Pith. sign in

REVIEW 5 major objections 6 minor 22 references

High-Speed Ultra-Energy-Efficient Memristor-Based Massive MIMO SIC Detector Circuit with Hybrid Analog-Digital Computing Architecture

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Memristor circuit performs MIMO SIC detection 43 times faster than a DSP chip, the paper reports.

desk verdict A genuinely new circuit architecture for memristor-based SIC with clean derivation and conservative timing, but the headline 43x/110x ratios rest on estimates that need opening up. read the letter →

arxiv 2506.03495 v1 pith:ZGJS4F3U submitted 2025-06-04 eess.SP

Jia-Hui Bi , Shaoshi Yang , Sheng Chen , Ping Zhang This is my paper
classification eess.SP
keywords memristorcrossbararrayin-memorycomputinganalogmatrixmassiveMIMOsuccessiveinterferencecancellationMMSE-SICdetectorhybridanalog-digitalslicercomputationalenergyefficiency
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

The paper sets out to show that successive interference cancellation, a nonlinear MIMO detection algorithm normally implemented digitally, can be realized in a memristor crossbar circuit. The design uses K stages of analog matrix computing modules paired with hybrid analog-digital slicers, where each slicer converts an analog voltage into a digital selection code and then outputs the nearest constellation voltage. For a 32 by 64 16-QAM system, the authors report a computing speed of about 5.5 TOPS and a computational energy efficiency of about 1.41 TOPS/W, which they compare against a DSP, an FPGA, and a GPU. If accurate, the result would extend memristor-based in-memory computing from linear detection and maximum likelihood detection to an iterative nonlinear detector, with order-of-magnitude speed and energy advantages over digital processors.

What carries the argument

The load-bearing object is the proposed matrix computing module, which consists of six memristor crossbar arrays, three sets of analog inverters, and two sets of operational amplifiers. Applying Ohm's law and Kirchhoff's current law at the operational amplifier inputs yields the identity vout = (D3^T D2 + λ1λ2 I)^-1 D3^T (λ0 vin1 - D1 vin2), so a single analog circuit computes a regularized least-squares estimate in steady state. The second key mechanism is the hybrid analog-digital slicer, built from voltage comparators and an analog multiplexer, which compares the incoming voltage against constellation thresholds and selects the nearest symbol value, closing the iterative SIC loop without converting the full data path to digital.

What would settle it

Fabricate or SPICE-model the first matrix computing module with the stated 500 MHz gain-bandwidth product operational amplifier, a realistic memristor parasitic capacitance of a few hundred femtofarads, and conductance matrices programmed with a 6-bit range and typical device-to-device variation, then measure the settling time of the first output to within one threshold spacing. If T2 exceeds 130 nanoseconds or if the paired conductance matrices differ by more than a few percent, the claimed 4.87 microsecond total time and 1.41 TOPS/W efficiency do not hold.

Watch

Extended reading notes

Core claim

The central claim is that the MMSE-SIC algorithm, which requires a sequence of regularized matrix inversions interleaved with symbol slicing, can be mapped onto a pipeline of analog modules. Each matrix computing module consolidates the computation into the steady-state relation vout = (D3^T D2 + λ1λ2 I)^-1 D3^T (λ0 vin1 - D1 vin2), where D1, D2, and D3 are differences of memristor conductance matrices. The paper shows that mapping the channel matrix onto D2 and D3, the interference-cancelled received signal onto the input voltages, and the noise variance onto λ1λ2 reproduces the SIC update; the slicer then quantizes the first element of the output vector. The authors demonstrate by simulation that at least 6-bit memristor precision is needed for near-digital BER, and they report a total computing time of about 4.87 microseconds and an energy consumption of 18.98 microjoules for the 32 by 64 scenario.

Load-bearing premise

The load-bearing premise is that the analog circuit behaves ideally: operational amplifiers have perfect virtual ground and zero input current, the memristor crossbars add no parasitic capacitance or resistance, and two separately programmed conductance matrices D2 and D3 can be made identical; if any of these fails, the modeled 130 nanosecond convergence time and the reported 43x and 110x margins erode.

Editorial extensions

If this is right

  • The SIC detector, previously considered too nonlinear for memristor crossbars, can be built from a K-stage pipeline of identical analog matrix modules plus slicers.
  • At 6-bit memristor precision, detection performance approaches the digital MMSE-SIC benchmark, suggesting that imperfect analog devices are acceptable for this application.
  • The reported 4.87 microsecond total time and 18.98 microjoule energy consumption are the first concrete power-speed numbers for analog SIC detection.
  • The architecture scales with the number of users K because each stage reuses the same module structure, implying larger MIMO systems would require proportionally more stages but no new circuit topology.
  • Higher-order QAM can be handled with the indirectly select slicer, which uses a combinational logic stage to reduce the multiplexer channel count from exponential to linear in the constellation size.

Reading between the lines

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

  • The same matrix computing module could be reused for any regularized least-squares step, meaning the design generalizes beyond SIC to other iterative linear estimators such as weighted least squares or Kalman update steps.
  • Because the claimed margins rest on ideal analog behavior, a natural next test is to simulate the circuit with realistic memristor parasitic capacitance and D2/D3 programming mismatch; the paper does not report such nonideality sweeps.
  • If the energy numbers hold at scale, the per-bit detection energy for a 32-user 16-QAM frame becomes several orders of magnitude lower than current baseband processors, which would matter for battery-limited user equipment only if the memristor array itself can be integrated at low cost.
  • The slice-and-cancel loop suggests a direct analog extension to iterative decision-feedback equalization in single-carrier and OFDM receivers, where the same comparator-multiplexer slicer could close the feedback path.
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

5 major / 6 minor

Summary. The manuscript proposes a memristor-crossbar-based circuit for MMSE successive interference cancellation (SIC) detection in massive MIMO. The architecture has K stages, each with a matrix computing module (six crossbars, three inverter sets, two operational-amplifier arrays) and two hybrid analog-digital slicers. The authors derive the module's input-output relation (10) from KCL, describe a conductance mapping scheme (11)-(12), simulate BER under memristor quantization, and estimate computing speed and energy efficiency in a 32x64 16-QAM scenario, reporting 5.5 TOPS and 1.41 TOPS/W and claiming 43x speed and 110x energy-efficiency advantages over an 8-core DSP.

Significance. If the performance estimates were fully substantiated, the work would be a significant step in extending analog in-memory computing to nonlinear MIMO detection. The KCL-based derivation of the matrix module is clean and self-consistent, and the hybrid slicer design is novel. The BER results showing that 6-bit memristor precision is sufficient are a useful design guideline. However, the central quantitative claims (speed and energy efficiency) are not yet auditable, and the comparison baselines are not task-matched.

major comments (5)
  1. [Section IV.B] The paper reports an equivalent computing speed of 5.5 TOPS and a computational energy efficiency of 1.41 TOPS/W, but the "equivalent FLOP number" used to compute these values is never disclosed. Without the exact FLOP count for the 32x64 16-QAM MMSE-SIC task, including the dimensions of each matrix operation and the number of multiply-accumulates, the speed and efficiency numbers cannot be reproduced. Please provide this count explicitly.
  2. [Section IV.B] The energy figure of 18.98 µJ is presented without any component-level breakdown. The text mentions "Joule dissipation on each memristor" and "typical energy consumption of all other components," but no numbers are given for the number of memristors, OAs, comparators, multiplexers, ADCs, or DACs, nor their individual energy contributions. Please provide a full breakdown and explicitly state whether memristor programming/configuration energy is included. Without this, the claimed energy efficiency and the 110x ratio over the DSP cannot be verified.
  3. [Section IV.B, Table II] The comparison with the DSP, FPGA, and GPU is not task-matched. References [20]-[22] are generic kernel benchmarks and datasheets, not implementations of the MMSE-SIC algorithm for a 32x64 16-QAM system. The claimed 43x/110x advantages over the DSP are therefore ratios of an analog estimate to unspecified digital baselines. Please either run the same detection algorithm on these platforms with optimized implementations, or provide a transparent derivation of the baseline time and energy values used in Table II.
  4. [Section III.B, after Eq. (10)] The mapping of the channel matrix and noise variance onto the conductance values is underspecified. For (10) to compute (G^T G + σ²I)^{-1}G^T y, the two crossbar pairs D2 and D3 must be programmed to the same matrix ~G, but the paper does not state how identical programming is achieved, how D2/D3 mismatch is handled, or how σ_n² is mapped to the product λ1λ2, nor how λ0, λ1, and λ2 are chosen relative to the conductance range. Please provide a complete mapping procedure and analyze the effect of D2/D3 mismatch, since this directly affects the BER results in Section IV.A.
  5. [Section IV.B, timing model] The timing model uses K(T2+Tslicer) as the maximum convergence time, and the SPICE validation is shown for a 4x4 noise-free system (Fig. 6) and for the first matrix module of a 32x64 system (Fig. 7), but not for the full 32-stage cascade. Please provide evidence that T2 < 130 ns holds for all stages in the 32x64 scenario, including the later stages whose inputs depend on slicer decisions, or state clearly that the claimed 4.87 µs includes an additional per-stage margin.
minor comments (6)
  1. [Section II.B, Eqs. (5)-(6)] Please specify the dimensions of the identity matrices in (5) and (6): they should be 2K x 2K and 2(K-k+1) x 2(K-k+1), respectively.
  2. [Section III.C, first paragraph] The directly select structure is described as using a "2W−1-channel" multiplexer, but the subsequent examples (W=4 uses 8 channels, W=8 uses 128 channels) indicate the intended expression is 2^{W−1}-channel. Please correct this notation.
  3. [Table II] The energy consumption values for the DSP, FPGA, and GPU (2.1 mJ, 343.1 µJ, 199.7 µJ) are given without citation to a specific benchmark or calculation; please provide the source of these numbers or the method used to derive them.
  4. [Section IV.A] Please specify the SNR range, number of channel realizations, and the definition of the digital benchmark in Fig. 5 to improve reproducibility.
  5. [Section IV.B] The term "FLOP" is used for both a floating-point operation and a throughput unit; please use "FLOPs" for count and "TOPS" or "FLOPS" consistently throughout.
  6. [Fig. 2] A sentence explaining the function of the second set of operational amplifiers (those with noninverting inputs) and the role of the ground connections would help readers understand the KCL derivation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the circuit derivation is self-contained and the headline performance numbers are external benchmark comparisons.

full rationale

The paper's central derivation is the KCL analysis of the matrix computing module: equations (8)-(10) express the output voltage vout as (D3^T D2 + λ1λ2 I)^-1 D3^T(λ0 vin1 - D1 vin2), and the design then maps the SIC quantities ~G, ~y, σ_n^2 onto D2, D3, vin1, λ1λ2, and the cancellation term onto D1 and vin2. This is a constructive mapping, not a circular one: the circuit output is derived from Kirchhoff's and Ohm's laws under the stated ideal-OA assumptions, and the algorithm (3)-(7) is the standard MMSE-SIC recursion. No fitted parameter is later renamed as a prediction: T2 is obtained from SPICE waveforms, the slicer delays are taken from component datasheets, and the total 4.87 µs and 18.98 µJ are sums of these component-level estimates. The two self-citations ([1], [10]) are background references and are not load-bearing for the claimed SIC implementation or for the benchmark ratios. The main quantitative claim is assessed against external DSP/FPGA/GPU numbers, so it does not reduce to the paper's own assumptions by construction. Concerns that the equivalent FLOP count is undisclosed, that the digital baselines may not be task-matched, and that parasitic/mismatch effects are idealized are correctness and reproducibility risks, not circularity, and do not warrant a nonzero circularity score under the stated rules.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claims rest on three kinds of unstated inputs: (i) scalar conductances (lambda0, lambda1, lambda2) and the normalization (beta) that bridge the mathematical SIC equations to the physical circuit, for which no numeric values or scaling rules are given; (ii) ideal circuit assumptions (perfect OA virtual ground, no parasitic crossbar capacitance, programmability of six crossbars with only quantization error); and (iii) a modeling choice for the comparison metric (equivalent FLOPs). No new physical entities are introduced. These inputs are engineering design parameters rather than fitted data, but they are load-bearing because they determine both the BER and the reported speed and efficiency.

free parameters (3)
  • lambda0, lambda1, lambda2 (scalar conductances) = not specified in the paper
    Chosen by hand to map y, sigma squared, and the SIC feedback onto the circuit of Eq. (10). The product lambda1*lambda2 is set to represent sigma squared, but the conversion between the MIMO model values and physical conductances is never stated. This choice affects the achievable BER and dynamic range.
  • beta (conductance normalization in Eq. (12)) = (alpha_max - alpha_min) / max{|o_ij|}
    Normalizes each mapped matrix O onto the conductance range [0.1, 30] uS in Eq. (12). It depends on the specific matrix being mapped and controls the quantization error, but no example values for the MIMO channel matrices are given.
  • v0 (reference voltage for slicer) = 0.1 V
    Reference voltage setting the 16 QAM constellation levels and comparator thresholds in the slicer (Section III.C). Chosen for simulation; its value is a design parameter, not derived from the system model.
assumptions (5)
  • standard math Ideal operational amplifier behavior: virtual short between inputs and zero input currents
    Used to derive Eqs. (8)-(10) in Section III.B; standard in analog circuit analysis.
  • domain assumption Memristor conductance can be programmed to any value in [alpha_min, alpha_max] and the six crossbar arrays are ideal (no parasitic capacitance or resistance, no mismatch between D2 and D3)
    Section III.B and the mapping scheme (11)-(12); the paper does not model parasitic effects, device-to-device variability beyond quantization, or the fact that D2 and D3 would be realized by physically different crossbars with mismatch.
  • domain assumption The total convergence time is K*(T2 + Tslicer) with a single worst-case T2 (<=130 ns) for every stage and a fixed slicer delay; all stages are sequential and no pipelining overlap is assumed
    Section IV.B, statement 'the maximum convergence time of the proposed circuit is K(T2 + Tslicer)'. This assumes the largest stage's settling time bounds all stages and that slicer transitions are instantaneous after Tslicer.
  • domain assumption The equivalent FLOP count of the SIC algorithm is the same for the analog circuit and the digital processors, and is computed from the algorithm's digital complexity
    Section IV.B defines computing speed as equivalent FLOPs divided by time. This metric is a modeling choice; the paper does not state the exact FLOP count formula or how it was derived for the 32x64 16-QAM case.
  • domain assumption The BER simulation error model: memristor precision is modeled as quantization of conductance values to bits, but the exact quantization and error injection procedure is not described
    Section IV.A and Fig. 5; the mapping scheme (11)-(12) is given but not the rounding rule, the distribution of channel matrices, or the number of Monte Carlo trials.

how reviews work

0 comments
Cite this review

Pith. "Pith review of High-Speed Ultra-Energy-Efficient Memristor-Based Massive MIMO SIC Detector Circuit with Hybrid Analog-Digital Computing Architecture." pith.science (2026). https://pith.science/paper/ZGJS4F3U

@misc{pith2026250603495,
  author       = {Pith},
  title        = {Pith review of: High-Speed Ultra-Energy-Efficient Memristor-Based Massive MIMO SIC Detector Circuit with Hybrid Analog-Digital Computing Architecture},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZGJS4F3U}},
  note         = {Machine review of arXiv:2506.03495}
}
read the original abstract

The emerging memristor crossbar array based computing circuits exhibit computing speeds and energy efficiency far surpassing those of traditional digital processors. This type of circuits can complete high-dimensional matrix operations in an extremely short time through analog computing, making it naturally applicable to linear detection and maximum likelihood detection in massive multiple-input multiple-output (MIMO) systems. However, the challenge of employing memristor crossbar arrays to efficiently implement other nonlinear detection algorithms, such as the successive interference cancellation (SIC) algorithm, remains unresolved. In this paper we propose a memristor-based circuit design for massive MIMO SIC detector. The proposed circuit comprises several judiciously designed analog matrix computing modules and hybrid analog-digital slicers, which enables the proposed circuit to perform the SIC algorithm with a hybrid analog-digital computing architecture. We show that the computing speed and the computational energy-efficiency of the proposed detector circuit are 43 times faster and 110 times higher, respectively, than those of a traditional 8-core digital signal processor (DSP), and also advantageous over the benchmark high-performance field programmable gate array (FPGA) and graphics processing unit (GPU).

Figures

Figures reproduced from arXiv: 2506.03495 by the authors.

Figure 1
Figure 1. Structure of the proposed memristor-based SIC detec [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Circuit design of the proposed matrix computing modu [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Structures of the proposed slicer: (a) the directly s [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: The proposed slicer with: (a) the directly select str [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 7
Figure 7. Figure 7: An example of the voltage waveforms of the 64 output no [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]
Figure 6
Figure 6. Figure 6: Voltage waveforms at input and output nodes of slicer [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages

  1. [20]

    Optimization and evaluation of image- and signal- processing kernels on the TI C6678 multi-core DSP ,

    B. Ramesh, et al. , “Optimization and evaluation of image- and signal- processing kernels on the TI C6678 multi-core DSP ,” in Proc. HPEC 2014 (Waltham, MA, USA), Sep. 9-11, 2014, pp. 1–6

    work page 2014

  2. [22]

    Data sheet: NVIDIA RTX A1000,

    “Data sheet: NVIDIA RTX A1000,” NVIDIA, 2024. [Online] . Available: https://www.nvidia.com/en-us/design-visualization/rtx-a1000/

    work page 2024

  3. [1]

    Fifty years of MIMO detection: The r oad to large-scale MIMOs,

    S. Y ang and L. Hanzo, “Fifty years of MIMO detection: The r oad to large-scale MIMOs,” IEEE Commun. Surveys Tuts. , vol. 17, no. 4, pp. 1941–1988, 4th Quart. 2015

    work page 1941

  4. [2]

    Direction-of-arrival esti mation for large antenna arrays with hybrid analog and digital archite ctures,

    R. Zhang, B. Shim, and W. Wu, “Direction-of-arrival esti mation for large antenna arrays with hybrid analog and digital archite ctures,” IEEE Trans. Signal Process. , vol. 70, pp. 72–88, 2022

    work page 2022

  5. [3]

    Channel training-aided target sensing for terahertz integrated sensing and massive MIMO communications,

    R. Zhang, et al. , “Channel training-aided target sensing for terahertz integrated sensing and massive MIMO communications,” IEEE Internet Things J. , vol. 12, no. 4, pp. 3755–3770, Feb. 2025

    work page 2025

  6. [4]

    Solving matrix equations in one step with cross-point resistive arrays,

    Z. Sun, et al. , “Solving matrix equations in one step with cross-point resistive arrays,” Proc. Nat. Acad. Sci , vol. 116, no. 10, pp. 4123–4128, Mar. 2019

    work page 2019

  7. [5]

    Researc h progress on memristor: From synapses to computing systems,

    X. Y ang, B. Taylor, A. Wu, Y . Chen, and L. O. Chua, “Researc h progress on memristor: From synapses to computing systems, ” IEEE Trans. Circuits Syst. I, Reg. Papers , vol. 69, no. 5, pp. 1845–1857, May. 2022

    work page 2022

  8. [6]

    Memristor crossbar-based ultra-efficient next-generation baseband processors,

    G. Y uan, et al., “Memristor crossbar-based ultra-efficient next-generation baseband processors,” in Proc. MWSCAS 2017 (Boston, MA, USA), Aug. 6-9, 2017, pp. 1121–1124

    work page 2017

Show all 22 references
  1. [7]

    An analogue i n-memory ridge regression circuit with application to massive MIMO a cceleration,

    P . Mannocci, E. Melacarne, and D. Ielmini, “An analogue i n-memory ridge regression circuit with application to massive MIMO a cceleration,” IEEE J. Emerg. Sel. Topics Circuits Syst. , vol. 12, no. 4, pp. 952–962, Dec. 2022

    2022

  2. [8]

    Realizing in-memory baseband processing for ultrafast and energy-efficient 6G,

    Q. Zeng, et al., “Realizing in-memory baseband processing for ultrafast and energy-efficient 6G,” IEEE Internet Things J. , vol. 11, no. 3, pp. 5169–5183, Feb. 2024

    2024

  3. [9]

    Rethinking massive M IMO detection: A memristor approach,

    Y . Fang, L. Chen, C. Y ou, and H. Yin, “Rethinking massive M IMO detection: A memristor approach,” IEEE Commun. Lett. , vol. 27, no. 12, pp. 3350–3354, Dec. 2023

    2023

  4. [10]

    Accelerating maximum- likelihood detection in massive MIMO: A new paradigm with me mristor crossbar based in-memory computing circuit,

    Y .-H. Ren, S. Y ang, J.-H. Bi, and Y .-X. Zhang, “Accelerating maximum- likelihood detection in massive MIMO: A new paradigm with me mristor crossbar based in-memory computing circuit,” IEEE Trans. V eh. Tech- nol., vol. 73, no. 12, pp. 19745–19750, Dec. 2024

    2024

  5. [11]

    Layered space-time architecture for w ireless commu- nication in a fading environment when using multi-element a ntennas,

    G. J. Foschini, “Layered space-time architecture for w ireless commu- nication in a fading environment when using multi-element a ntennas,” Bell Labs Tech. J. , vol. 1, no. 2, pp. 41–59, Autumn 1996

    1996

  6. [12]

    Y . S. Cho, J. Kim, W. Y . Y ang, and C. G. Kang, MIMO-OFDM Wireless Communications with MATLAB . Hoboken, NJ, USA: Wiley, 2010

    2010

  7. [13]

    Data sheet: LT1016,

    “Data sheet: LT1016,” Analog Devices, Inc., 2018. [Onl ine]. Available: https://www.analog.com/media/en/technical-documentation/data-sheets/ lt1016.pdf

    2018

  8. [14]

    Data sheet: ADG1608/ADG1609,

    “Data sheet: ADG1608/ADG1609,” Analog Devices, Inc., 2015. [Online]. Available: https://www.analog.com/media/en/ technical- documentation/data-sheets/ADG1608 1609.pdf

    2015

  9. [15]

    Data sheet: MAX900–MAX903,

    “Data sheet: MAX900–MAX903,” Analog Devices, Inc., 20 05. [Online]. Available: https://www.analog.com/media/en/ technical- documentation/data-sheets/MAX900-MAX903.pdf

  10. [16]

    Data sheet: ADG708/ADG709,

    “Data sheet: ADG708/ADG709,” Analog Devices, Inc., 20 14. [Online]. Available: https://www.analog.com/media/en/ technical- documentation/data-sheets/ADG708 709.pdf

  11. [17]

    Algorithm/hardware co-des ign config- urable SAR ADC with low power for computing-in-memory in 28n m CMOS,

    Z. Guo, D. Chen, and X. Xue, “Algorithm/hardware co-des ign config- urable SAR ADC with low power for computing-in-memory in 28n m CMOS,” in Proc. ASICON 2021 (Kunming, China), Oct. 26-29, 2021, pp. 1–4

    2021

  12. [18]

    Design of low power 8-bit DAC using PTM-LP technology,

    S. M. Ishraqul Huq, S. Islam, N. Saqib, and S. N. Biswas, “ Design of low power 8-bit DAC using PTM-LP technology,” in Proc. ICRTEECT 2017 (Warangal, India), Jul. 30-31, 2017, pp. 64–69

    2017

  13. [19]

    An analog preconditioner for solving linear sys- tems,

    B. Feinberg, et al. , “An analog preconditioner for solving linear sys- tems,” in Proc. HPCA 2021 (Seoul, Korea), Feb. 27-Mar. 3, 2021, pp. 761–774

    2021

  14. [21]

    GPU vs FPGA performance comparison,

    “GPU vs FPGA performance comparison,” BERTEN DSP S.L., White Paper, 2016. [Online]. Available: https://www.bertendsp .com/gpu-vs- fpga-performance-comparison/

    2016

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.