Intel Demos Chip To Compute With Encrypted Data
Source: Slashdot
Background
Fully homomorphic encryption (FHE) allows computation on encrypted data without ever decrypting it. The major drawback has been performance: FHE can be thousands to tens of thousands times slower on conventional CPUs and GPUs compared to processing plaintext. To close this gap, universities, startups, and major processor manufacturers have been developing specialized chips.
Intel’s Heracles Chip
At the IEEE International Solid‑State Circuits Conference (ISSCC) in San Francisco, Intel unveiled Heracles, a chip designed to accelerate FHE workloads. According to Sanu Mathew, who leads security circuits research at Intel, Heracles is the first hardware that works at scale:
- Size & Process: Approximately 20 × larger than typical FHE research chips (around 200 mm²) and fabricated using Intel’s 3‑nm FinFET technology.
- Memory: Integrated with two 24‑GB high‑bandwidth memory (HBM) chips in a liquid‑cooled package, a configuration usually reserved for AI GPUs.
- Performance Claim: Up to 5,000‑fold speedup over a top‑of‑the‑line Intel Xeon server CPU for FHE tasks.
Demonstration of Compute Scaling
The live demo at ISSCC showcased a private query use case:
- A voter encrypts her ID and vote and sends it to a government database that stores encrypted voter records.
- The server, without decrypting the data, determines whether the encrypted vote matches the stored record and returns an encrypted result.
- The voter decrypts the answer locally.
- Xeon CPU: 15 ms per query.
- Heracles: 14 µs per query.
While the per‑query difference is imperceptible to a single user, scaling to 100 million ballots translates to more than 17 days of CPU time versus about 23 minutes on Heracles.
Read more of this story at Slashdot.