Erroneous measurement
On occasion, criterion-cycles-per-byte returns a value that visually appears extremely close to 264. I would imagine this is the result of an overflow. It should go without saying, but even a single instance of this occurring eliminates any usefulness of that specific benchmark.