and the balancing/accuracy tradeoff.
  • Phase-2 block size B_s = not reported
    Algorithm 2 processes columns in blocks of B_s; the value is not stated in the paper, leaving an unquantified hyperparameter inherited from GPTQ.
assumptions (5)
  • domain assumption Per-lane execution cost equals the product of Booth term counts, #PP = η(w)·η(a)
    Invoked in Eq. (4) and throughout Sec. 2.2. Arithmetically true for pairwise term products, but treating it as the lane's cycle count assumes the datapath has no other serialization overhead.
  • domain assumption Lockstep lane groups: G lanes advance only when the slowest lane finishes
    Used to define PE utilization in Eq. (5) and to motivate equalization in Algorithm 1; it is a design choice of the accelerator, not a physical law.
  • domain assumption Activation term-count profile τ from 128 calibration samples generalizes to inference
    Sec. 3.1.1 uses the mean Booth term count over S samples; Sec. 4.2 sets S=128. CBP fixes weights offline, so runtime distribution shifts would break the balance and the >90% utilization claim.
  • domain assumption GPTQ Hessian-based error compensation recovers accuracy after weight replacement
    Phase 2 (Algorithm 2) relies on the inverse input-correlation Hessian and block error propagation following GPTQ [7]; the paper provides no independent proof of convergence or accuracy recovery.
  • standard math Booth encoding's signed non-adjacent form reduces non-zero term counts vs binary
    Standard arithmetic property used to justify the choice of Radix-2 Booth encoding in Sec. 2.2; not introduced by this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BRIM: Workload-Balanced Dual-Sided Bit-Serial Sparse Inference Accelerator." pith.science (2026). https://pith.science/paper/6G34OA2V

@misc{pith2026260719431,
  author       = {Pith},
  title        = {Pith review of: BRIM: Workload-Balanced Dual-Sided Bit-Serial Sparse Inference Accelerator},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6G34OA2V}},
  note         = {Machine review of arXiv:2607.19431}
}
read the original abstract

Bit-serial accelerators exploit bit-level sparsity to reduce DNN inference cost, but existing designs exploit sparsity on only one operand, bounding the speedup. Extending sparsity exploitation to both operands simultaneously yields compounding reductions in partial products but introduces a critical new bottleneck: workload imbalance. Because each concurrent weight - activation pair's execution cost depends on the product of two independently varying operand non-zero bit counts, pairs that must complete together finish at vastly different times, leaving faster computations idle. We show this limits PE utilization to 56 - 64% in existing dual-sided designs. We present BRIM, a hardware - software co-designed dual-sided bit-serial sparse accelerator that directly targets this bottleneck. BRIM combines two integrated mechanisms: 1) Cyclic-Balanced Pruning (CBP), a post-training weight optimization that reshapes weight representations based on profiled activation statistics to equalize expected workloads across concurrently processed pairs offline; and 2) Pairwise Slot Donation, a lightweight hardware mechanism that absorbs residual runtime imbalance with negligible area overhead. Evaluated across CNNs, ViTs, and LLMs under iso-area constraints, BRIM achieves over 90% PE utilization, up to 2.37x speedup, and up to 1.63x energy efficiency improvement over prior dual-sided designs.

Figures

Figures reproduced from arXiv: 2607.19431 by the authors.

Figure 1
Figure 1. Workload imbalance in dual-sided bit-serial exe [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Layer-wise PE utilization of a dual-sided bit-serial [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. End-to-end BRIM pipeline. term count at its position. This ensures the per-lane partial product counts 𝜂(𝑤) · 𝜂(𝑎) are equalized within each lockstep group and with weight replacement carried out efficiently via lookup tables (Section 3.1.1). Phase 2 (Column-wise Error Compensation) realizes these targets by replacing weights with the nearest LUT candidate at the assigned term count, then propagating the resulting e… view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: Top-level architec￾ture. A 32 × 32 PE array is fed by banked activation and weight buffers through Booth encoders (Enc.). An output￾stationary dataflow keeps par￾tial sums local to each PE. The PE is organized into two stages: (1) a slot dona￾tion front-end that dynam￾…
Figure 6
Figure 6. Figure 6: Processing element. Stage 1: slot donation — each [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 8
Figure 8. Figure 8: Average Speedup Comparison of various schemes [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 9
Figure 9. Figure 9: Average Energy Efficiency Comparision of various [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]
Figure 10
Figure 10. Figure 10: Group size (𝐺) design space on OPT-6.7B (W8A8). The Pareto curve plots perplexity increase against normal￾ized EDP improvement (both relative to 𝐺=4); the inset re￾ports the fraction of weights CBP modifies at each group size. 𝐺=16 (star) achieves the best EDP–accurac…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 6 linked inside Pith

  1. [1]

    Jorge Albericio, Alberto Delmás, Patrick Judd, Sayeh Sharify, Gerard O’Leary, Roman Genov, and Andreas Moshovos. 2017. Bit-Pragmatic Deep Neural Net- work Computing. In2017 50th Annual IEEE/ACM International Symposium on Microarchitecture (MICRO). 382–394

    2017

  2. [2]

    Andrew D. Booth. 1951. A Signed Binary Multiplication Technique.The Quarterly Journal of Mechanics and Applied Mathematics4, 2 (1951), 236–240. doi:10.1093/ qjmam/4.2.236

    1951

  3. [3]

    Yunji Chen, Tao Luo, Shaoli Liu, Shijin Zhang, Liqiang He, Jia Wang, Ling Li, Tianshi Chen, Zhiwei Xu, Ninghui Sun, et al . 2014. Dadiannao: A machine- learning supercomputer. In2014 47th Annual IEEE/ACM international symposium on microarchitecture. IEEE, 609–622

    2014

  4. [4]

    Yuzong Chen, Jian Meng, Jae-sun Seo, and Mohamed S Abdelfattah. 2024. BBS: Bi- directional bit-level sparsity for deep learning acceleration. In2024 57th IEEE/ACM International Symposium on Microarchitecture (MICRO). IEEE, 551–564

    2024

  5. [5]

    Yu-Hsin Chen, Tushar Krishna, Joel S Emer, and Vivienne Sze. 2016. Eyeriss: An energy-efficient reconfigurable accelerator for deep convolutional neural networks.IEEE journal of solid-state circuits52, 1 (2016), 127–138

    2016

  6. [6]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929(2020)

    arXiv 2020

  7. [7]

    Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. 2022. Gptq: Accurate post-training quantization for generative pre-trained transformers. arXiv preprint arXiv:2210.17323(2022)

    arXiv 2022

  8. [8]

    Amir Gholami, Zhewei Yao, Sehoon Kim, Coleman Hooper, Michael W Mahoney, and Kurt Keutzer. 2024. Ai and memory wall.IEEE Micro44, 3 (2024), 33–39

    2024

Show all 28 references
  1. [9]

    Ashish Gondimalla, Noah Chesnut, Mithuna Thottethodi, and T. N. Vijaykumar

  2. [10]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 770–778

    2016

  3. [11]

    Patrick Judd, Jorge Albericio, Tayler Hetherington, Tor M Aamodt, and Andreas Moshovos. 2016. Stripes: Bit-serial deep neural network computing. In2016 49th Annual IEEE/ACM International Symposium on Microarchitecture (MICRO). IEEE, 1–12

    2016

  4. [12]

    Seunghyun Lee, Dongho Ha, Sungbin Kim, Sungwoo Kim, Hyunwuk Lee, and Won Woo Ro. 2025. BitL: A Hybrid Bit-Serial and Parallel Deep Learning Acceler- ator for Critical Path Reduction. InProceedings of the 58th IEEE/ACM International Symposium on Microarchitecture. 1565–1578

    2025

  5. [13]

    Jiayao Ling, Gang Li, Xiaolong Lin, Cheng Gu, Xing Li, Jian Cheng, and Xiaoyao Liang. 2025. An Efficient Bit-Sparse DNN Accelerator Exploiting Adaptive Bit- Serial Computations.IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems44, 11 (2025), 4249–426...

    2025

  6. [14]

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. 2021. Swin transformer: Hierarchical vision transformer using shifted windows. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 10012–10022

    2021

  7. [15]

    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 Inter- national Symposium on Microarchitecture. 963–976

    2021

  8. [16]

    Olin Lowe Macsorley. 2007. High-speed arithmetic in binary computers.Proceed- ings of the IRE49, 1 (2007), 67–91

    2007

  9. [17]

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

    2009

  10. [18]

    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 Microarchi...

    2017

  11. [19]

    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 n...

    2017

  12. [20]

    David Patterson, Joseph Gonzalez, Quoc Le, Chen Liang, Lluis-Miquel Munguia, Daniel Rothchild, David So, Maud Texier, and Jeff Dean. 2021. Carbon emissions and large neural network training.arXiv preprint arXiv:2104.10350(2021)

    2021 arXiv

  13. [21]

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners.OpenAI blog 1, 8 (2019), 9

    2019

  14. [22]

    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

  15. [23]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yas- mine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhos- ale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models.arXiv preprint arXiv:2307.09288(2023)

    2023 arXiv

  16. [24]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2023. Attention Is All You Need. arXiv:1706.03762 [cs.CL] https://arxiv.org/abs/1706.03762

    2023 arXiv

  17. [25]

    Gang Wang, Siqi Cai, Zhenyu Li, Wenjie Li, Dongxu Lyu, Yanan Sun, Jianfei Jiang, and Guanghui He. 2025. BitPattern: Enabling Efficient Bit-Serial Acceleration of Deep Neural Networks through Bit-Pattern Pruning. In2025 62nd ACM/IEEE Design Automation Conference (DAC). 1–7. doi...

    2025

  18. [26]

    Ruokai Yin, Youngeun Kim, Yuhang Li, Abhishek Moitra, Nitin Satpute, Anna Hambitzer, and Priyadarshini Panda. 2023. Workload-Balanced Pruning for Sparse Spiking Neural Networks.IEEE Transactions on Emerging Topics in Computa- tional Intelligence8 (2023), 2897–2907. https://api...

    2023

  19. [27]

    Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, Todor Mihaylov, Myle Ott, Sam Shleifer, Kurt Shuster, Daniel Simig, Punit Singh Koura, Anjali Sridhar, Tianlu Wang, and Luke Zettlemoyer. 2...

    2022 arXiv

  20. [2019]

    InProceedings of the 52nd Annual IEEE/ACM International Symposium on Mi- croarchitecture(Columbus, OH, USA)(MICRO-52)

    SparTen: A Sparse Tensor Accelerator for Convolutional Neural Networks. InProceedings of the 52nd Annual IEEE/ACM International Symposium on Mi- croarchitecture(Columbus, OH, USA)(MICRO-52). Association for Computing Machinery, New York, NY, USA, 151–165. doi:10.1145/3352460.3358291

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.