753d4d4

Adam's Bridge Hardware Specification
Version 3.0
Scope
This document defines technical specifications for a Adam's Bridge Post-Quantum Cryptography (PQC ML-DSA and ML-KEM) subsystem used in the Open Compute Project (OCP). This document shall comprise the Adam's Bridge technical specification.
Overview
This document provides definitions and requirements for a Adam's Bridge Post-Quantum Cryptography (PQC ML-DSA and ML-KEM) subsystem. The document then relates these definitions to existing technologies, enabling device and platform vendors to better understand those technologies in trusted computing terms.
Introduction
The advent of quantum computers poses a serious challenge to the security of cloud infrastructures and services, as they can potentially break the existing public-key cryptosystems, such as RSA and elliptic curve cryptography (ECC). Even though the gap between today’s quantum computers and the threats they pose to current public-key cryptography is large, the cloud landscape should act proactively and initiate the transition to the post-quantum era as early as possible. To comply with that, the U.S. government issued a National Security Memorandum in May 2022 that mandated federal agencies to migrate to PQC by 2035 [1].
The long-term security of cloud computing against quantum attacks depends on developing lattice-based cryptosystems, which are among the most promising PQC algorithms that are believed to be hard for both classical and quantum computers. The American National Institute of Standards and Technology (NIST) recognized this and selected CRYSTALS-KYBER (ML-KEM) [2] and CRYSTALS-Dilithium (ML-DSA) [3], two lattice-based algorithms, as standards for post-quantum key-establishment and digital signatures, respectively, in July 2022. These cryptosystems are constructed on the hardness of the module learning-with-errors problem (M-LWE) in module lattices.
To transition to PQC, we must develop hybrid cryptosystems to maintain industry or government regulations, while PQC updates will be applied thoroughly. Therefore, classical cryptosystems, e.g. ECC, cannot be eliminated even if PQC will significantly be developed.
Adam’s bridge was a mythological structure that existed to cross the formidable gulf that existed between two land masses. Asymmetric cryptography to post quantum is a similar formidable gap that exists in the world of cryptography and Adam’s bridge is the work undertaken to bridge the gap by building post quantum cryptographic accelerators.
This document shares the architectural characteristics of the proposed post-quantum Adams Bridge implementation. The proposed work divides the operations in the algorithms into multiple stages and executes them using pipelined processing architecture. An optimized cascading method is used within each stage and fine-tune each module individually to exploit multi-levels of parallelism to accelerate post-quantum Dilithium computation on hardware platforms to address performance and complexity challenges of PQC implementation. The proposed architecture uses various optimization techniques, including multi-levels of parallelism, designing reconfigurable cores, and implementing interleaved and pipelined architecture achieving significant speedup while maintaining high security and scalability. This work can facilitate the adoption and deployment of PQC in cloud computing and enhance the security and efficiency of cloud services and applications in the post-quantum era.
Documentation
For the Adams Bridge V2.0 architecture and evaluation, see "Adams Bridge (V2.0) Accelerator: Bridging the Post-Quantum Transition" [4].
The project contains comprehensive documentation of all submodules for ML-DSA and ML-KEM:
Memory requirement
The following table shows the required memory instances for Adam's Bridge:
| Instance | Depth | Data Width | Strobe Width |
|---|---|---|---|
| abr_sk_mem_bank0 | 596 | 32 | |
| abr_sk_mem_bank1 | 596 | 32 | |
| abr_w1_mem | 512 | 4 | |
| abr_mem_inst0_bank0 | 800 | 96 | |
| abr_mem_inst0_bank1 | 800 | 96 | |
| abr_mem_inst1 | 64 | 96 | |
| abr_mem_inst2 | 1536 | 96 | |
| abr_sig_z_mem | 224 | 160 | 20 |
| abr_pk_mem | 64 | 320 | 40 |
All memories are modeled as 1 read 1 write port RAMs with a flopped read data.
See abr_1r1w_ram.sv and abr_1r1w_be_ram.sv for examples.
The abr_top parameter SRAM_LATENCY (default 1) sets the read latency in clock cycles; the reference wrapper abr_mem_top.sv adds SRAM_LATENCY - 1 read-data pipeline stages to model it.
Strobe width is the number of byte-enable strobe signals (data width / 8); each strobe enables 8 bits. All strobed memories are byte enabled in the design.
Masking-protected memory (MASKING_EN = 1)
When the top-level MASKING_EN parameter is set to 1, four additional SRAM instances are generated to hold the second share of all masked operands. These mirror the dimensions of the corresponding regular memory instances:
| Instance | Depth | Data Width | Strobe Width |
|---|---|---|---|
| abr_mem_inst0_bank0_masked | 800 | 96 | |
| abr_mem_inst0_bank1_masked | 800 | 96 | |
| abr_mem_inst1_masked | 64 | 96 | |
| abr_mem_inst2_masked | 1536 | 96 |
Regular and masked SRAMs together carry the two-share representation of each coefficient: the regular memory holds share0 (uniform random) and the masked memory holds share1 where share0 + share1 = data. When MASKING_EN = 0 the masked instances are not instantiated and their interface read data is tied to zero.
Zeroize
The ZEROIZE bit in the control register clears all internal data-path registers that hold or have held secret-derived values. This prevents residual sensitive data from leaking through side-channel analysis or from being consumed by a subsequent operation.
Firmware must issue zeroize:
- After every completed operation, before starting the next command.
- After any error or aborted operation, before re-issuing a command or reading results.
Hardware behavior:
- Zeroize takes the highest priority after reset.
- The ZEROIZE control-register field is a pulse command: firmware writes a 1; the register auto-clears after one clock cycle. However, the hardware zeroize operation itself takes multiple cycles because on-chip SRAM is scrubbed one address per cycle. Firmware must poll the STATUS register and wait for ready before issuing the next command.
- After zeroize completes, the STATUS register returns to ready, and the engine accepts a new command.
Area Results
The following table shows the abr_top area of Adams Bridge (ML-DSA-87 + ML-KEM-1024) in gate count, split into standard-cell logic and SRAM. ABR 2.x is included for comparison.
| Configuration | Std cell [GE] | SRAM [GE] | Total [GE] |
|---|---|---|---|
| ABR 2.x | 2,389,140 | 587,724 | 2,976,864 |
ABR 3.0, MASKING_EN = 0, SHA3_MASKING_EN = 0 | 428,497 | 481,105 | 909,602 |
ABR 3.0, MASKING_EN = 1, SHA3_MASKING_EN = 0 | 554,624 | 749,814 | 1,304,438 |
ABR 3.0, MASKING_EN = 1, SHA3_MASKING_EN = 1 (default) | 648,230 | 749,814 | 1,398,044 |
MASKING_EN = 1adds the four_maskedSRAM instances (see Memory requirement) and the second NTT engine.SHA3_MASKING_EN = 1adds the two-share masked Keccak; it does not change the SRAM requirement.
Physical area at 5nm is available for the core-masking configuration (MASKING_EN = 1, SHA3_MASKING_EN = 0): abr_top standard-cell logic occupies 12,387 µm² and the SRAMs occupy an additional 33,500 µm², for a total Adams Bridge area of approximately 45,900 µm² (0.0459 mm²).
For per-algorithm performance breakdowns, see the Performance and Area Results sections in AdamsBridge_MLDSA.md and AdamsBridge_MLKEM.md.
Technology-Specific Primitive Instantiation
Adams Bridge uses redundant logic to counteract Fault Injection (FI) and Side Channel Analysis (SCA). In particular, the Domain-Oriented Masking (DOM) datapath relies on primitive cells (buffers, enabled flops, XOR gates) whose logic must not be optimized, merged, or reordered by synthesis, otherwise the masking guarantees are broken. This logic is built from generic primitive modules that act as placeholders, instantiated only through abstract prim wrappers (abr_prim_buf, abr_prim_flop, abr_prim_flop_en, abr_prim_xor2) so that no design or wrapper RTL needs to be edited to retarget a technology.
For the environment variables, build-time macros, and worked examples needed to replace these generic primitives with technology-specific cells (including how this interacts with a Caliptra-integrated build), see AdamsBridgeIntegrationSpecification.md.
References:
[1] The White House, "National Security Memorandum on Promoting United States Leadership in Quantum Computing While Mitigating Risks to Vulnerable Cryptographic Systems," 2022. [Online]. Available: White House.
[2] NIST, "FIPS 203 Module-Lattice-Based Key-Encapsulation Mechanism Standard," August 13, 2024.
[3] NIST, "FIPS 204 Module-Lattice-Based Digital Signature Standard," August 13, 2024.
[4] Mojtaba Bisheh-Niasar, Emre Karabulut, Kiran Upadhyayula, Michael Norris, and Bharat Pillilli, "Adams Bridge (V2.0) Accelerator: Bridging the Post-Quantum Transition," IACR Cryptology ePrint Archive, Report 2026/256, 2026. [Online]. Available: https://eprint.iacr.org/2026/256.