The Digester

Intel's Heracles accelerates fully homomorphic encryption up to 5,547x

Mar 11th 2026

Intel's Heracles is a purpose-built PCIe accelerator that runs fully homomorphic encryption on encrypted data, claiming 1,074 to 5,547 times the FHE math throughput of a 24-core Xeon by using an 8192-way SIMD engine, 48 GB HBM3, and dedicated NTT and bootstrapping hardware.

  • Heracles performs fully homomorphic encryption end to end, processing data while it remains encrypted and outputting encrypted results.
  • Intel reports Heracles is 1,074 to 5,547 times faster than a 24-core Intel Xeon W7-3455 for seven FHE math operations used in benchmarks.
  • The accelerator uses an 8192-way SIMD engine made from 64 tile-pairs in an 8x8 mesh with 32-bit arithmetic slices and hardware NTT support, plus automorphisms and bootstrapping.
  • Memory and bandwidth are built for heavy parallel FHE math with 48 GB of HBM3 across two stacks, 64 MB internal scratchpad, and custom data paths delivering terabytes per second internally.
  • Peak throughput is about 29.5 TOPS for butterfly primitives, 9.8 TOPS for modular arithmetic, and multi-terabit per second transform performance, and the chip supports BGV, BFV, and CKKS schemes.
  • Heracles is a purpose-built PCIe accelerator running at 1.2 GHz, fabricated on Intel 3, sized at 197 mm², rated 176 W, requires liquid cooling, and cannot run general-purpose software or an operating system.

Articles