From 8ef4a845ed275955e52ced2982f61771c5e868cd Mon Sep 17 00:00:00 2001
From: Romain PEREIRA <rpereira@anl.gov>
Date: Mon, 16 Mar 2026 10:44:27 -0500
Subject: [PATCH] Added documentation on experimental support for multi-gpu
 systems through XK.BLAS

---
 docs/make.jl          |  1 +
 docs/src/multi-gpu.md | 25 +++++++++++++++++++++++++
 2 files changed, 26 insertions(+)
 create mode 100644 docs/src/multi-gpu.md

diff --git a/docs/make.jl b/docs/make.jl
index b7735b525..4da9cc7e7 100644
--- a/docs/make.jl
+++ b/docs/make.jl
@@ -27,6 +27,7 @@ makedocs(
            "Storage requirements" => "storage.md",
            "Preconditioners" => "preconditioners.md",
            "GPU support" => "gpu.md",
+           "Multi-GPU support" => "multi-gpu.md",
            "Warm-start" => "warm-start.md",
            "Matrix-free operators" => "matrix_free.md",
            "Callbacks" => "callbacks.md",
diff --git a/docs/src/multi-gpu.md b/docs/src/multi-gpu.md
new file mode 100644
index 000000000..f887d054c
--- /dev/null
+++ b/docs/src/multi-gpu.md
@@ -0,0 +1,25 @@
+# [Multi-GPU support](@id multi-gpu)
+
+Krylov.jl has an experimental support for executing on multi-GPU systems through the [BLAS package of XK.jl](https://github.com/anlsys/xk.jl).
+Multi-GPU interfaces are analogous to single-GPU's --- relying on Julia's types dispatcher on `XKVector` and `XKMatrix` to target multi-GPUs.
+XK.jl handles work distribution and communications automatically and lazily.
+
+```julia
+using Krylov, XK
+
+# CPU Arrays
+A_cpu, y_cpu = symmetric_definite(n)
+
+# XK.jl Arrays
+A_xk = XKMatrix(A_cpu)
+y_xk = XKVector(y_cpu)
+
+# Run a conjugate gradient
+(x_cpu, stats) = cg(A_xk, b_xk)
+
+# At this point of the execution, 'x' may be distributed across multiple memories.
+# The next line triggers and wait for copies (e.g., D2H transfers) required so `x_cpu` holds a coherent replica on the host memory.
+XK.memory_coherent_sync(x_cpu)
+```
+
+See [XK.jl repository](https://github.com/anlsys/XK.jl/tree/main/examples/Krylov) for examples.