velocitatem/PHANTOM
1
0
Fork 0
mirror of https://github.com/velocitatem/PHANTOM.git synced 2026-05-31 16:43:36 +00:00
Code Issues Packages Projects Releases Wiki Activity

updating computation power graph

Browse Source
This commit is contained in:
Daniel Rosel
2026-03-08 14:22:54 +01:00
parent 17c128cbc0
commit 28dbcacd95
6 changed files with 142 additions and 114 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
paper/src/chapters/03-methodology.tex
View File

@@ -210,8 +210,7 @@ The simulator has multiple configurable factors. We design a multi-factor study
% Power analysis plan: apply a two-sample Mann-Whitney U (or permutation test) on per-session (delta_H - delta_A) divergence scores comparing the human and agent groups. Compute minimum detectable effect size at alpha=0.05, power=0.8, given n=18 per group. Bootstrap confidence intervals on mean KL are a cleaner complement given the non-normality of divergence distributions.
While this scale is generally expensive for reinforcement learning, we execute it on a large TPU cluster to make the sweep tractable.
% TODO: cite in the apendix the math to get to 160 petaflops of compute
Our training budget is provisioned through TPU Research Cloud and spans 384 chips across TPU v4, v5e, and v6e generations, with a spot-heavy allocation plus an on-demand reserve. At peak BF16 throughput this corresponds to approximately 160 PFLOPS of aggregate compute, which makes repeated seeds, ablations, and sensitivity sweeps feasible within practical wall-clock limits. We allocate v6e capacity to the highest-intensity policy training jobs, use v5e for wider hyperparameter exploration where throughput-per-dollar is favorable, and reserve on-demand v4 capacity for runs that should not be interrupted.
Our training budget is provisioned through TPU Research Cloud and spans 384 chips across TPU v4, v5e, and v6e generations, with a spot-heavy allocation plus an on-demand reserve. At peak BF16 throughput this corresponds to approximately 160\,PFLOPS of aggregate compute (derivation in Appendix~\ref{app:compute_budget}), which makes repeated seeds, ablations, and sensitivity sweeps feasible within practical wall-clock limits. We allocate v6e capacity to the highest-intensity policy training jobs, use v5e for wider hyperparameter exploration where throughput-per-dollar is favorable, and reserve on-demand v4 capacity for runs that should not be interrupted.
\begin{table}[ht]
\centering