Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

spelling nit #917

Open
wants to merge 1 commit into
base: master
Choose a base branch
from
Open

spelling nit #917

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion general/faq/training/neuron-training.rst
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -34,7 +34,7 @@ For simplicity, you should consider each NeuronCore within your instances as an
What are the time to train advantages of Trn1?
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

While the answer is largely model defendant, training performance on Trn1 is fast due thanks for multiple system wide optimizations working in concert. Dependent on the data type, you should expect between 1.4-5X higher throughput on Trn1 as compared to the latest GPUs instances (P4d). For distributed workloads, 800Gbps EFA gives customers lower latency, and 2x the throughput as compared to P4d. (a Trn1n 1.6Tb option is coming soon). Each Trainium also has a dedicated collective compute (CC) engine, which enables running the CC ops in parallel to the NeuronCores compute. This enables another 10-15% acceleration of the overall workload. Finally, stochastic rounding enables running at half precision speeds (BF16) while maintaining accuracy at near full precision, this is not only simplifying model development (no need for mixed precision) it also helps the loss function converge faster and reduce memory footprint.
While the answer is largely model dependent, training performance on Trn1 is fast due thanks for multiple system wide optimizations working in concert. Dependent on the data type, you should expect between 1.4-5X higher throughput on Trn1 as compared to the latest GPUs instances (P4d). For distributed workloads, 800Gbps EFA gives customers lower latency, and 2x the throughput as compared to P4d. (a Trn1n 1.6Tb option is coming soon). Each Trainium also has a dedicated collective compute (CC) engine, which enables running the CC ops in parallel to the NeuronCores compute. This enables another 10-15% acceleration of the overall workload. Finally, stochastic rounding enables running at half precision speeds (BF16) while maintaining accuracy at near full precision, this is not only simplifying model development (no need for mixed precision) it also helps the loss function converge faster and reduce memory footprint.

What are some of the training performance results for Trn1?
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Expand Down