This depends on what your workload is doing. Guava's cache is built on top of Java 5's ConcurrentHashMap with a default concurrency level of 4. This setting is because that hash table is segmented into multiple smaller tables, so more segments allows for higher concurrency at a cost of a larger memory footprint. A setting of 64 is more appropriate for performance sensitive scenarios.
Java 8 rewrote the hash table implementation for a 2x speedup and no longer relies on coarse segmentation. To leverage this required a complete rewrite of Guava's cache, which was done in Caffeine. Given the ability to start from scratch and build on everything we learned in Guava, there are many other performance improvements that were planned but not realized in Guava's implementation. This rewrite has about 33% of the read throughput of an unbounded ConcurrentHashMap and can obtain a higher hit rate than Guava's LRU.
The last time that I performed an in-depth benchmark was in 2015, but I do not believe any changes have been dramatic enough to invalidate these results.
In short, first try increasing the concurrencyLevel and, if not satisfactory, then switch to Caffeine.