Parallel implementation of GCM on GPUs
Open accessAbstract
This paper presents the first fully parallelized optimization of GCM in a GPU environment. As the era of IoT emerges, a large number of clients communicate with servers, necessitating encrypted communications for security. GCM is a type of AEAD and is currently used in various security protocols , including TLS 1.3 and IPsec . Due to the burden of performing encrypted communication with numerous clients, there has been significant research on utilizing GPUs for high-speed parallel processing in encryption. However, to date, there has been no fully parallelized implementation of GCM on GPUs. This paper proposes a method for parallelizing the challenging GHASH computation in GCM mode, leading to a high-speed parallel implementation of AES-GCM that can exceed 400Gb/s, meeting the requirements of next-generation communication systems . The proposed approach is algorithm-independent and can be applied to any block ciphers . Our implementation on an RTX 4090 demonstrates a performance improvement of × 15 . 38 compared to the maximum processing throughput of a multi-threaded Intel(R) Core(TM) i7-13700K. It also achieves a × 17 . 87 improvement compared to a hybrid CPU–GPU system. Compared to the most researched FPGA implementation for GCM, specifically Xilinx Ultrascale FPGA , our implementation achieves × 1 . 11 better performance. For not only throughput but also power efficiency also better than other implementation, it achieves × 3 . 33 compared to CPU implementation on Intel Xe…
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The standard GHASH algorithm is inherently sequential, making it ill-suited for the massive parallelism of GPUs.
“Consequently, the standard GHASH algorithm is inherently sequential, making it ill-suited for the massive parallelism of Graphics Processing Units (GPUs) (Manavski, 2007; Lee et al., 2025).”
From A high-throughput AES-GCM implementation on GPUs for secure, policy-based access to massive astronomical catalogs (2026) · cites this paper as Lee et al., 2025
Following the strategy described by Lee et al. (2025), we transform the linear O(N) GHASH computation into a scalable, logarithmic-time O(log N) process.
“Following the strategy described by Lee et al. (2025), we transform the linear O ( N ) GHASH computation into a scalable, logarithmictime O (log N ) process.”
From A high-throughput AES-GCM implementation on GPUs for secure, policy-based access to massive astronomical catalogs (2026) · cites this paper as Lee et al., 2025
The implementation utilizes a parallel reduction strategy operating in GF(2^128), inspired by recent optimizations for parallel architectures.
“To address this, the implementation utilizes a parallel reduction strategy operating in the Galois Field GF (2 128 ), inspired by recent optimizations for parallel architectures (Lee et al., 2025).”
From A high-throughput AES-GCM implementation on GPUs for secure, policy-based access to massive astronomical catalogs (2026) · cites this paper as Lee et al., 2025
On Ada Lovelace hardware, the L40S reaches 664 Gbps, exceeding the 536 Gbps reported for the RTX 4090 by approximately 24%.
“On Ada Lovelace hardware, the L40S reaches 664 Gbps, exceeding the 536 Gbps reported for the RTX 4090 by approximately 24%.”
From A high-throughput AES-GCM implementation on GPUs for secure, policy-based access to massive astronomical catalogs (2026) · cites this paper as Lee et al. (2025)
This is also the choice made by Lee et al. (2024) and Lee et al. (2025), whose implementation we adopt as our cryptographic core.
“This is also the choice made by Lee et al. (2024) and Lee et al. (2025), whose implementation we adopt as our cryptographic core.”
From A high-throughput AES-GCM implementation on GPUs for secure, policy-based access to massive astronomical catalogs (2026) · cites this paper as Lee et al. (2025)
On the RTX 3090, GpuFitsCrypt achieves 472 Gbps versus 247 Gbps reported by Lee et al., a factor of approximately 1.91× higher.
“On the RTX 3090, GpuFitsCrypt achieves 472 Gbps versus 247 Gbps reported by Lee et al., a factor of approximately 1.91 × higher.”
From A high-throughput AES-GCM implementation on GPUs for secure, policy-based access to massive astronomical catalogs (2026) · cites this paper as Lee et al. (2025)
Recent state-of-the-art benchmarks for fully parallelized AES-GCM on GPUs (Lee et al., 2025) demonstrate a speedup factor of ≈15× compared to highly optimized multi-threaded CPU implementations.
“Recent state-ofthe-art benchmarks for fully parallelized AES-GCM on GPUs (Lee et al., 2025) demonstrate a speedup factor of ≈ 15 × compared to highly optimized multi-threaded CPU implementations (Intel Core i7 with AES-NI instructions using OpenSSL).”
From A high-throughput AES-GCM implementation on GPUs for secure, policy-based access to massive astronomical catalogs (2026) · cites this paper as Lee et al., 2025
The most direct comparison is with Lee et al. (2025), who measure GHASH-only throughput using 16 KB pre-staged message blocks, a scope equivalent to our GHashKernelOnly measurement.
“The most direct comparison is with Lee et al. (2025), who measure GHASH-only throughput using 16 KB pre-staged message blocks, a scope equivalent to our GHashKernelOnly measurement.”
From A high-throughput AES-GCM implementation on GPUs for secure, policy-based access to massive astronomical catalogs (2026) · cites this paper as Lee et al. (2025)
The same study highlights a power efficiency advantage of over 20× for the GPU architecture.
“Furthermore, the same study highlights a power efficiency advantage of over 20 × for the GPU architecture.”
From A high-throughput AES-GCM implementation on GPUs for secure, policy-based access to massive astronomical catalogs (2026) · cites this paper as Lee et al., 2025
Cites
A fast quantum mechanical algorithm for database search
Grover, L. K. (1996) · Proceedings of the twenty-eighth annual ACM symposium on Theory of computing - STOC '96 · doi:10.1145/237814.237866 · Cited by 8,830 · see citing works
Flagged first, then by citations
Cited by
Sorted by citations
Secure and Efficient Block Cipher Mode Design for Parallel Processing and Reliable Security
Vatsavayi, V. K., & Bommireddy, D. R. (2026) · Cryptography · doi:10.3390/cryptography10010013 · Cited by 2 · see citing works
Lemes-Perera, S., R. Alarcon, M., Caballero-Gil, P., & Serra-Ricart, M. (2026) · Astronomy and Computing · doi:10.1016/j.ascom.2026.101153 · Cited by 1 · see citing works
- Authors
- Lee, J., Kim, D., & Seo, S. C.
- Affiliations
- Kookmin University (KR)
- Corresponding author
- Seo, Seog Chung
- Venue
- ICT Express
- Publisher
- Elsevier BV
- Published
- 2025-04
- Volume
- 11
- Issue
- 2
- Pages
- 310-316
- ISSN
- 2405-9595 (ISSN-L)
- OpenAlex
- W4407583012
- Cited by
- 7
- References
- 1
- Citations by year
- 2026: 5 · 2025: 2
- FWCI
- 2.83
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- Locations
- ICT Express (publishedVersion, cc-by-nc-nd)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- Funders
- National Research Foundation of Korea (2022R1C1C1013368)
- IITP (2021-0-00796)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Cryptographic Implementations and Security
- Physical Sciences › Computer Science › Information Systems › Cryptography and Residue Arithmetic
- Physical Sciences › Computer Science › Hardware and Architecture › Parallel Computing and Optimization Techniques
- Keywords
- AES-GCM, GPU acceleration, parallel encryption, authenticated encryption, high-throughput encryption, power efficiency, TLS 1.3, IPsec, FPGA comparison
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
