From 91c0ea0216a7d7b400b7dd171ca25cb1102c9f6c Mon Sep 17 00:00:00 2001
From: Avik Pal <avikpal@mit.edu>
Date: Fri, 25 Aug 2023 17:49:25 -0400
Subject: [PATCH] Add documentation entry

---
 docs/src/solvers/solvers.md | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/docs/src/solvers/solvers.md b/docs/src/solvers/solvers.md
index f7e52c2a7..fe8a8b6e4 100644
--- a/docs/src/solvers/solvers.md
+++ b/docs/src/solvers/solvers.md
@@ -72,6 +72,14 @@ choice of Krylov method should be the one most constrained to the type of operat
 has, for example if positive definite then `Krylov_CG()`, but if no good properties then
 use `Krylov_GMRES()`.
 
+!!! tip
+
+    If your materialized operator is a uniform block diagonal matrix, then you can use
+    `SimpleGMRES(; blocksize = <known block size>)` to further improve performance.
+    This often shows up in Neural Networks where the Jacobian wrt the Inputs (almost always)
+    is a Uniform Block Diagonal matrix of Block Size = size of the input divided by the 
+    batch size.
+
 ## Full List of Methods
 
 ### RecursiveFactorization.jl
@@ -106,6 +114,7 @@ LinearSolve.jl contains some linear solvers built in for specailized cases.
 ```@docs
 SimpleLUFactorization
 DiagonalFactorization
+SimpleGMRES
 ```
 
 ### FastLapackInterface.jl