{"id":"44179c20-9b23-4fc6-9ebc-356ec5ba06c1","arxiv_id":"2501.18371","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"The paper proposes FLASH-FHE, a heterogeneous FHE accelerator with bootstrappable and swift clusters, and reports speedups of 1.4x to 11.2x for deep workloads and up to 8x for parallel shallow workloads in cycle-accurate simulation.","lead":"FLASH-FHE is a chip design that puts two kinds of computing engines on one fully homomorphic encryption (FHE) accelerator: one type for deep encrypted computations that need bootstrapping and one cheaper type for shallow computations that do not. The authors report speedups over earlier FHE chips of 1.4x to 11.2x for deep workloads and up to 8x throughput for running many shallow jobs at once.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 8x shallow-workload speedup is attributed to heterogeneity, but CraterLake's 8 computation groups could plausibly run 8 jobs in parallel under a different scheduler; the paper provides no analysis ruling this out, so the central claim is not yet established.","rationale":"Reader identified the same weakest assumption, and I agree it is the load-bearing point. The paper's headline claim has two parts: 1.4x/11.2x deep speedup and up to 8x shallow speedup. The deep speedup vs CraterLake is explicitly attributed to cache size (Section 6.3), and vs F1+ to F1+'s unoptimized bootstrapping; neither is a novel architectural contribution. The shallow speedup is the evidence for the paper's central thesis that heterogeneity is needed for mixed workloads. That evidence is confounded: the comparison is FLASH-FHE with a multi-job scheduler against CraterLake with a single-job scheduler. Since CraterLake has 8 computation groups, the relevant counterfactual is CraterLake with a multi-job scheduler. The paper never tests this counterfactual nor provides an architectural argument that it is impossible. This is not a matter of consensus vs. the paper; it is a missing control in the evaluation of the causal claim. The concrete test I propose would settle it. I do not find an internal inconsistency in the architecture itself that would make the design non-functional, and the RTL synthesis details are a point in the paper's favor even without artifacts. The verdict remains CONDITIONAL: the architecture is plausible, but the headline performance attribution needs this test before it can be taken at face value.","tokens_in":22481,"tokens_out":4792,"duration_ms":44429,"concrete_test":"Model CraterLake with a batch scheduler: modify the cycle-accurate simulator used for Figure 12 (or re-implement CraterLake's architecture from [41]) so that up to 8 independent shallow workloads are dispatched to the 8 computation groups, and measure the average completion time for the 1-10 workload mixes in Figure 12. Compare against (a) the paper's reported CraterLake sequential results and (b) FLASH-FHE's parallel results. If the multi-job CraterLake achieves a speedup close to 8x over its sequential baseline, the 8x claim is a scheduling artifact and FLASH-FHE's heterogeneity is not the cause. A cheaper analytical check: verify whether CraterLake's computation groups have independent instruction fetch and per-group memory ports; if they do, a multi-job scheduler is feasible without hardware changes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline 8x shallow-workload speedup is the paper's central differentiator, and it is attributed to heterogeneity in the title and abstract ('delivering up to 8.0x speedup for shallow workloads due to its heterogeneous architecture'). But the comparison baseline CraterLake is a homogeneous accelerator with eight computation groups, each containing 256-lane NTT engines and local register files. The paper's argument that such an accelerator 'can only handle one job simultaneously' (Section 3.3) and 'can only use a sequential scheduling policy due to its homogeneous design' (Section 6.3) is asserted, not demonstrated. Nothing in the published CraterLake architecture prevents a multi-job scheduler from assigning one shallow workload per computation group; indeed, the paper's own Figure 2 shows that cooperating clusters beyond four gives diminishing returns for a single shallow job, which is exactly why a batch scheduler is the natural fix. If such a scheduler were added to CraterLake, it could plausibly run up to eight shallow jobs concurrently and capture most of the 8x throughput gain. The paper also acknowledges that its deep-workload advantage over CraterLake (1.4x) comes from a larger 320MB cache, not from heterogeneity. Thus the unique claim that heterogeneity, rather than scheduling policy, drives the headline result rests on an unproven premise. Without an ablation that keeps the scheduler fixed and varies only the hardware, or an analysis showing CraterLake's memory system and instruction dispatch cannot support multi-job execution, the central causal claim is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"FLASH-FHE proposes a heterogeneous FHE accelerator that combines eight 'bootstrappable' clusters (optimized for deep workloads with 2^16-point NTT and BConv pipelines) and sixteen 'swift' clusters (optimized for shallow workloads with 2^14-point NTT pipelines), organized into eight cluster affiliations. A scheduling scheme assigns one shallow workload per affiliation and decomposes the bootstrappable cluster into multiple small-point NTT pipelines, while deep workloads use all bootstrappable clusters jointly. The paper reports an RTL implementation synthesized at 7nm and 14/12nm, a cycle-accurate simulator evaluation on seven workloads, a 1.4x average deep-workload speedup over CraterLake, an 11.2x speedup over F1+, and up to 8.0x shallow-workload speedup attributed to the heterogeneous architecture, with less than 7% extra chip area for the swift engines.","tokens_in":22831,"tokens_out":5017,"duration_ms":47793,"significance":"If the architectural claims are substantiated, this is a useful and timely contribution: it is among the first FHE accelerator designs to explicitly target mixed shallow/deep workloads, and it provides a concrete heterogeneous organization with full RTL, synthesis results at two technology nodes, area/power breakdowns, and a stated path for integrating the ideas into existing accelerators such as CraterLake, ARK, and SHARP. The paper also names its limitations honestly, including the lack of support for combining swift clusters for deep workloads (Section 7) and the disclosure that the deep-workload advantage over CraterLake comes mainly from the larger 320MB cache rather than from heterogeneity (Section 6.3). The main unresolved issue is whether the headline 8x shallow-workload speedup is genuinely caused by the heterogeneous design or is instead a batch-parallelism effect that a homogeneous multi-group accelerator could also achieve under a different scheduler.","major_comments":[{"comment":"The central claim that the up-to-8.0x shallow-workload speedup is 'due to its heterogeneous architecture' is not yet established. The paper asserts in §3.3 that existing accelerators 'can only handle one job simultaneously' and in §6.3 that CraterLake 'can only use a sequential scheduling policy due to its homogeneous design,' but no architectural or software-stack analysis is provided to rule out a multi-job scheduler for CraterLake. Since CraterLake already contains eight computation groups with independent 256-lane NTT engines and local register files, a batch scheduler assigning one shallow job per group could plausibly capture most of the same throughput gain; in that case the 8x figure reflects parallelism available to any multi-group accelerator, not the heterogeneous design. Please add a fair CraterLake scheduler to the simulation, or provide a structural proof that CraterLake's hardware and software stack cannot support multi-job execution.","section":"§3.3, §6.3 (Fig. 12)"},{"comment":"The key mechanism of dynamically decomposing a 2^16-point (i)NTT pipeline into multiple smaller pipelines via multi-entry/multi-exit ports is assumed to be throughput-preserving with negligible overhead, but no microbenchmark or area/latency analysis isolates this effect. The shallow-workload speedup depends directly on the throughput of a bootstrappable cluster in decomposed mode, so the paper should report the effective NTT throughput (or lack of throughput loss) for the decomposed configuration and quantify the added mux/delay logic in the synthesized area.","section":"§4.1 (Fig. 4), §4.2"},{"comment":"The single-workload shallow comparison is not apples-to-apples: FLASH-FHE uses L=6 for the LoLa-MNIST variants while CraterLake and F1+ use L=4, and the paper acknowledges that FLASH-FHE is slower in these settings. The subsequent multi-workload claim should therefore be stated as a comparison of batch throughput under a fixed (larger) parameter set, not as a general shallow-workload speedup; otherwise the headline 8x result conflates parameter choice with architectural benefit.","section":"§6.3 (Fig. 9)"}],"minor_comments":[{"comment":"The abstract and introduction refer to 'F1' while the evaluation compares against 'F1+'; please use consistent naming throughout.","section":"Abstract, §1, §6.1"},{"comment":"The title contains an unexplained typographical artifact '/b♀t'; if this is intended to evoke big.LITTLE, please spell it out in words.","section":"Title"},{"comment":"The phrase 'speedup of 0.4–2.2x' is a slowdown at the lower end; please describe these results as relative performance ratios rather than speedups when the ratio is below 1.","section":"§6.3"},{"comment":"The statement that the swift clusters consume less than 7% of chip area refers ahead to Table 3; please cite the table directly and clarify whether this is logic-only area, since the swift clusters share L1/L2 memory.","section":"§4.2, Table 3"},{"comment":"The evaluation methodology would be strengthened by a short description of the cycle-accurate simulator's memory model, including HBM bandwidth and memory-controller contention, since the performance comparisons rely on it.","section":"§6.1"},{"comment":"Figure 2 lacks a caption in the manuscript text; please add one that clearly states that diminishing returns set in above four cooperating computation clusters, since this observation drives the motivation for parallel shallow-workload scheduling.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"This is a well-engineered architecture paper with full RTL synthesis and a transparent evaluation, and the heterogeneity idea is likely to be of interest to the FHE architecture community. The main weakness is the unsupported assumption that a homogeneous accelerator such as CraterLake cannot run multiple jobs concurrently; this is fixable with a fair baseline scheduler or a structural analysis, but it is load-bearing for the headline claim and should be addressed before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"FLASH-FHE is a real architectural contribution, not a repackaging: heterogeneous computation clusters for FHE, a multi-exit (i)NTT pipeline that splits a 2^16-point engine into smaller parallel engines, and a multi-job scheduler are all new to the FHE accelerator literature. The authors also did the implementation work: full RTL, synthesis at 7nm and 14/12nm, and an area breakdown showing the swift clusters cost less than 7% of total chip area. They are honest about one thing: the deep-workload 1.4x speedup over CraterLake is mostly a cache effect (320MB vs 256MB), not a heterogeneity effect. That claim is fine.\n\nThe soft spot is the headline shallow-workload claim. The paper says CraterLake can only handle one job at a time and can only use sequential scheduling due to its homogeneous design, but it never shows that CraterLake's instruction dispatch or memory system actually prevents multi-job execution. CraterLake has eight 256-lane computation groups. Adding a multi-job scheduler that assigns one shallow job per group could plausibly capture most of the 8x throughput gain. Without an ablation that varies only the hardware and fixes the scheduler, or an analysis of CraterLake's dispatch, the paper does not establish that heterogeneity, rather than scheduling policy, produces the 8x. This is a core gap because the title explicitly attributes the speedup to heterogeneous architecture.\n\nThere is also an internal inconsistency: Section 3.2 says Logistic Regression cannot run with N < 2^15, but Section 6.1 sets N = 2^13 for the same workload. Minor, but it makes me want to recheck the reported numbers. No artifacts are released, so the cycle-accurate simulator results are not reproducible.\n\nIf the authors add an ablation or a careful argument about CraterLake's multi-job limits, this becomes a strong architecture paper. The multi-exit NTT pipeline alone is worth a serious look. This deserves peer review, but the evaluation needs major revision before the performance claims should be taken at face value.","headline":"A genuinely new heterogeneous FHE architecture with a clever multi-exit NTT design, but the 8x shallow speedup is not convincingly attributed to heterogeneity because the baseline's multi-job capability is never analyzed.","tokens_in":23374,"tokens_out":4589,"would_cite":true,"duration_ms":40676,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"FLASH-FHE is presented as the first FHE accelerator with a heterogeneous architecture, pairing bootstrappable clusters for deep workloads with swift clusters that run shallow workloads in parallel.","keywords":["FHE accelerator","heterogeneous architecture","bootstrapping","NTT pipeline","multi-job scheduling","shallow workloads","deep workloads","hardware accelerator"],"falsifier":"If a multi-job scheduler added to a homogeneous accelerator with eight bootstrapping computation groups yields nearly the same shallow-workload throughput as FLASH-FHE on the same benchmarks at the same technology node, then the central heterogeneity claim is falsified.","tokens_in":22282,"feed_emoji":"⚡","tokens_out":11500,"duration_ms":98255,"temperature":0.7,"pith_summary":"Real fully homomorphic encryption (FHE) workloads mix shallow computations, which can run without bootstrapping, and deep computations, which need the expensive bootstrapping refresh; existing accelerators optimize for only one of the two. FLASH-FHE's claim is that a single heterogeneous chip can serve both: it pairs large bootstrappable clusters with smaller swift clusters, and schedules deep jobs across all clusters while running up to eight shallow jobs in parallel. The paper reports average speedups of 1.4x over CraterLake and 11.2x over F1+ on deep workloads, and up to 8.0x on shallow workloads, with the swift engines costing under 7% extra chip area. If the design holds up, FHE acceleration no longer has to force a choice between deep capacity and shallow throughput.","feed_headline":"Heterogeneous FHE chip runs shallow jobs up to 8x faster","feed_subtitle":"Small fast-NTT clusters and bootstrapping engines share one chip, speeding deep and shallow encrypted jobs together.","key_machinery":"The central mechanism is the cluster affiliation: one bootstrappable cluster paired with two swift clusters, repeated eight times. The bootstrappable cluster is a pipelined NTT/iNTT-BConv-NTT engine for $N=2^{16}$; the swift clusters are $N=2^{14}$ NTT engines. Multi-level transpose hardware (L1/L2/L3) and explicit NTT entrance and exit points let a single large NTT circuit be split into multiple small parallel NTTs, so the same physical datapath accelerates both deep and shallow workloads. The scheduler then either distributes one deep job over all eight bootstrappable clusters or assigns one shallow job to each affiliation, while shared L1/L2 caches keep the shallow engines from duplicating memory.","core_discovery":"The paper's central claim is that heterogeneity, not more homogeneous compute, is the missing ingredient for mixed FHE workloads. The chip is organized into eight cluster affiliations, each with one bootstrappable cluster built around a $2^{16}$-point NTT-to-BConv-to-NTT pipeline and two swift clusters with $2^{14}$-point NTT pipelines. Deep workloads engage all bootstrappable clusters; shallow workloads are assigned one per affiliation, and the bootstrappable NTT is dynamically decomposed into smaller parallel NTT pipelines to add throughput. A multi-level transpose and shared hierarchical cache keep the added parallelism cheap. As a result the paper reports 1.4x and 11.2x average deep-workload gains over CraterLake and F1+, and up to 8.0x shallow-workload gains, at under 7% extra area for the swift cluster logic.","pith_inferences":["The one-bootstrappable-to-two-swift ratio is dictated by NTT circuit geometry (a $2^8$ versus $2^7$ radix), not by measured deep-to-shallow workload mix; a deployment dominated by shallow jobs might want a different ratio, a tradeoff the paper does not explore.","The 8x shallow speedup fundamentally depends on having at least several shallow jobs in flight at once; under sparse or bursty job arrivals, queueing and preemption policy, not peak throughput, will determine the realized gain.","The same multi-exit NTT decomposition idea could be used to time-share a single large NTT datapath across job sizes even in a homogeneous chip, suggesting part of the benefit may be obtainable without the second cluster type."],"forward_implications":["Deep FHE workloads gain an average 1.4x over CraterLake and 11.2x over F1+ across packed bootstrapping, ResNet-20, logistic regression, and LSTM.","Shallow FHE workloads can be processed up to eight in parallel, reaching up to 8.0x throughput when enough independent shallow jobs are present.","The added swift cluster logic costs less than 7% of total chip area, since the swift and bootstrappable clusters share on-chip L1/L2 caches.","At the 7nm node FLASH-FHE delivers 1.49x to 1.78x better performance per area than ARK on the two compared deep workloads, while SHARP's short-word optimization remains faster in absolute terms.","The heterogeneous organization can be layered onto existing deep-workload accelerators by adding multi-exit NTT pipelines, a multi-level transpose, and the multi-job scheduler."],"supporting_citations":[{"why":"Supplies the CraterLake baseline, its homogeneous single-job scheduling behavior, and the F1+ extension used for the deep-workload comparison.","marker":"[41]"},{"why":"Supplies the original F1 accelerator whose parameter choices define the shallow-optimized baseline.","marker":"[40]"},{"why":"Provides the iNTT-to-BConv-to-NTT bootstrapping pipeline structure that the bootstrappable clusters adopt.","marker":"[30]"},{"why":"Supplies the ARK 7nm accelerator baseline with the runtime data-generation and key-reuse co-design used for deep-workload comparisons.","marker":"[29]"},{"why":"Supplies the SHARP short-word accelerator baseline and its chip-area and performance numbers at 7nm.","marker":"[28]"},{"why":"Establishes that bootstrapping accounts for over 80% of deep FHE runtime, motivating a dedicated deep-optimized cluster type.","marker":"[47]"},{"why":"Provides the encrypted logistic-regression training workload used as one of the four deep benchmarks.","marker":"[24]"},{"why":"Provides the encrypted LSTM inference workload used as one of the four deep benchmarks.","marker":"[36]"},{"why":"Provides the LoLa low-latency private inference workloads used as the shallow benchmarks.","marker":"[11]"},{"why":"Provides the ResNet-20 encrypted inference workload used as one of the four deep benchmarks.","marker":"[33]"}],"fun_headline_variants":["Heterogeneous FHE chip speeds deep jobs 11x, shallow 8x","First heterogeneous FHE accelerator: 1.4x/11.2x deep, 8x shallow","For mixed FHE jobs, heterogeneity beats homogeneity: 1.4x, 11.2x, 8x","Heterogeneous FHE accelerator: 8x faster shallow, 11x deep","Heterogeneous design makes FHE fast for both deep and shallow workloads"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a homogeneous accelerator like CraterLake cannot be given a multi-job scheduling policy and therefore must process shallow workloads sequentially, so the up-to-8x shallow gain is really caused by heterogeneous hardware rather than by a scheduler that could have been added to the existing homogeneous design.","fun_headline_variants_meta":{"raw":{"variants":["Heterogeneous FHE chip speeds deep jobs 11x, shallow 8x","First heterogeneous FHE accelerator: 1.4x/11.2x deep, 8x shallow","For mixed FHE jobs, heterogeneity beats homogeneity: 1.4x, 11.2x, 8x","Heterogeneous FHE accelerator: 8x faster shallow, 11x deep","Heterogeneous design makes FHE fast for both deep and shallow workloads"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000669,"raw_usage":{"total_tokens":3084,"prompt_tokens":1014,"completion_tokens":2070,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":630,"completion_tokens_details":{"reasoning_tokens":1949}},"tokens_in":630,"tokens_out":2070,"duration_ms":14748,"temperature":1.0,"reasoning_tokens":1949,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T23:44:34.773376+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"If a multi-job scheduler added to a homogeneous accelerator with eight bootstrapping computation groups yields nearly the same shallow-workload throughput as FLASH-FHE on the same benchmarks at the same technology node, then the central heterogeneity claim is falsified.","supporting_citations":[{"cited_title":"Craterlake: a hardware accelerator for efficient un- bounded computation on encrypted data","cited_arxiv_id":null,"evidence_quote":"Supplies the CraterLake baseline, its homogeneous single-job scheduling behavior, and the F1+ extension used for the deep-workload comparison."},{"cited_title":"Dreslinski, Christopher Peikert, and Daniel Sánchez","cited_arxiv_id":null,"evidence_quote":"Supplies the original F1 accelerator whose parameter choices define the shallow-optimized baseline."},{"cited_title":"BTS: an accelerator for bootstrappable fully homo- morphic encryption","cited_arxiv_id":null,"evidence_quote":"Provides the iNTT-to-BConv-to-NTT bootstrapping pipeline structure that the bootstrappable clusters adopt."},{"cited_title":"ARK: fully homomorphic encryption accelerator with runtime data generation and inter-operation key reuse","cited_arxiv_id":null,"evidence_quote":"Supplies the ARK 7nm accelerator baseline with the runtime data-generation and key-reuse co-design used for deep-workload comparisons."},{"cited_title":"SHARP: a short-word hierarchical accelerator for robust and prac- tical fully homomorphic encryption","cited_arxiv_id":null,"evidence_quote":"Supplies the SHARP short-word accelerator baseline and its chip-area and performance numbers at 7nm."},{"cited_title":"Sok: Fully homomor- phic encryption accelerators","cited_arxiv_id":null,"evidence_quote":"Establishes that bootstrapping accounts for over 80% of deep FHE runtime, motivating a dedicated deep-optimized cluster type."},{"cited_title":"Logistic regression on homomorphic encrypted data at scale","cited_arxiv_id":null,"evidence_quote":"Provides the encrypted logistic-regression training workload used as one of the four deep benchmarks."},{"cited_title":"Classification of encrypted word embeddings using recurrent neural networks","cited_arxiv_id":null,"evidence_quote":"Provides the encrypted LSTM inference workload used as one of the four deep benchmarks."},{"cited_title":"Low latency privacy preserving inference","cited_arxiv_id":null,"evidence_quote":"Provides the LoLa low-latency private inference workloads used as the shallow benchmarks."},{"cited_title":"Privacy-preserving machine learn- ing with fully homomorphic encryption for deep neural network","cited_arxiv_id":null,"evidence_quote":"Provides the ResNet-20 encrypted inference workload used as one of the four deep benchmarks."}],"review_version":1}