From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S262329AbUBXRcz (ORCPT ); Tue, 24 Feb 2004 12:32:55 -0500 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S262322AbUBXRcz (ORCPT ); Tue, 24 Feb 2004 12:32:55 -0500 Received: from rwcrmhc11.comcast.net ([204.127.198.35]:17632 "EHLO rwcrmhc11.comcast.net") by vger.kernel.org with ESMTP id S262336AbUBXRcm (ORCPT ); Tue, 24 Feb 2004 12:32:42 -0500 Subject: Re: Intel vs AMD x86-64 From: Albert Cahalan To: Dave Jones Cc: Albert Cahalan , linux-kernel mailing list , davem@redhat.com, Linus Torvalds In-Reply-To: <20040224164404.GB10157@redhat.com> References: <1077590524.8084.237.camel@cube> <20040224164404.GB10157@redhat.com> Content-Type: text/plain Organization: Message-Id: <1077635481.8120.300.camel@cube> Mime-Version: 1.0 X-Mailer: Ximian Evolution 1.2.4 Date: 24 Feb 2004 10:11:22 -0500 Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org On Tue, 2004-02-24 at 11:44, Dave Jones wrote: > On Mon, Feb 23, 2004 at 09:42:05PM -0500, Albert Cahalan wrote: > > Setting up only one of the IO-MMUs would be neat. > > AGPv3 standard mandates that you MUST keep all GARTs coherent, > so this isn't possible. The amd64 GART driver goes to great > lengths to make sure it does update the northbridges on every > CPU whenever something changes. Of course. That's the easy way; you won't need to worry about memory interleave or out-of-bounds prefetch if you keep everything coherent. I'm just saying it would be neat, and potentially useful, to intentionally violate this. Of greatest interest would be the 2-way Opteron boards that only have RAM connected to the CPU closest to PCI. The sidecar CPU :-) could be ignored.