No mention of the VPU’s for hardware video decoding (H.264/H.265)/encoding(H.265). Or that the GPU is more powerful than the one in the RPi 4B. And the one benchmark that worked (libjpeg-turbo tjbench 2.1.0) is definitely not using the VPU’s for hardware jpeg decoding acceleration. You can be sure that a generic encryption/decryption benchmark would totally ignore the encryption/decryption engines of JH7110. The real problem with generic benchmarks is that they do not typically use dedicated hardware acceleration unless access is builtin to the library API that they are using.
Well software vs software comparison is not useless, as long as you compare apples to apples.
You can compare ARM with SIMD vs RISC-V without SIMD/Vector, but that’s not really fair, you should also include ARM without SIMD enabled in the test list.
Also this is mentioned in the article, compiler support for the RISC-V is still fresh, and you can see dramatic performance difference going through different version of GCC and Clang, and the lack of hand optimised code for some use-cases is clearly an issue, especially for library like jpeg decoding/encoding.