From d7869b09b9db15bc609e8806ee23099a59d13131 Mon Sep 17 00:00:00 2001
From: osmarks <osmarks@mycorrhiza>
Date: Thu, 25 Dec 2025 23:14:22 +0000
Subject: [PATCH] =?UTF-8?q?Edit=20=E2=80=98autogollark=E2=80=99?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 autogollark.myco | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/autogollark.myco b/autogollark.myco
index 18308e8..5a1a921 100644
--- a/autogollark.myco
+++ b/autogollark.myco
@@ -38,7 +38,7 @@ Autogollark currently comprises the dataset, the search API server and the [[htt
 * https://arxiv.org/abs/2507.01335
 * https://github.com/d0rc/egg.c and https://eshyperscale.github.io/. Does this actually work (at scale)? Why? Would be really nice for using AMX units.
 * Maybe compute grants are available for training.
-* Substantial bandwidth bottleneck on CPU. Specdec/MTP would be useful.
+* Substantial bandwidth bottleneck on CPU (230GB/s nominal; 200GB/s benchmarked; 100GB/s per NUMA node, which llama.cpp handles awfully). Specdec/MTP would be useful.
 }
 * Search over conversations with non-gollark simulacra? Should find //something// to use spare parallelism on local inference. Best-of-n? https://arxiv.org/abs/2505.10475
 * {Longer context, mux several channels.