{"schema_version":"0.1","map_id":"paper-64-map","publication_id":64,"publication_anchor":"paper-64","slug":"paper-64","canonical_path":"/knowledge/papers/paper-64/","machine_path":"/knowledge/papers/paper-64.json","root_node_id":"paper-64","stage":"mapped_draft","contribution_type_vocabulary_version":"0.1","contribution_types":["algorithm"],"title":"CraterLake: A Hardware Accelerator for Efficient Unbounded Computation on Encrypted Data","short_title":"CraterLake","year":2022,"venue":"49th International Symposium on Computer Architecture (ISCA)","publication_status":"Published","topic":"secure-encrypted-computation","labels":["Applied"],"authors":["Nikola Samardzic","Axel Feldmann","Aleksandar Krastev","Nathan Manohar","Nicholas Genise","Srinivas Devadas","Karim Eldefrawy","Chris Peikert","Daniel Sánchez"],"keywords":["fully homomorphic encryption","FHE accelerator","bootstrapping","boosted key switching","hardware compiler co-design"],"research_question":"Can a programmable hardware accelerator remove the refresh bottleneck that had limited prior FHE accelerators to bounded-depth programs and efficiently execute deep encrypted workloads?","central_answer":"CraterLake combines boosted key switching, a 2,048-lane vector architecture, specialized residue and key-switch generation units, and a scheduling compiler. In cycle-accurate simulation and RTL synthesis it supports CKKS, BGV, and GSW computations with bootstrapping and substantially outperforms the evaluated CPU and prior-accelerator baselines on deep workloads; these are modeled and synthesized results, not measurements from fabricated silicon.","curation":{"drafted_at":"2026-07-11","drafted_by":[{"actor_type":"ai","name":"OpenAI Codex","role":"full-text extraction, claim-evidence mapping, and initial assessment"}],"method":"Source-grounded review of the complete 15-page author copy, including visual inspection of the title page and evaluation pages, cross-checked against the official DOI and public-access record.","source_scope":"full_source_audit","approval":{"status":"pending","note":"AI-authored source map awaiting author verification; architecture interpretations, benchmark readings, and ratings remain provisional until approval."}},"sources":[{"id":"source-paper-64-local-pdf","type":"author_hosted_copy","title":"CraterLake: A Hardware Accelerator for Efficient Unbounded Computation on Encrypted Data","url":"/pubs/2022/craterlake-isca2022.pdf","provenance_category":"author","retrieved_from":"https://people.csail.mit.edu/sanchez/papers/2022.craterlake.isca.pdf","media_type":"application/pdf","sha256":"3800eb135cdf91e1c86b959ecdf86df15cf72a8d64d97d8c820040cdbad9e5de","page_count":15},{"id":"source-paper-64-official","type":"official_publication_record","title":"ACM ISCA 2022 publication record","url":"https://doi.org/10.1145/3470496.3527393","provenance_category":"official"},{"id":"source-paper-64-public-access","type":"public_archive_copy","title":"NSF Public Access copy","url":"https://par.nsf.gov/servlets/purl/10350313","provenance_category":"archive"},{"id":"source-paper-64-citations","type":"citation_index_snapshot","title":"OpenAlex record W4281792301","url":"https://api.openalex.org/works/W4281792301","accessed_at":"2026-07-11"}],"source_anchors":[{"id":"anchor-paper-64-problem","source_id":"source-paper-64-local-pdf","label":"FHE depth barrier and CraterLake contributions","locator":"Abstract and Sections 1-2, PDF pages 1-4","url":"/pubs/2022/craterlake-isca2022.pdf#page=1"},{"id":"anchor-paper-64-fhe-model","source_id":"source-paper-64-local-pdf","label":"Supported schemes, ciphertext representation, and key switching","locator":"Section 2, PDF pages 2-4","url":"/pubs/2022/craterlake-isca2022.pdf#page=2"},{"id":"anchor-paper-64-boosted","source_id":"source-paper-64-local-pdf","label":"Boosted key-switching algorithm and data reduction","locator":"Section 3, PDF pages 4-6","url":"/pubs/2022/craterlake-isca2022.pdf#page=4"},{"id":"anchor-paper-64-architecture","source_id":"source-paper-64-local-pdf","label":"Vector organization, transpose network, residue conversion, and key-switch generation","locator":"Sections 4-5, PDF pages 6-9","url":"/pubs/2022/craterlake-isca2022.pdf#page=6"},{"id":"anchor-paper-64-compiler","source_id":"source-paper-64-local-pdf","label":"Compiler scheduling, chaining, and data movement","locator":"Section 6, PDF pages 9-10","url":"/pubs/2022/craterlake-isca2022.pdf#page=9"},{"id":"anchor-paper-64-implementation","source_id":"source-paper-64-local-pdf","label":"RTL synthesis, physical estimates, and simulation method","locator":"Sections 7-8, PDF pages 10-11","url":"/pubs/2022/craterlake-isca2022.pdf#page=10"},{"id":"anchor-paper-64-performance","source_id":"source-paper-64-local-pdf","label":"Deep-workload latency and baseline comparisons","locator":"Section 9 and Table 3, PDF pages 11-13","url":"/pubs/2022/craterlake-isca2022.pdf#page=11"},{"id":"anchor-paper-64-energy-security","source_id":"source-paper-64-local-pdf","label":"Energy and higher-security-parameter sensitivity","locator":"Section 9, PDF pages 12-13","url":"/pubs/2022/craterlake-isca2022.pdf#page=12"},{"id":"anchor-paper-64-conclusion","source_id":"source-paper-64-local-pdf","label":"Conclusions and stated scope","locator":"Section 11, PDF page 14","url":"/pubs/2022/craterlake-isca2022.pdf#page=14"},{"id":"anchor-paper-64-publication","source_id":"source-paper-64-official","label":"Official publication identity","locator":"ISCA 2022, DOI 10.1145/3470496.3527393","url":"https://doi.org/10.1145/3470496.3527393"},{"id":"anchor-paper-64-citation-count","source_id":"source-paper-64-citations","label":"Dated citation-count snapshot","locator":"OpenAlex cited_by_count was 177 when accessed 2026-07-11","url":"https://api.openalex.org/works/W4281792301"}],"nodes":[{"id":"paper-64","kind":"paper","parent_id":null,"order":1,"epistemic_status":"published","title":"CraterLake","summary":"A hardware-software co-design for bootstrappable, programmable FHE whose evaluated contribution is efficient execution of workloads that exceed a fixed multiplicative-depth budget.","source_anchor_ids":["anchor-paper-64-problem"]},{"id":"paper-64-question","kind":"question","parent_id":"paper-64","order":1,"epistemic_status":"research_question","title":"Research question","summary":"Can an FHE accelerator efficiently refresh very large ciphertexts and thereby support general computations of unbounded multiplicative depth?","source_anchor_ids":["anchor-paper-64-problem"]},{"id":"paper-64-answer","kind":"contribution","parent_id":"paper-64","order":2,"epistemic_status":"source_asserted","title":"Central answer","summary":"The design makes bootstrapping a first-class workload through a compact key-switch method, wide parallel datapath, on-chip specialization, and compiler-controlled reuse; the resulting simulated system executes deep CKKS workloads much faster than the evaluated baselines.","source_anchor_ids":["anchor-paper-64-boosted","anchor-paper-64-architecture","anchor-paper-64-performance"]},{"id":"paper-64-model","kind":"model","parent_id":"paper-64","order":3,"epistemic_status":"defined","title":"Cryptographic and workload model","summary":"The architecture targets ring-based FHE operations and supports CKKS, BGV, and GSW, with evaluation centered on CKKS. Deep parameter sets contain 64K ring elements and roughly 1,600-bit coefficients, making one ciphertext about 25 MB.","source_anchor_ids":["anchor-paper-64-fhe-model","anchor-paper-64-problem"]},{"id":"paper-64-model-security","kind":"assumption","parent_id":"paper-64-model","order":1,"epistemic_status":"parameter_dependent","title":"Security and correctness are parameter dependent","summary":"CraterLake accelerates cryptographic computations chosen by the application and compiler; semantic security, numerical precision, circuit correctness, and parameter adequacy remain properties of the selected FHE scheme and parameter set.","source_anchor_ids":["anchor-paper-64-fhe-model","anchor-paper-64-energy-security"]},{"id":"paper-64-design","kind":"system","parent_id":"paper-64","order":4,"epistemic_status":"constructed","title":"Hardware-software construction","summary":"The system integrates algorithmic key-switch compression, a wide vector processor, specialized functional units, memory organization, and static compiler scheduling.","source_anchor_ids":["anchor-paper-64-boosted","anchor-paper-64-architecture","anchor-paper-64-compiler"]},{"id":"paper-64-design-boosted","kind":"algorithm","parent_id":"paper-64-design","order":1,"epistemic_status":"constructed","title":"Boosted key switching","summary":"The boosted formulation reduces projected auxiliary key material for a deep parameter set from roughly 1.4 GB to about 50 MB, making reuse and movement feasible; KSHGen can generate half of a hint on demand and reduce resident hint storage to about 25 MB.","source_anchor_ids":["anchor-paper-64-boosted","anchor-paper-64-architecture"]},{"id":"paper-64-design-datapath","kind":"architecture","parent_id":"paper-64-design","order":2,"epistemic_status":"constructed","title":"Wide-vector datapath","summary":"A 2,048-lane uniprocessor, fixed transpose network, and specialized residue-conversion and key-switch-generation units expose the parallelism and data transforms used by large polynomial operations.","source_anchor_ids":["anchor-paper-64-architecture"]},{"id":"paper-64-design-compiler","kind":"method","parent_id":"paper-64-design","order":3,"epistemic_status":"implemented","title":"Compiler-directed scheduling","summary":"The compiler schedules operations, reuses data, overlaps or decouples movement from computation, and chains kernels so intermediate values need not repeatedly traverse costly memory paths.","source_anchor_ids":["anchor-paper-64-compiler"]},{"id":"paper-64-claims","kind":"claim_group","parent_id":"paper-64","order":5,"epistemic_status":"source_asserted","title":"Principal claims","summary":"The paper claims practical acceleration of unbounded-depth FHE and reports advantages in latency and energy over its evaluated CPU and F1+ baselines.","source_anchor_ids":["anchor-paper-64-performance","anchor-paper-64-energy-security"]},{"id":"paper-64-claim-depth","kind":"claim","parent_id":"paper-64-claims","order":1,"epistemic_status":"supported_by_evaluation","title":"Refresh enables deep workloads","summary":"Unlike accelerators restricted to circuits within a fixed noise budget, the represented design includes bootstrapping and therefore runs the evaluated deep ResNet, LSTM, and other workloads without a bounded-depth architectural ceiling.","source_anchor_ids":["anchor-paper-64-problem","anchor-paper-64-performance"]},{"id":"paper-64-claim-speed","kind":"claim","parent_id":"paper-64-claims","order":2,"epistemic_status":"measured_in_model","title":"Large modeled speedups on deep workloads","summary":"Across the paper's deep benchmark set, CraterLake reports an 11.2x geometric-mean speedup over the scaled F1+ baseline and 4,611x over the CPU baseline; reported examples include 249.45 ms for ResNet-20 and 3.91 ms for packed bootstrapping.","source_anchor_ids":["anchor-paper-64-performance"]},{"id":"paper-64-claim-energy","kind":"claim","parent_id":"paper-64-claims","order":3,"epistemic_status":"estimated","title":"Energy advantage","summary":"The modeled deep-workload evaluation reports about 18x lower energy and about 201x higher performance per joule than scaled F1+.","source_anchor_ids":["anchor-paper-64-energy-security"]},{"id":"paper-64-evidence","kind":"evidence_group","parent_id":"paper-64","order":6,"epistemic_status":"documented","title":"Evaluation evidence","summary":"The evidence combines RTL synthesis and cycle-accurate architectural simulation with a 32-core AMD Threadripper CPU and a scaled improved F1+ design as baselines.","source_anchor_ids":["anchor-paper-64-implementation","anchor-paper-64-performance"]},{"id":"paper-64-evidence-scaling","kind":"evidence","parent_id":"paper-64-evidence","order":1,"epistemic_status":"sensitivity_analysis","title":"Security-parameter sensitivity","summary":"Relative to the base 80-bit parameter evaluation, the paper reports a 1.36x geometric-mean slowdown at 128-bit security and 2.60x at 200-bit security, the latter using a larger 128K-element hardware configuration.","source_anchor_ids":["anchor-paper-64-energy-security"]},{"id":"paper-64-boundaries","kind":"limitation_group","parent_id":"paper-64","order":7,"epistemic_status":"explicitly_bounded","title":"Scope and limitations","summary":"The results establish a simulated and synthesized architecture under selected cryptographic parameters and workloads, not universal FHE performance or deployment readiness.","source_anchor_ids":["anchor-paper-64-implementation","anchor-paper-64-performance","anchor-paper-64-conclusion"]},{"id":"paper-64-boundary-silicon","kind":"limitation","parent_id":"paper-64-boundaries","order":1,"epistemic_status":"evidence_boundary","title":"No fabricated-silicon measurement","summary":"Area, power, timing, energy, and workload latency are derived from synthesis and simulation rather than a fabricated chip, so manufacturing and deployment effects remain unevaluated.","source_anchor_ids":["anchor-paper-64-implementation"]},{"id":"paper-64-boundary-generalization","kind":"limitation","parent_id":"paper-64-boundaries","order":2,"epistemic_status":"benchmark_bounded","title":"Benchmark and baseline dependence","summary":"Speedups depend on the selected workloads, parameter sets, compiler schedules, and scaled comparison designs; shallow workloads sometimes narrow or reverse the advantage over F1+.","source_anchor_ids":["anchor-paper-64-performance","anchor-paper-64-energy-security"]},{"id":"paper-64-resources","kind":"artifact_group","parent_id":"paper-64","order":8,"epistemic_status":"source_available","title":"Auditable resources","summary":"The complete author copy is checked into this site with page count and SHA-256; official and NSF public-access records establish publication identity and an independent access route.","source_anchor_ids":["anchor-paper-64-problem","anchor-paper-64-publication"]},{"id":"paper-64-scrutiny","kind":"scrutiny","parent_id":"paper-64","order":9,"epistemic_status":"venue_reviewed","title":"External scrutiny","summary":"ISCA publication provides venue-level scrutiny. No public review reports, fabricated-hardware reproduction, correction, or independent benchmark replication was located in this audit.","source_anchor_ids":["anchor-paper-64-publication"]},{"id":"paper-64-lineage","kind":"lineage","parent_id":"paper-64","order":10,"epistemic_status":"documented","title":"Architectural lineage","summary":"The evaluation explicitly positions CraterLake against CPU FHE software and the prior F1/F1+ accelerator line, while changing the target from bounded-depth acceleration to refresh-capable deep computation.","source_anchor_ids":["anchor-paper-64-problem","anchor-paper-64-performance"]}],"relations":[{"id":"paper-64-rel-answer-question","type":"addresses","from_id":"paper-64-answer","to_id":"paper-64-question"},{"id":"paper-64-rel-design-answer","type":"realizes","from_id":"paper-64-design","to_id":"paper-64-answer"},{"id":"paper-64-rel-boosted-design","type":"component_of","from_id":"paper-64-design-boosted","to_id":"paper-64-design"},{"id":"paper-64-rel-datapath-design","type":"component_of","from_id":"paper-64-design-datapath","to_id":"paper-64-design"},{"id":"paper-64-rel-compiler-design","type":"component_of","from_id":"paper-64-design-compiler","to_id":"paper-64-design"},{"id":"paper-64-rel-evidence-speed","type":"supports","from_id":"paper-64-evidence","to_id":"paper-64-claim-speed"},{"id":"paper-64-rel-evidence-depth","type":"supports","from_id":"paper-64-evidence","to_id":"paper-64-claim-depth"},{"id":"paper-64-rel-evidence-energy","type":"supports","from_id":"paper-64-evidence","to_id":"paper-64-claim-energy"},{"id":"paper-64-rel-model-claims","type":"qualifies","from_id":"paper-64-model","to_id":"paper-64-claims"},{"id":"paper-64-rel-boundaries-claims","type":"qualifies","from_id":"paper-64-boundaries","to_id":"paper-64-claims"},{"id":"paper-64-rel-resources-evidence","type":"enables_audit_of","from_id":"paper-64-resources","to_id":"paper-64-evidence"},{"id":"paper-64-rel-lineage-paper","type":"contextualizes","from_id":"paper-64-lineage","to_id":"paper-64"}],"assessment":{"id":"paper-64-assessment-2026-07-11","rubric_version":"0.2","assessed_at":"2026-07-11","status":"ai_draft_author_review_pending","note":"These dimensions describe documented support and process, not truth, correctness, or a universal ranking. No composite score is calculated.","axes":[{"id":"epistemic_evidence","level":"high","rationale":"The full paper documents algorithms, architecture, synthesis and simulation methods, baselines, benchmark results, and parameter sensitivity. The evidence is strong for the evaluated model but is not fabricated-silicon validation.","basis_source_anchor_ids":["anchor-paper-64-boosted","anchor-paper-64-implementation","anchor-paper-64-performance"]},{"id":"auditability","level":"high","rationale":"A complete checked-in author copy with fixity and page count, plus official and public-archive records, makes assumptions, methods, and reported measurements directly inspectable.","basis_source_anchor_ids":["anchor-paper-64-problem","anchor-paper-64-publication"]},{"id":"production_provenance","level":"medium","rationale":"Named authorship, venue, DOI, manuscript identity, and public copies are documented; contributor roles, revision history, tool use, and artifact-version lineage are not.","basis_source_anchor_ids":["anchor-paper-64-problem","anchor-paper-64-publication"]},{"id":"external_scrutiny","level":"medium","rationale":"ISCA publication establishes venue review, but public reports, independent silicon reproduction, correction history, and adversarial re-evaluation were not located.","basis_source_anchor_ids":["anchor-paper-64-publication"]},{"id":"reception","level":"high","rationale":"The dated OpenAlex snapshot located 177 citations. Under the author-defined rule, 11 or more located citations is High; citation counts do not establish correctness.","basis_source_anchor_ids":["anchor-paper-64-citation-count"]},{"id":"contribution_significance","level":"high","rationale":"The source presents the first accelerator targeting efficient unbounded-depth FHE and demonstrates large deep-workload improvements against contemporary baselines; the rating remains an AI draft pending author review.","basis_source_anchor_ids":["anchor-paper-64-problem","anchor-paper-64-performance"]}]},"reception_snapshot":{"as_of":"2026-07-11","method":"OpenAlex work lookup by DOI 10.1145/3470496.3527393","citation_count":177,"source_url":"https://api.openalex.org/works/W4281792301","signals":["OpenAlex reported 177 citing works for the matched publication record."],"limitation":"Citation counts are index- and date-dependent and measure visibility rather than validity, implementation quality, or practical adoption."}}