They presented an approach which was not cache optimized which also showed significant gains. And of course they used a lot of cores, they're comparing against

They presented an approach which was not cache optimized which also showed significant gains. And of course they used a lot of cores, they're comparing against a v100, which is a $7k gpu (maybe $5k if you're lucky).