I think it boils down to what sense of "reality" one holds to. Personally, I have to wonder why test for something that never is? I suppose if I were in the sciences, that would not be a true statement, but as a person who has done multimedia of for over 20 years, my world doesn't necessarily need the "hard knocks" prime95 seems to like to pass out with this AVX offset stuff as currently offered. Besides, AVX would seem to be a method to use high overclocks on lighter loads with a facility to downclock the heaver ones automatically. That would mean trying to stress AVX past a reasonable point would really be using an opposite philosophy to the purpose for AVX -- Sort of an antithesis to it.
I'm just glad I can do overclocking without crashing now testing with prime96 small FFT's. It also opens up the possibility of testing for higher overclocks, using prime95, which I have used since I found it some many years ago.
-Rodger
If you think Prime95 AVX/FMA3 utilizes the CPU completely, You should try Intel linpack benchtesting. https://software.intel.com/en-us/articles/intel-linpack-benchmark-download-license-agreement
Prime95 with the AVX/FMA3 instructions utilizes the floating point extensively and light utilization on the integer. Where as linpack utilizes floating point and integer fully. So with prime95 you can brows the web without slowdown in tasks. Then when linpack is bursting the browsing tasks slow and stop sometimes. Programs are mostly integer based unless they have calculations for math and science then that is floating point mostly. Games use integer and floating point however the data is slowed by system memory and the process of multitasking. Howsoever, when stressing floating point unit mostly the basic numbers fit into high speed cache, so not using the main memory much causes a slow down of data fed to the processor like most integer programs. So linpack and prime95 AVX/FMA3 is worst case, it keeps the FPU transistors busy.