NVIDIA CUDA 13.3 adds GPU instruction aimed at faster cryptography workloads

NVIDIA has used the CUDA 13.3 technical blog to highlight a small but important addition for developers building cryptography, coding-theory and proof-system workloads on GPUs: a new PTX instruction called clmad. The company says the carryless multiply-accumulate instruction is available on NVIDIA Ampere and newer GPUs, closing a gap with x86 CPUs, which have had dedicated carryless multiplication support for years.

Why the instruction matters

Carryless multiplication is a low-level arithmetic primitive, but it sits under security and integrity features that show up across modern systems. NVIDIA points to authenticated encryption, error-correcting codes and zero-knowledge proofs as examples of software that can depend on efficient binary extension field arithmetic. In practice, that means the change is less about a new end-user product and more about making existing GPU hardware a better target for specialized cryptographic pipelines.

The clearest benchmark in NVIDIA's post is GHASH, the authentication component used in AES-GCM. NVIDIA says GHASH reaches up to an 18.8x speedup and 6.3 TB/s on a B200 GPU compared with bitsliced circuits, approaching DRAM bandwidth. The company also reports similar throughput improvements on the GeForce RTX 5090, which matters because it suggests the feature is not confined to only one data-center accelerator.

Developer impact

For teams working on privacy-preserving computation, the update could be especially useful. NVIDIA says sum-check protocols over GF(2^128), a building block used in some zero-knowledge proof systems, were accelerated by 3x to 13x with clmad. That does not automatically make every proof system GPU-bound or production-ready, but it gives library authors a hardware primitive that can reduce the amount of hand-optimized workaround code needed for large, parallel arithmetic workloads.

The announcement is also a reminder that cryptography performance often depends on hardware details that are invisible to most users. CUDA 13.3 does not change the security properties of AES-GCM or zero-knowledge protocols by itself. Instead, it gives developers a faster implementation path on supported NVIDIA GPUs, with the biggest benefits likely to show up first in libraries, research prototypes, data-center services and high-throughput systems where many operations can be batched across thousands of GPU cores.

Sources

• developer.nvidia.com

Cover photo by Adriano Ponte Abreu on Pexels, used under the Pexels License.