search

harvard-acc / smaug

SMAUG: Simulating Machine Learning Applications Using Gem5-Aladdin
https://harvard-acc.github.io/smaug_docs
BSD 3-Clause "New" or "Revised" License
96 stars 27 forks source link

Can we run multiple instances of NVDLA minerva model on smaug . #109

Open sadia-4 opened 1 year ago

sadia-4 commented 1 year ago

If we can run multiple instances 4-layer Minerva model at a time ? If yes how can we?

xyzsam commented 1 year ago

Aladdin simulates a single accelerated kernel in isolation. Multi-tenant systems are not possible; however, you can instantiate four copies of the NVDLA module and run a single model on each one in parallel. This is supported with the --num-accels flag. You would use that to generate multiple traces , then configure the system to use all of them in simulation.

xyzsam commented 1 year ago

If I wanted to estimate metrics like power and time for NVDLA models.How do we do it on smaug.

NVDLA is modeled using Aladdin, so Aladdin provides the power and cycles modeling for you.