NVIDIA CUDA 13.3 Brings Hardware-Accelerated Cryptography

NVIDIA has introduced hardware-accelerated carryless multiplication with the release of CUDA 13.3, a new feature aimed at enhancing cryptographic performance on Ampere and newer GPUs. This long-awaited addition allows developers to utilize the clmad instruction, a parallel thread execution (PTX) feature, to perform high-speed binary extension field multiplications. Previously, developers had to rely on slower, software-based methods such as bitsliced circuits to achieve similar functionality.

Carryless multiplication is a cornerstone operation for modern cryptography, underpinning protocols like GHASH in AES-GCM encryption (used in TLS and VPNs) and advanced zero-knowledge (ZK) proving systems. NVIDIA’s benchmarks demonstrate just how transformative this hardware support can be. On the NVIDIA B200 GPU, GHASH throughput has reached an astonishing ~6.3 TB/s—a nearly 19x improvement over prior methods—while ZK proving systems see speedups of up to 13x.

Hardware support for carryless multiplication has been standard in x86 CPUs since Intel introduced the PCLMULQDQ instruction in 2010. NVIDIA’s move to bring similar capabilities to their GPUs aligns with the increasing computational demands of cryptography-heavy workloads, such as data center encryption and blockchain applications.

Why This Matters

Cryptographic operations like AES-GCM are critical to securing internet communications and storage systems, while ZK proofs are becoming a linchpin for privacy-preserving blockchain protocols. Carryless multiplication in hardware drastically reduces the computational overhead, making these technologies more scalable and cost-effective. For example, the GHASH operation in AES-GCM—previously a bottleneck—can now process data at speeds closer to DRAM bandwidth limits.

Furthermore, the inclusion of clmad widens the range of cryptographic workloads that can run efficiently on NVIDIA GPUs. These include error-correcting codes like Reed–Solomon and BCH, post-quantum cryptographic schemes, and even quantum stabilizer codes. This positions Ampere-based GPUs as a versatile choice for industries ranging from telecommunications to blockchain development.

Performance Benchmarks

The impact of CUDA 13.3’s carryless multiplication is clear in NVIDIA’s tests:

• GHASH Throughput: ~6.3 TB/s on the B200 GPU, a nearly 19x improvement over the previous bitslicing methods.
• Zero-Knowledge Proofs: Up to 13x speedups in sum-check protocol performance, crucial for scaling blockchain and privacy applications.
• Real-World Applications: On an NVIDIA GeForce RTX 5090, peak GHASH throughput of ~1,300 GB/s demonstrates high performance even on consumer-grade GPUs.

Industry Context

As cryptographic workloads grow more demanding, hardware acceleration is becoming a competitive differentiator. Intel and AMD have long supported carryless multiplication in x86 processors, with Intel recently reaffirming its commitment to cryptography-friendly extensions in its upcoming Nova Lake CPUs. NVIDIA’s move to bring similar capabilities to GPUs is a logical next step, especially as GPUs increasingly handle workloads beyond graphics, such as AI and blockchain computation.

The introduction of CUDA 13.3 puts NVIDIA in a stronger position against competitors like AMD, which has also been enhancing its GPU compute capabilities. For developers and enterprises already utilizing NVIDIA’s hardware, this update represents a significant opportunity to optimize cryptographic pipelines without additional hardware investment.

Looking Ahead

With CUDA 13.3 now available for download, developers can start integrating clmad into their cryptographic workloads immediately. The implications are significant for industries relying on high-throughput encryption, ZK proofs, and error-correcting codes. As cryptographic demands continue to rise, NVIDIA’s hardware acceleration could help redefine the cost and performance dynamics of secure computing.

The race to optimize cryptographic performance is far from over. With competitors like Intel continuing to push innovations in this space, the market for high-performance, secure computing hardware is only set to heat up.

Image source: Shutterstock