You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
In perftest.jl the number used for andwidth is 1355: println("So that's cool. We are getting close to hardware limit, running at $(round(T_eff/1355*100), sigdigits=2) % of memory copy! 🚀")
Is that right. From what I can find online, the bandwith of A100 40GB is 1555. Also the readme states 1.55 TB/s.
Shall it be changed to 1555 in perftest.jl?
The text was updated successfully, but these errors were encountered:
1.55 TB/s is the vendor announced BW. Practically, memcopy often slightly differs from that theoretical number, reason I have the set the memcopy measured BW as ref here.
In perftest.jl the number used for andwidth is 1355:
println("So that's cool. We are getting close to hardware limit, running at $(round(T_eff/1355*100), sigdigits=2) % of memory copy! 🚀")
Is that right. From what I can find online, the bandwith of A100 40GB is 1555. Also the readme states 1.55 TB/s.
Shall it be changed to 1555 in perftest.jl?
The text was updated successfully, but these errors were encountered: