Pith. sign in

REVIEW 4 major objections 4 minor 44 references

Hardware-Aware Data and Instruction Mapping for AI Tasks: Balancing Parallelism, I/O and Memory Tradeoffs

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read The paper argues that predictable deep-learning inference can be compiled into a self-sequencing message stream, with 97% of VGG-19 messages generated on-chip and 88-92% utilization.

desk verdict A coherent extension of the authors' MAVeC program with useful scheduling tables, but the headline numbers are not established because the simulator is unvalidated and 'utilization' is undefined. read the letter →

arxiv 2509.03846 v1 pith:BDPEQEX2 submitted 2025-09-04 cs.AR cs.LG

classification cs.ARcs.LG
keywords AIacceleratorstreamingarchitecturemessage-drivenexecutiondatareuseoff-chiptrafficreductionVGG-19reconfigurablearraydeterministicrouting
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

This paper tries to establish that deep-learning inference is regular enough to be compiled into a single stream of messages that carry both the data and the instruction for what to do next, so the hardware can run layer after layer without the host stepping in. The target is the MAVeC fabric, a grid of small processing sites connected by multicast and reduction buses; each 64-bit packet contains an opcode, a destination address, a 32-bit operand, and the next opcode and destination. On VGG-19 inference the authors report that over 97% of all messages are produced on-chip, about 89% of cycles are spent moving messages between nearby sites rather than fetching from off-chip, and a 64x64 array keeps 88-92% of sites busy while sustaining over 1 TFLOP/s. If this holds, DNN accelerator performance would no longer be gated primarily by PCIe and DRAM bandwidth, because the fabric is primed once and then streams to completion.

What carries the argument

The load-bearing object is the 64-bit message: a 4-bit present opcode, 12-bit present address, 32-bit payload (weight, activation, or partial sum), 4-bit next opcode, and 12-bit next address. The message is both instruction and data; consuming it at a SiteO produces a new message for the next SiteO, so sequencing is carried by the packet itself. It is supported by a fixed C0-to-C3 routing plan in which active columns multiply, reserved columns perform staged reductions, and results offload to on-chip L1, plus workload-pattern bits (TStream, Shift, Identity) that let one packet spawn shift and inter-tile forwarding packets. This combination is what removes host-driven load-execute-store bound

What would settle it

Run the same VGG-19 layers on an RTL or FPGA implementation of a 64x64 MAVeC array with real FIFO backpressure, bus arbitration, and memory timing; compare total cycles, message counts, and SiteO utilization with the simulator. If measured on-chip-generated messages fall well below 97% or utilization below 88% under identical mapping and input, the self-sequencing claim is not supported.

Watch

Extended reading notes

Core claim

The central claim is that by co-packing control and data into one 64-bit message, a spatially programmable message-passing accelerator can execute an entire inference workload as a self-sequencing stream. Each SiteO decodes the present opcode and address, performs the operation on the 32-bit operand, then rewrites the message with the next opcode, the next address, and the result, so every communication step also schedules the following one. The paper shows a deterministic four-stage reduction pipeline (C0 multiply, C1 column sum, C2 depth sum, C3 multi-depth aggregation) with weights held stationary in SiteO registers, activations multicast across rows, and partial sums reduced in place bef

Load-bearing premise

The load-bearing premise is stated in Section IV.A: measurements come from a cycle-accurate simulator built on analytic models in [36] and [37], not from fabricated hardware, so if that simulator is optimistic about traffic flow on the fixed C0-to-C3 routing and multicast/reduction buses, the reported 97% on-chip message share and 88-92% utilization would not hold on real silicon.

Editorial extensions

If this is right

  • If the central claim holds, a DNN accelerator can be primed once and then execute a layer stack without host-driven DMA epochs; host link speed stops being the main latency bottleneck.
  • The near-linear throughput scaling with array size (roughly doubling from 16x16 to 32x32 and exceeding 1 TFLOP/s at 64x64) makes the number of SiteOs, not off-chip bandwidth, the primary scaling lever.
  • Traffic reductions up to about 100 MB per layer from stationary weights, multicast, and staged reduction mean designs can trade expensive high-bandwidth DRAM for more on-chip buffers.
  • Deterministic scheduling with an Identity bit suggests the same streaming model can handle skip connections and layer boundaries without flushing the fabric, so end-to-end inference can run as a single resident pipeline.

Reading between the lines

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

  • The same reasoning would suggest that other predictable workloads, such as transformers with fixed sequence lengths or graph convolution with static topology, could also be compiled into self-sequencing message streams, though the paper only demonstrates VGG-19.
  • The near-insensitivity to DRAM family implies a design consequence the authors do not spell out: memory cost and power could be cut by choosing the cheapest DRAM that still meets priming bandwidth, rather than the fastest available.
  • A natural extension is to vary batch size or input resolution after compilation; if the schedule is truly self-sequencing, only the image-injection rate should change, with no recompilation of the routing plan.
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. This manuscript presents MAVeC, a programmable message-passing accelerator and a deterministic mapping framework for DNN inference. The key idea is a unified 64-bit packet that packs opcode, addresses, and data so each processing SiteO executes and forwards a message, enabling the array to self-sequence after host priming. The paper claims on VGG-19 that >97% of messages are generated on-chip, ~89% of cycles are spent on message movement, SiteO utilization reaches 88-92% on a 64x64 array, and convolution throughput exceeds 1 TFLOP/s. Results are obtained from a cycle-accurate, message-level simulator built on the authors' analytic models from [36] and [37], with a case study of a 4x4x4 convolution.

Significance. If substantiated, the work would be a useful contribution: co-encoding control and data into a single stream is a clean way to remove host-paced execution, and deterministic C0-C1-C2-C3 routing with hierarchical multicast/reduction gives a concrete recipe for locality. The paper also makes an honest attempt to quantify reuse traffic per layer. However, the central quantitative claims are currently unsupported: the simulator is unvalidated, the utilization metric is undefined and appears inconsistent with the cycle breakdown, and the evaluation covers only the convolutional layers of VGG-19. These are not presentation issues; they are load-bearing for the paper's headline results.

major comments (4)
  1. [Section IV.A, Fig. 6 and Fig. 8] The sole evidence for the headline metrics is 'a cycle-accurate, message-level simulator leveraging analytic models in [36], [37]'. No validation against RTL simulation, FPGA measurement, or an independent architecture model is reported, and no equations or assumptions of the simulator are given here. Because the analytic models are by the same authors, the >97% on-chip message and 88-92% utilization results are largely entailed by the model definitions rather than demonstrated. To make the claims load-bearing, the authors should validate the simulator against a concrete baseline (e.g., Eyeriss/Timeloop-style mappings or RTL) and provide a sensitivity analysis of the model's assumptions (FIFO stalls, bus contention, multicast timing).
  2. [Section IV.B, Fig. 6(b) vs Fig. 8(a)] 'Average SiteO utilization' is never defined. Figure 6(b) reports 88.5% of cycles as message movement and only 8.7% as arithmetic; Figure 8(a) reports 88-92% utilization. If utilization counts any cycle in which a SiteO forwards, stores, or consumes a message, the number reflects bus activity, not useful compute, and the claim as stated is misleading. If it counts arithmetic, the numbers cannot both be correct. Please define the metric precisely and reconcile it with the cycle breakdown.
  3. [Table 4 and Section II] The evaluation is limited to the 'VGG-19 convolutional stack', yet the abstract and introduction claim 'applied to VGG-19', and Section II states the pipeline sustains high utilization 'across convolutional, pooling, and fully connected stages'. Table 2 includes pooling (CMP) and fully connected (FC) messages, but no pooling or FC layer results are reported. The paper should either evaluate the full VGG-19 network including FC/pooling layers or explicitly scope all claims to convolution layers; the current text overstates the system-level conclusion.
  4. [Section III.A and Fig. 6(a)] The message-count decomposition (97.85% on-chip-generated) is not, by itself, evidence of reduced off-chip traffic. Because every reduction hop and multicast fan-out is counted as a generated message, the metric can be high even if the same bytes traverse the fabric many times. The paper should report byte-level off-chip traffic (or compare with a dataflow baseline) and show that the on-chip message generation translates into actual reductions in host/DRAM traffic. Without that, the central I/O-bottleneck claim is not established.
minor comments (4)
  1. [General] Typos and formatting issues: 't he' in the abstract, 'IDIB-IDFF' in Table 2 header is unexplained, and Table 3 has dense notation (O(NF)^C_Index, IDPS/PSShape) that should be defined in a legend.
  2. [Section IV.A] The baseline uses 'DDR7' as off-chip memory; if DDR7 is speculative, state that explicitly or use currently available memory standards (e.g., DDR5/LPDDR5/GDDR6) to avoid confusion.
  3. [Section II] Memory capacity numbers need checking: 0.125 KB per L2 Tile buffer is extremely small, and the claimed 24.5 MB per Quad is not obviously derivable from the stated L1/L0 sizes (96 KB per SiteM, 64B per SiteO). Please clarify the hierarchy and per-array capacities.
  4. [References] Reference [35] is a preprint with no venue; if a published version exists, cite it. Also, [36] and [37] are central to the simulator and should be described sufficiently in the text.

Circularity Check

1 steps flagged · score 6.0 of 10

Headline performance metrics rest on a self-referential simulator built from the authors' own analytic models, with no independent validation.

  1. self citation load bearing [Section IV.A (Evaluation Methodology)]
    "Measurements are obtained with a cycle-accurate, message-level simulator leveraging analytic models in [36], [37]."

    The simulator is the sole source for every headline quantitative claim (97% on-chip messages, 88–92% utilization, >1 TFLOP/s, ~89% transfer cycles). References [36] and [37] are prior works by the same authors describing the same MAVeC architecture and its analytic models. No independent RTL implementation, FPGA prototype, or third-party simulator is reported, and the analytic models are not shown to be validated against external baselines. Thus the evaluation is a self-referential loop: the framework is judged by a simulator whose underlying models are the authors' own prior constructs, so the central performance claims reduce to a self-citation chain rather than to an independent, falsifiable test.

full rationale

The paper's mapping and streaming design is described in enough detail (Section III, Tables 2–3, Figures 2–5) to be a concrete algorithmic contribution, and the VGG-19 message schedule is not itself a fitted artifact. However, the performance evaluation that supports every headline number comes exclusively from a simulator that 'leverages analytic models in [36], [37]' — both prior works by the same authors on the same MAVeC architecture. No external validation is provided, so the quantitative results are load-bearing on a self-referential evaluation loop. Additionally, 'average SiteO utilization' is never defined; with the reported cycle mix (8.7% arithmetic vs 88.5% message movement), the 88–92% utilization figure may be counting packet forwarding as useful activity, making it a renamed measure of bus occupancy rather than compute utilization. That is a reporting gap rather than a circular step, but it compounds the evaluation's unreliability. The paper's internal derivations (C0→C1→C2→C3 routing, the example walkthrough) are consistent, so this is partial circularity rather than a fully definitional collapse: score 6.

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

No numeric free parameters are fitted; array sizes and PCIe/DRAM configurations are swept inputs. However, the simulator's analytic models themselves are unvalidated, which is captured in the axioms. No new physical entities are introduced; MAVeC and the message-passing model come from the authors' prior work.

assumptions (4)
  • domain assumption The deterministic C0->C1->C2->C3 routing and multicast/reduction buses operate without unmodeled contention or FIFO stalls.
    Section III.B describes fixed routing; Section IV.A reports cycle counts from an analytic simulator. No congestion modeling is described, so the transfer-dominated cycle breakdown rests on this assumption.
  • domain assumption The analytic models in self-cited works [36], [37] accurately predict MAVeC performance.
    Section IV.A: 'leveraging analytic models in [36], [37]' with no independent validation or comparison to standard simulators or hardware.
  • domain assumption Filter weights fit in on-chip L0/L1 storage and remain stationary across each fold, so each weight is loaded from host exactly once.
    Section III.D says weights are stationary to maximize temporal reuse; Section IV.B(a) reports 2.13% host-injected weight messages. If capacity were exceeded, reloads would raise host traffic.
  • domain assumption A 64-bit packet with 4-bit opcodes, 12-bit addresses, and 32-bit payload can encode all VGG-19 layer transitions including padding, stride, and pooling.
    Section III.A introduces workload pattern bits (TStream, Shift, Identity) with 'unused bits left available for future extensions'; the sufficiency of this encoding for all layers is asserted, not proven.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hardware-Aware Data and Instruction Mapping for AI Tasks: Balancing Parallelism, I/O and Memory Tradeoffs." pith.science (2026). https://pith.science/paper/BDPEQEX2

@misc{pith2026250903846,
  author       = {Pith},
  title        = {Pith review of: Hardware-Aware Data and Instruction Mapping for AI Tasks: Balancing Parallelism, I/O and Memory Tradeoffs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BDPEQEX2}},
  note         = {Machine review of arXiv:2509.03846}
}
read the original abstract

We introduce a mapping framework for deep learning inference that takes advantage of predictable neural network behavior to plan both computation and communication ahead of time. The framework generates a unified stream of instructions and data, enabling the hardware to execute operations and route information on its own, without frequent involvement from the host and with minimal off-chip memory use. This naturally reduces reliance on I/O, off-chip memory, and host control. By leveraging fine-grained message passing on a programmable, message-based compute architecture, the framework keeps data movement local and coordinates computation across the array using techniques such as stationary-weight reuse, in-array multicasting, and staged reductions. Applied to VGG-19, the framework sustains high utilization (88 to 92 percent), with over 97 percent of messages generated internally and nearly 89 percent of time consumed on-chip transfers. Computation throughput scales beyond 1 TFLOP/s on larger arrays, while traffic reductions from reuse and local aggregation reach up to 100 MB per layer. Overall, the results highlight the effectiveness of streaming-based computation and show how our mapper enables this execution style by tightly coordinating data and instruction flow across the hardware.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 31 canonical work pages

  1. [36]

    Greedy Prefetch for Reducing Off -Chip Memory Accesses in Convolutional Neural Network Inference,

    D. Yang and L. Chen, “Greedy Prefetch for Reducing Off -Chip Memory Accesses in Convolutional Neural Network Inference,” Information, vol. 16, no. 3, p. 164, Feb. 2025, doi: 10.3390/info16030164

  2. [37]

    Present and Future, Challenges of High Bandwith Memory (HBM),

    K. Kim and M. Park, “Present and Future, Challenges of High Bandwith Memory (HBM),” in 2024 IEEE International Memory Workshop (IMW), IEEE, May 2024, pp. 1 –4. doi: 10.1109/IMW59701.2024.10536972

  3. [1]

    Eyeriss: An Energy -Efficient Reconfigurable Accelerator for Deep Convolutional Neural Networks,

    Y.-H. Chen, T. Krishna, J. S. Emer, and V. Sze, “Eyeriss: An Energy -Efficient Reconfigurable Accelerator for Deep Convolutional Neural Networks,” IEEE J Solid -State Circuits, vol. 52, no. 1, pp. 127 –138, Jan. 2017, doi: 10.1109/JSSC.2016.2616357

  4. [2]

    H. Kwon, A. Samajdar, and T. Krishna, “MAERI,” ACM SIGPLAN Notices, vol. 53, no. 2, pp. 461 –475, Nov. 2018, doi: 10.1145/3296957.3173176

  5. [3]

    DianNao,

    T. Chen et al. , “DianNao,” ACM SIGARCH Computer Architecture News , vol. 42, no. 1, pp. 269–284, Apr. 2014, doi: 10.1145/2654822.2541967

  6. [4]

    Then, IF#2-IF#4 inject Col3, Col4, and Col5 for channels {0, 1}; overlapping columns are forwarded laterally (blue arrows), so only the new column is fetched each time

    This pipeline (shift → multicast → multiply → reduction) continues until one output column over all filters is generated. Then, IF#2-IF#4 inject Col3, Col4, and Col5 for channels {0, 1}; overlapping columns are forwarded laterally (blue arrows), so only the new column is fetched each time. After these four shifts the four output columns for filters 0-3 ov...

  7. [5]

    Immediately after programming , A_MULS compute messages are multicast to all C-0 SiteOs (entry 2), each carrying an image data and the convolution-pattern bits

    to initialize the offload address (OA) with the initial multi- depth summation, A_ADDS for intermediate folds ( entry 7) to keep accumulating into OA, and A_ADD for the last fold (entry 6) to finish the accumulation and hold the result. Immediately after programming , A_MULS compute messages are multicast to all C-0 SiteOs (entry 2), each carrying an imag...

  8. [6]

    DaDianNao: A Machine-Learning Supercomputer,

    Y. Chen et al., “DaDianNao: A Machine-Learning Supercomputer,” in 2014 47th Annual IEEE/ACM International Symposium on Microarchitecture , IEEE, Dec. 2014, pp. 609 –622. doi: 10.1109/MICRO.2014.58

Show all 44 references
  1. [7]

    ShiDianNao,

    Z. Du et al., “ShiDianNao,” in Proceedings of the 42nd Annual International Symposium on Computer Architecture , New York, NY, USA: ACM, Jun. 2015, pp. 92 –104. doi: 10.1145/2749469.2750389

  2. [8]

    Krishna, H

    T. Krishna, H. Kwon, A. Parashar, M. Pellauer, and A. Samajdar, Data Orchestration in Deep Learning Accelerators . Cham: Springer International Publishing, 2020. doi: 10.1007/978 - 3-031-01767-4. Figure 8. Per-layer performance profiling under message-driven exec ution. (a) Av...

  3. [9]

    V. Sze, Y. -H. Chen, T. -J. Yang, and J. S. Emer, Efficient Processing of Deep Neural Networks . Cham: Springer International Publishing, 2020. doi: 10.1007/978-3-031-01766-7

  4. [10]

    An Evaluation of Edge TPU Accelerators for Convolutional Neural Networks,

    K. Seshadri, B. Akin, J. Laudon, R. Narayanaswami, and A. Yazdanbakhsh, “An Evaluation of Edge TPU Accelerators for Convolutional Neural Networks,” in 2022 IEEE International Symposium on Workload Characterization (IISWC), IEEE, Nov. 2022, pp. 79–91. doi: 10.1109/IISWC55918.2022.00017

  5. [11]

    Custom AI Streaming Accelerator Architecture,

    X. Y. Niu, C. Xiong, L. Jiao, and K. H. Tsoi, “Custom AI Streaming Accelerator Architecture,” in 2023 Asia Symposium on Image Processing (ASIP), IEEE, Jun. 2023, pp. 114 –122. doi: 10.1109/ASIP58895.2023.00026

  6. [12]

    Optimizing Off-Chip Memory Access for Deep Neural Network Accelerator,

    Y. Zheng, H. Yang, Y. Shu, Y. Jia, and Z. Huang, “Optimizing Off-Chip Memory Access for Deep Neural Network Accelerator,” IEEE Transactions on Circuits and Systems II: Express Briefs , vol. 69, no. 4, pp. 2316 –2320, Apr. 2022, doi: 10.1109/TCSII.2022.3150030

  7. [13]

    Minimizing Off-Chip Memory Access for CNN Accelerators,

    S. Tewari, A. Kumar, and K. Paul, “Minimizing Off-Chip Memory Access for CNN Accelerators,” IEEE Consumer Electronics Magazine , vol. 11, no. 3, pp. 95 –104, May 2022, doi: 10.1109/MCE.2021.3097697

  8. [14]

    EGCN: An Efficient GCN Accelerator for Minimizing Off-Chip Memory Access,

    Y. Han, K. Park, Y. Jung, and L.-S. Kim, “EGCN: An Efficient GCN Accelerator for Minimizing Off-Chip Memory Access,” IEEE Transactions on Computers , pp. 1 –13, 2022, doi: 10.1109/TC.2022.3211413

  9. [15]

    AI and Memory Wall,

    A. Gholami, Z. Yao, S. Kim, C. Hooper, M. W. Mahoney, and K. Keutzer, “AI and Memory Wall,” IEEE Micro, vol. 44, no. 3, pp. 33–39, May 2024, doi: 10.1109/MM.2024.3373763

  10. [16]

    Hitting the memory wall,

    Wm. A. Wulf and S. A. McKee, “Hitting the memory wall,” ACM SIGARCH Computer Architecture News, vol. 23, no. 1, pp. 20–24, Mar. 1995, doi: 10.1145/216585.216588

  11. [17]

    Occam: Optimal Data Reuse for Convolutional Neural Networks,

    A. Gondimalla, J. Liu, M. Thottethodi, and T. N. Vijaykumar, “Occam: Optimal Data Reuse for Convolutional Neural Networks,” ACM Transactions on Architecture and Code Optimization, vol. 20, no. 1, pp. 1–25, Mar. 2023, doi: 10.1145/3566052

  12. [18]

    Continuous Convolution Accelerator with Data Reuse based on Systolic Architecture,

    J. Park, S. An, J. Kim, and S. E. Lee, “Continuous Convolution Accelerator with Data Reuse based on Systolic Architecture,” in 2023 20th International SoC Design Conference (ISOCC) , IEEE, Oct. 2023, pp. 319 –320. doi: 10.1109/ISOCC59558.2023.10396060

  13. [19]

    MAESTRO: A Data - Centric Approach to Understand Reuse, Performance, and Hardware Cost of DNN Mappings,

    H. Kwon, P. Chatarasi, V. Sarkar, T. Krishna, M. Pellauer, and A. Parashar, “MAESTRO: A Data - Centric Approach to Understand Reuse, Performance, and Hardware Cost of DNN Mappings,” IEEE Micro, vol. 40, no. 3, pp. 20–29, May 2020, doi: 10.1109/MM.2020.2985963

  14. [20]

    A Survey of Coarse -Grained Reconfigurable Architecture and Design,

    L. Liu et al. , “A Survey of Coarse -Grained Reconfigurable Architecture and Design,” ACM Comput Surv, vol. 52, no. 6, pp. 1–39, Nov. 2020, doi: 10.1145/3357375

  15. [21]

    A Survey on Coarse-Grained Reconfigurable Architectures From a Performance Perspective,

    A. Podobas, K. Sano, and S. Matsuoka, “A Survey on Coarse-Grained Reconfigurable Architectures From a Performance Perspective,” IEEE Access, vol. 8, pp. 146719 –146743, 2020, doi: 10.1109/ACCESS.2020.3012084

  16. [22]

    Karunaratne, A

    M. Karunaratne, A. K. Mohite, T. Mitra, and L.-S. Peh, “HyCUBE,” in Proceedings of the 54th Annual Design Automation Conference 2017 , New York, NY, USA: ACM, Jun. 2017, pp. 1 –6. doi: 10.1145/3061639.3062262

  17. [23]

    Integrating NVIDIA Deep Learning Accelerator (NVDLA) with RISC -V SoC on FireSim,

    F. Farshchi, Q. Huang, and H. Yun, “Integrating NVIDIA Deep Learning Accelerator (NVDLA) with RISC -V SoC on FireSim,” in 2019 2nd Workshop on Energy Efficient Machine Learning and Cognitive Computing for Embedded Applications (EMC2) , IEEE, Feb. 2019, pp. 21 –

  18. [24]

    Timeloop: A Systematic Approach to DNN Accelerator Evaluation,

    A. Parashar et al. , “Timeloop: A Systematic Approach to DNN Accelerator Evaluation,” in 2019 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS), IEEE, Mar. 2019, pp. 304 –315. doi: 10.1109/ISPASS.2019.00042

  19. [25]

    doi: 10.1109/EMC249363.2019.00012

  20. [26]

    Ultra-Elastic CGRAs for Irregular Loop Specialization,

    C. Torng, P. Pan, Y. Ou, C. Tan, and C. Batten, “Ultra-Elastic CGRAs for Irregular Loop Specialization,” in 2021 IEEE International Symposium on High -Performance Computer Architecture (HPCA), IEEE, Feb. 2021, pp. 412 –

  21. [27]

    DRESC: a retargetable compiler for coarse -grained reconfigurable architectures,

    Bingfeng Mei, S. Vernalde, D. Verkest, H. De Man, and R. Lauwereins, “DRESC: a retargetable compiler for coarse -grained reconfigurable architectures,” in 2002 IEEE International Conference on Field-Programmable Technology,

  22. [28]

    Scratchpad Memory Management for Deep Learning Accelerators,

    S. Zouzoula, M. A. Maleki, M. W. Azhar, and P. Trancoso, “Scratchpad Memory Management for Deep Learning Accelerators,” in Proceedings of the 53rd International Conference on Parallel Processing, New York, NY, USA: ACM, Aug. 2024, pp. 629 –639. doi: 10.1145/3673038.3673115

  23. [29]

    Deep Learning Model Compression With Rank Reduction in Tensor Decomposition,

    W. Dai, J. Fan, Y. Miao, and K. Hwang, “Deep Learning Model Compression With Rank Reduction in Tensor Decomposition,” IEEE Trans Neural Netw Learn Syst, vol. 36, no. 1, pp. 1315– 1328, Jan. 2025, doi: 10.1109/TNNLS.2023.3330542

  24. [30]

    Design of a Convolutional Neural Network Accelerator Based on On -Chip Data Reordering,

    Y. Liu et al., “Design of a Convolutional Neural Network Accelerator Based on On -Chip Data Reordering,” Electronics (Basel), vol. 13, no. 5, p. 975, Mar. 2024, doi: 10.3390/electronics13050975

  25. [31]

    Twenty Years of Automated Methods for Mapping Applications on CGRA,

    K. J. M. Martin, “Twenty Years of Automated Methods for Mapping Applications on CGRA,” in 2022 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW) , IEEE, May 2022, pp. 679 –686. doi: 10.1109/IPDPSW55747.2022.00118

  26. [32]

    Neurostream: Scalable and Energy Efficient Deep Learning with Smart Memory Cubes,

    E. Azarkhish, D. Rossi, I. Loi, and L. Benini, “Neurostream: Scalable and Energy Efficient Deep Learning with Smart Memory Cubes,” IEEE Transactions on Parallel and Distributed Systems, vol. 29, no. 2, pp. 420 –434, Feb. 2018, doi: 10.1109/TPDS.2017.2752706

  27. [33]

    Evaluating Modern GPU Interconnect: PCIe, NVLink, NV-SLI, NVSwitch and GPUDirect,

    A. Li et al. , “Evaluating Modern GPU Interconnect: PCIe, NVLink, NV-SLI, NVSwitch and GPUDirect,” Mar. 2019, doi: 10.1109/TPDS.2019.2928289

  28. [34]

    Accelerating PageRank Algorithmic Tasks with a new Programmable Hardware Architecture,

    M. R. Hossain Chowdhury and M. Rahman, “Accelerating PageRank Algorithmic Tasks with a new Programmable Hardware Architecture,” in 2024 IEEE International Conference on Rebooting Computing (ICRC) , IEEE, Dec. 2024, pp. 1 –6. doi: 10.1109/ICRC64395.2024.10937012

  29. [35]

    Loom: Exploiting Weight and Activation Precisions to Accelerate Convolutional Neural Networks,

    S. Sharify, A. D. Lascorz, K. Siu, P. Judd, and A. Moshovos, “Loom: Exploiting Weight and Activation Precisions to Accelerate Convolutional Neural Networks,” in 2018 55th ACM/ESDA/IEEE Design Automation Conference (DAC), IEEE, Jun. 2018, pp. 1–6. doi: 10.1109/DAC.2018.8465915

  30. [38]

    The Evolution of the PCI Express (PCIe) Specification: In its Sixth Generation, Third Decade and Still Going Strong,

    PCI-SIG, “The Evolution of the PCI Express (PCIe) Specification: In its Sixth Generation, Third Decade and Still Going Strong,” https://pcisig.com/blog/evolution-pci-express- specification-its-sixth-generation-third-decade- and-still-going-strong

  31. [39]

    Diversification of DRAM Application and Memory Hierarchy,

    SK hynix Newsroom, “Diversification of DRAM Application and Memory Hierarchy,” https://news.skhynix.com/diversification-of- dram-application-and-memory-hierarchy/

  32. [40]

    Messaging-based Intelligent Processing Unit (m- IPU) for next generation AI computing,

    Md. R. H. Chowdhury and M. Rahman, “Messaging-based Intelligent Processing Unit (m- IPU) for next generation AI computing,” Oct. 2024

  33. [41]

    Implications of memory embedding and hierarchy on the performance of MAVeC AI accelerators,

    M. R. H. Chowdhury and M. Rahman, “Implications of memory embedding and hierarchy on the performance of MAVeC AI accelerators,” Memories - Materials, Devices, Circuits and Systems , vol. 10, p. 100131, Apr. 2025, doi: 10.1016/j.memori.2025.100131

  34. [42]

    Demystifying the 7 -D Convolution Loop Nest for Data and Instruction Streaming in Reconfigurable AI Accelerators,

    M. R. H. Chowdhury and M. Rahman, “Demystifying the 7 -D Convolution Loop Nest for Data and Instruction Streaming in Reconfigurable AI Accelerators,” Jul. 2025, [Online]. Available: http://arxiv.org/abs/2507.20420

  35. [425]

    doi: 10.1109/HPCA51647.2021.00042

  36. [2002]

    Proceedings

    (FPT). Proceedings. , IEEE, pp. 166 –173. doi: 10.1109/FPT.2002.1188678

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.