1. AVX512 allows a 64-bit conditional mask to be used over the 64-bytes (aka a 512-bit operation). Like GPUs (NVidia or AMD) before it, this allows for far simpler implementations of if-statements (or loops) in practice.

2. AVX512 has vgather and vscatter, aka vector-load and vector-store. While no computer accelerates this process, this still simplifies programming and read/writes to memory can be thought of on a per-lane basis.

3. Compress and Expand, new instructions not found in GPU-space, perform the stream-compression operation (and inverse operation). See: https://www.cse.chalmers.se/~uffe/streamcompaction.pdf , lots of practical applications here, nice to see just a simple "instruction" that handles this detail. Though GPU programmers can prefix-sum and build this primitive themselves in practice, its a common enough operation that a proper-instruction is likely a good idea.

------------

Just a few tidbits of AVX512 ISA that shows that Intel really knows what they're doing here.

#1 can be seen as a 64-bit generalization to the cmov conditional instructions that are in Intel's normal 64-bit instruction set. Except instead of being a "1-bit" compare or 1-bit if-statement, might as well generalize it to a 64-bit operation across 64x8-bit bytes that naturally map into a 512-bit register, right?

RVV has masking for everything (masks being stored as packed bits in a vector register), memory gather & scatter, compress (& easily emulatable expand via an instruction that expands 0,1,2,3,… by which you can then shuffle), and in addition to that is scalable (i.e. same code can run on hardware with anything between 128-bit and 65536-bit vector registers; plus something alike "hardware unrolling" where operations can be configured to run over groups of 1/2/4/8 registers; and a vector length option in addition to masking allowing trivially handling the tail of loops with the same body as the main loop, without affecting masking), and is still more orthogonal than AVX-512 (e.g. AVX-512 still only has 8- & 16-bit saturating add/sub, missing 8-bit shifts, whereas RVV has orthogonal everything (even high half of 64-bit product, allowing for two-instruction 64×64→128-bit multiplication)). Granted, it's significantly newer than AVX-512.