The real FLOPS & Vector use on the iMac Pro

i'm interested in knowing how processor hardware embedded vector instruction set improves flops number apple's claimed number on imp web page ?

suspect geekbench 4 not, @ time, have code recompiled employ vector instructions , therefore show/display apple's flops claim.

guess can write simple program exploit new vector instructions , have report flops compare against geekbench 4 reports.

thoughts.... ?

 

https://www.geekbench.com/doc/geekbench4-cpu-workloads.pdf

quote

floating point workloads gemm

gemm (general matrix multiplication) computes result c = ab + c, a, b, , c matrices.

gemm workload uses 512x512 single-precision matrices. gemm implementations written using hand-tuned vector instructions following instruction sets:

• avx + fma
• avx
• sse3
• sse2

• armv8 neon • armv7 neon

click expand...

so-- think incorporated. see https://www.intel.com/content/www/us/en/architecture-and-technology/avx-512-overview.html

quote

with ultra-wide 512-bit vector operations capabilities, intel® avx-512 can handle demanding computational tasks.

applications can pack 32 double precision , 64 single precision floating point operations per second per clock cycle within 512-bit vectors, 8 64-bit , sixteen 32-bit integers, 2 512-bit fused-multiply add (fma) units, doubling width of data registers, doubling number of registers, and, doubling width of fma units, compared intel® advanced vector extensions 2.0 (intel® avx2).3

click expand...

is there anomalously high gemm sub-score among cpus support avx-512?

 

Forums Macs Desktops iMac

• iPhone
• Mac OS & System Software
• iPad
• Apple Watch
• Notebooks
• iTunes
• Apple ID
• iCloud
• Desktop Computers
• Apple Music
• Professional Applications
• iPod
• iWork
• Apple TV
• iLife
• Wireless