{"id":"57000a67-4259-4641-aff7-d5da91da19a8","arxiv_id":"2411.19430","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A core placement method using deep reinforcement learning is claimed to reduce communication cost and improve throughput for SNN training on many-core near-memory systems in unreleased simulator experiments.","lead":"Reinforcement learning is used to choose where to place the pieces of a spiking neural network on the cores of a many-core near-memory chip, with the goal of reducing communication and training time. The paper reports simulated communication-cost reductions of roughly 20 to 50 percent over simple and learned baselines, but ships no code, no data, and no power measurements.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline training-time and power reductions depend on an unverified Section 4.3 assumption that delay and power are linear in communication cost, so all reported gains are projections of communication improvements through a proxy.","rationale":"The paper's intended contribution is a placement method that reduces SNN training time, communication cost, and power on a many-core near-memory architecture. The strongest defensible result is a reduction in inter-core communication: the reward is defined on communication, and Figures 6(a) and 8(a) report such reductions. The transition from communication reduction to training-time and power reduction is the weak link. Section 4.3 assumes linearity without support, and the simulator is not available, so the reported latency, throughput, and power numbers cannot be independently verified. The internal DDPG/PPO inconsistency (abstract/title vs. Section 3.3 and Eq. 5) is real and should be corrected, but it is not the most load-bearing issue: even a correctly labeled PPO or DDPG method would still need the linearity assumption and simulator validation. Because the reader's conditional verdict already reflects this gap, I recommend keeping the verdict unchanged pending code/data release and validation.","tokens_in":12689,"tokens_out":6187,"duration_ms":53989,"concrete_test":"Release the simulator (or a reproducible binary) and, for the 32-core Spike-ResNet18 training workload, measure end-to-end training cycles and power for both the optimized mapping and the Zigzag baseline using an added cycle-accurate power model or an FPGA prototype of the Section 4.1 architecture. Compare these measurements with values predicted from the reported communication-cost reduction under the Section 4.3 linearity assumption; if the measured training-time or power improvement deviates from the predicted value by more than 10%, the proxy assumption fails and the headline claim needs to be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is carried by an unverified proxy. Section 4.3 (Reward) states: 'The power consumption and delay are linear with respect to communication cost, so the reward function of Actor Network only needs to consider the communication between computing cores.' No derivation, measurement, or citation is provided for this linear relation. The policy is therefore trained on communication cost alone; the reported reductions in training time (Section 5.2, Figures 6(b)-(c), 8(b)) and the promised power reduction (Abstract, Conclusion) are not direct measurements but are inferred from lower communication cost through that linear relation. If true delay/power is dominated by compute, off-chip storage access, or pipeline bubbles, or if network energy is nonlinear under congestion, the RL agent may simply be optimizing a surrogate that does not transfer to the claimed end-to-end quantities. The simulator that produced all numbers is unreleased and no RTL/board validation is provided, so this bridge cannot currently be checked. This is load-bearing because without it the reward does not target the headline metrics.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a deep reinforcement learning method for physical core placement of spiking neural network training workloads on a many-core near-memory architecture, combined with a balanced storage/computation partitioning method and FPDeep-style pipelining. The authors evaluate on simulated 32-core and 64-core systems with Spike-ResNet18, Spike-VGG16, and Spike-ResNet50, reporting reductions in communication cost, latency, hop count, and improved throughput relative to Zigzag, Sigmate, random search, and Policy. The central claim is that the method reduces SNN training time and power consumption by optimizing inter-core communication.","tokens_in":12833,"tokens_out":5158,"duration_ms":40932,"significance":"If the results hold, the paper addresses an important problem—mapping logical SNN training graphs to physical many-core systems—and would extend RL-based placement from inference to training. The use of graph convolution to encode the logical graph and architecture topology is a reasonable and potentially impactful approach, and the idea of jointly optimizing partitioning and placement is credible. The paper also provides useful comparison against existing placement methods and visualizes hotspot reduction. However, the significance is tempered by the fact that the headline gains in latency and power are not directly measured but projected through an unverified linear relationship with communication cost, and the simulator is unreleased.","major_comments":[{"comment":"The assertion that 'power consumption and delay are linear with respect to communication cost' is load-bearing for the paper's central claims. The reported latency reductions (Section 5.2, Figures 6(b)-6(c), 8(b)) and the promised power reduction (Abstract, Conclusion) are not direct measurements but consequences of this assumption. No derivation, measurement, or citation is provided for this linear relation, and the statement 'the reward function of Actor Network only needs to consider the communication between computing cores' does not follow from linearity alone. Moreover, Eq. (4) is not readable: the terms in the max expression are placeholders (??????? etc.) and are never defined, so the reward is not reproducible. The authors should either measure delay and power directly in the simulator or provide empirical evidence (e.g., scatter plots of delay/power vs. communication cost for representative mappings) that the linear model holds.","section":"Section 4.3 (Reward), Eq. (4)"},{"comment":"The algorithm identity is inconsistent. The abstract and Section 5.1 state 'Off-policy Deterministic Actor-Critic' and 'based on DDPG', while Section 3.3 says 'We employ the Proximal Policy Optimization (PPO) algorithm' and Section 4.3 says 'optimizing the surrogate loss with PPO'. DDPG and PPO are distinct algorithms, and the update rule in Eq. (5) resembles PPO's clipped surrogate, not DDPG. The paper should state clearly which algorithm is implemented, reconcile the terminology (including 'near-end policy optimization' in the abstract), and specify how the Actor-Critic architecture with graph convolution is trained under the chosen algorithm.","section":"Sections 3.3 and 4.3 vs. Abstract and Section 5.1"},{"comment":"The evaluation lacks error bars or multiple-seed reporting for the RL method. Because the policy is stochastic and the discretization/collision-resolution procedure introduces randomness, the reported improvements (e.g., 28.84%, 32.92%, 18.89% training communication cost reduction) could be seed-dependent or tuned. At minimum, the authors should report mean and standard deviation over at least 5 random seeds for each model and core count. In addition, the simulator is not released, and no validation against RTL simulation or a hardware prototype is described, so the absolute values in Figures 6-8 cannot be externally checked; the paper should include a description of the simulator's accuracy or make it available.","section":"Section 5.1 and Figures 6-8"},{"comment":"The reward function is communication cost, and the headline evaluation metric is also communication cost (Figures 6(a), 8(a)). While the paper also reports latency, throughput, and hop distance, these are said to be reduced as a consequence of the same linear assumption, rather than measured independently. To break the circularity, the simulator should report cycle-accurate latency and energy/power directly for the final placements, and ideally rank placements by those metrics without reference to the reward proxy. Please also clarify in the text which quantities are measured and which are derived from the linear model.","section":"Section 5.2 (metric independence and circularity)"},{"comment":"The paper claims that the FPDeep-style pipelining 'significantly reduc[es] model training time' and improves utilization, but Figure 9 only shows qualitative waveforms of core utilization and does not report training time or throughput numbers with and without pipelining. Since 'reducing model training time' is part of the central claim, the evaluation should include quantitative end-to-end training latency or cycles for the layer-wise vs. FPDeep pipelines, ideally combined with the proposed placement.","section":"Section 5.2 and Figure 9"}],"minor_comments":[{"comment":"The phrase 'near-end policy optimization' should read 'proximal policy optimization'.","section":"Abstract"},{"comment":"Equation (4) contains undefined placeholders; the variables should be spelled out and the connection between the reward text and the equation made explicit.","section":"Equation (4)"},{"comment":"The acronym 'EER' is introduced without definition; it likely means 'energy efficiency ratio,' and the units (fps/W) suggest so. Please define it at first use.","section":"Table 1"},{"comment":"Definitions A-C contain garbled mathematical symbols (e.g., '???'); please typeset all definitions properly.","section":"Section 3.1"},{"comment":"The sentence 'the latency and throughput of the proposed method is reduced by about 8%' is contradictory; throughput should increase while latency decreases. Please correct.","section":"Section 5.2, line near Figures 6(b)-6(c)"},{"comment":"The caption 'The computing resource usage Layer-wise and FPDeep' is missing prepositions; also, the figure panels are not all referenced in the text with sufficient detail.","section":"Figure 9 caption"},{"comment":"The sentence 'which is 0.167 less than the average transmission distance of single data packet of the Policy method' is redundant with the preceding number; please tighten.","section":"Section 5.3"},{"comment":"References [32] and [33] appear mis-ordered in the text: SemiMap is described after Prime, but the citation [32] points to TrueNorth in the reference list; verify the citation numbering.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper builds closely on prior RL-based placement work (Wu et al., Myung et al.), and the novelty boundary is not clearly drawn in the text; I would ask the authors to specify what is new beyond applying an RL-based mapper to SNN training with a graph-convolutional policy. The unverified linearity assumption and the unreleased simulator are the main blockers; however, the core mapping idea is plausible and the required fixes are within scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Couple of things you should know. The genuinely new bit is applying RL-based core placement to SNN training, not inference, and pairing it with a storage/computation-balanced partitioning scheme. That is a real extension of the Policy/Core Placement line, and the authors clearly know that literature. The FPDeep-style pipelining discussion is a nice addition, though not deeply evaluated.\n\nWhat they do well: the problem is clearly stated, the partitioning idea targets a genuine issue (uniform partitioning leaves later SNN layers unbalanced), and they compare against Zigzag, Sigmate, random search, and Policy on 32- and 64-core simulated systems. The hop-distance and hotspot visualizations are helpful.\n\nThe soft spots are around the evidence. Section 4.3 says power and delay are linear in communication cost, so the reward only considers inter-core communication. No derivation or measurement is provided. That assumption carries a lot of weight: the training-time and power reductions in Section 5 are not measured; they are inferred from lower communication cost. If network energy or latency is dominated by compute or off-chip access, the headline numbers fall apart. The reward being the same metric as the main evaluation also means part of the reported gain is built into the objective. The simulator is unreleased, there are no error bars or multiple seeds, and power is never actually measured. There is also a DDPG/PPO identity mismatch: the abstract says off-policy deterministic actor-critic, but the body updates via PPO. Minor, but sloppy.\n\nNone of this is fatal to the core idea. Reducing inter-core communication during SNN training is plausibly a good objective, and the balanced partitioning is reasonable. But the paper overstates what is demonstrated. It needs code/data release, a corrected algorithm description, and a sensitivity check on the linearity assumption before the quantitative claims stand.\n\nWho this is for: neuromorphic architects and people working on many-core mapping/compilation. It is a legitimate engineering paper. If you are deciding whether to referee it, I would send it out, but flag that the evaluation evidence needs major strengthening. The direction is timely and the method is sensible; the claims just outrun the evidence.","headline":"A sensible extension of RL-based core placement to SNN training, but the headline numbers are projections through an unverified linearity assumption.","tokens_in":13447,"tokens_out":1670,"would_cite":false,"duration_ms":15245,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A reinforcement-learning placement method cut simulated inter-core communication during spiking neural network training by up to 32.92%, beating Zigzag, Sigmate, random search, and Policy baselines.","keywords":["spiking neural networks","SNN training","core placement","reinforcement learning","many-core architecture","near-memory computing","deployment optimization","graph convolution"],"falsifier":"Measure wall-clock training time and chip power on a physical or cycle-accurately calibrated many-core near-memory system running Spike-ResNet18, Spike-VGG16, and Spike-ResNet50 with the proposed placement versus Zigzag. If the training-time and power reductions are substantially smaller than the communication-cost reductions (28.84%, 32.92%, 18.89%), the linear-reward assumption that carries the argument is false.","tokens_in":12477,"feed_emoji":"🧠","tokens_out":4403,"duration_ms":34413,"temperature":0.7,"pith_summary":"The paper claims that mapping the logical cores of a spiking neural network (SNN) training workload onto a many-core near-memory chip is a combinatorial optimization problem that can be solved with deep reinforcement learning better than with heuristics. It proposes a three-part method: a balanced partitioning that splits each layer unevenly along input and output channels to equalize computation and storage latency; an actor-critic policy network that takes the logical graph and the chip's on-chip network topology as inputs and outputs a placement via continuous values discretized to physical cores; and an FPDeep-style fine-grained pipeline that overlaps computation between layers. On a simulated 32-core system, the method reports training communication cost reduced by 28.84% (Spike-ResNet18), 32.92% (Spike-VGG16), and 18.89% (Spike-ResNet50) relative to Zigzag, Sigmate, random search, and the Policy baseline, with similar gains on 64 cores. The sympathetic reading is that reinforcement-learning-based placement, previously applied to inference, can also cut communication and improve utilization during the more demanding training phase.","feed_headline":"RL placement cuts SNN training traffic by up to 33%","feed_subtitle":"Graph-aware actor-critic mapping beats Zigzag, Sigmate, and Policy on 32-core near-memory SNN training.","key_machinery":"The load-bearing mechanism is an off-policy deterministic actor-critic framework with a graph-convolutional policy network. The logical SNN graph and the physical on-chip network topology are encoded as features; the actor outputs continuous coordinates for each logical core, which are clipped and discretized to physical core positions, with ties broken by a clockwise shortest-path search. The reward is a function of inter-core communication data volume only, based on the assumption that power and delay are linear in communication cost. A balanced partitioning step splits each layer along input and output channels so that per-core computation plus storage transmission time is balanced, and an FPDeep-style pipeline overlaps layer computations to raise core utilization.","core_discovery":"The central claim is that SNN training deployment on many-core near-memory systems can be optimized by treating placement as a graph-aware reinforcement learning problem and by balancing both computation and storage when partitioning the model. The paper reports that its deployment scheme reduces training communication cost by 28.84% for Spike-ResNet18, 32.92% for Spike-VGG16, and 18.89% for Spike-ResNet50 on a 32-core simulated architecture, outperforming Zigzag, Sigmate, random search, and the Policy baseline; the method also generalizes to 64 cores and improves ANN inference placement, beating Policy by 6.5% in average communication cost. Because power consumption and delay are assumed linear in communication cost, these communication reductions are presented as reductions in training time and chip power.","pith_inferences":["The paper's reported training-time and power gains rest on the unverified linear relation between communication cost, delay, and power; on hardware where compute or off-chip memory dominates, the placement gains would be smaller than claimed.","The clockwise shortest-path routing assumption is static; under adaptive or congestion-aware routing, the placement optimized for one routing policy may not be optimal, so the method's gains are routing-policy-dependent.","A direct test on a physical many-core near-memory prototype, measuring wall-clock training time and power rather than simulated communication cost, would settle whether the linearity assumption holds."],"forward_implications":["If correct, SNN training on many-core near-memory chips becomes more feasible: communication-heavy training can be mapped to reduce data movement and hotspot formation.","The method extends beyond SNNs: the same placement approach improves ANN inference on a Tianjic-like architecture, beating the Policy baseline by 6.5%.","The approach scales from 32 to 64 cores with communication cost reduced by more than 22.64%, suggesting it generalizes to larger many-core systems.","Fine-grained pipelining combined with placement increases the number of cores computing simultaneously, improving utilization and shortening training clock cycles.","Better placement reduces average hop count (by 0.67 per packet vs Zigzag), which should lower both latency and energy per transmission."],"supporting_citations":[{"why":"Supplies the Policy baseline method and the reward-function design that this paper extends from inference to training.","marker":"[16]"},{"why":"Provides the Core Placement balanced partitioning approach and the multi-chip many-core deployment problem formulation.","marker":"[7]"},{"why":"Contributes the FPDeep deep-pipeline computation method the paper adapts for fine-grained pipelining.","marker":"[25]"},{"why":"Motivates the many-core near-memory SNN training architecture with forward, backward, and weight-update engines.","marker":"[13]"},{"why":"Supplies the Tianjic many-core architecture used as the simulation platform for the ANN inference comparison.","marker":"[14]"}],"fun_headline_variants":["RL placement cuts SNN training traffic by up to 33%","Graph-aware actor-critic beats Zigzag on 32-core SNN training","Deep RL maps SNN training cores to cut communication cost 33%","DRL placement reduces SNN training comms 33% on many-core chips"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results assume that power consumption and delay are linear in inter-core communication cost, so the reward function only needs to minimize communication; if that linearity does not hold, the reported training-time and power reductions would not follow from the measured communication savings.","fun_headline_variants_meta":{"raw":{"variants":["RL placement cuts SNN training traffic by up to 33%","Graph-aware actor-critic beats Zigzag on 32-core SNN training","Deep RL maps SNN training cores to cut communication cost 33%","DRL placement reduces SNN training comms 33% on many-core chips"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000252,"raw_usage":{"total_tokens":1591,"prompt_tokens":1005,"completion_tokens":586,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":505}},"tokens_in":621,"tokens_out":586,"duration_ms":4999,"temperature":1.0,"reasoning_tokens":505,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:11:43.932585+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure wall-clock training time and chip power on a physical or cycle-accurately calibrated many-core near-memory system running Spike-ResNet18, Spike-VGG16, and Spike-ResNet50 with the proposed placement versus Zigzag. If the training-time and power reductions are substantially smaller than the communication-cost reductions (28.84%, 32.92%, 18.89%), the linear-reward assumption that carries the argument is false.","supporting_citations":[{"cited_title":"Policy gradient-based core placement optimization for multichip many-core systems","cited_arxiv_id":null,"evidence_quote":"Supplies the Policy baseline method and the reward-function design that this paper extends from inference to training."},{"cited_title":"Core placement optimization for multichip many-core neural network systems with reinforcement learning","cited_arxiv_id":null,"evidence_quote":"Provides the Core Placement balanced partitioning approach and the multi-chip many-core deployment problem formulation."},{"cited_title":"Fpdeep: Scalable acceleration of cnn training on deeply-pipelined fpga clusters","cited_arxiv_id":null,"evidence_quote":"Contributes the FPDeep deep-pipeline computation method the paper adapts for fine-grained pipelining."},{"cited_title":"H2learn: High-efficiency learning accelerator for high-accuracy spiking neural networks","cited_arxiv_id":null,"evidence_quote":"Motivates the many-core near-memory SNN training architecture with forward, backward, and weight-update engines."},{"cited_title":"Tianjic: A unified and scalable chip bridging spike-based and continuous neural computation[J]","cited_arxiv_id":null,"evidence_quote":"Supplies the Tianjic many-core architecture used as the simulation platform for the ANN inference comparison."}],"review_version":1}