Return to BSD News archive
Received: by minnie.vk1xwt.ampr.org with NNTP id AA1244 ; Tue, 23 Feb 93 14:32:02 EST Path: sserve!manuel.anu.edu.au!munnari.oz.au!hp9000.csc.cuhk.hk!saimiri.primate.wisc.edu!zaphod.mps.ohio-state.edu!howland.reston.ans.net!spool.mu.edu!uunet!math.fu-berlin.de!mailgzrz.TU-Berlin.DE!gmdtub!bigfoot!tmh From: tmh@condor.first.gmd.de (Thomas Hoberg) Newsgroups: comp.unix.bsd Subject: Re: ATI Mach32 and Xfree86 or XS3? Message-ID: <TMH.93Feb13043134@condor.first.gmd.de> Date: 13 Feb 93 03:31:34 GMT References: <1993Jan29.193936.13423@netcom.com> <1993Jan29.225541.20260@Informatik.TU-Muenchen.DE> <TMH.93Feb6014407@keks.first.gmd.de> <1993Feb6.051805.276@netcom.com> Sender: news@bigfoot.first.gmd.de Organization: GMD-FIRST, Berlin Lines: 78 In-reply-to: hasty@netcom.com's message of 6 Feb 93 05:18:05 GMT Amancio Hasty writes: For a chipset that is not quite a good fit for the X architecture it has generated faster graphics that any PC X server which I heard of and with the 928 it will be faster than the sun workstation ipx/gx by more than 30%. BTW: the 801 at 1024x728 60MHz is faster than the ATI Ultra Pro. On the Dos arena please read Jan 93' Byte, S3 vs. ATI Ultra pro, the 801 won. Well, I don't trust any benchmark that I've run myself or whose sources I haven't read. I usually rely on x11perf when comparing graphics architectures, because it doesn't just give you a single number (perfect for ads) but lots and lots of numbers (perfect for weading out strenghts and weaknesses. XS3 does use vga banking and we are looking into mapping the entire videoram memory. Perhaps, at an address range higher than 32MB. (X8514 used the banked CFB code as it came. I took it out: it's not needed there) As for it being a 16bit interface on the local bus we can address the 805 and the 928 using a 32 bit interface. I am not sure that this will help performance. It shouldn't. It might actually hurt, depending on the chipset (which has to translate those memory cycles into two consecutive 16 bit cycles). On local bus, the bottleneck on the fixed graphic accelarateror chips is more on the VRAM or DRAM refresh cycle. The latter according to Thomas Roell suffers more as the dot clock increases. Also, this has been empirically confirmed by Jon Tombs and I. It's called bus contention. It doesn't happen so much on VRAM, because graphics output gets in the way of an ordinary memory access only once in 256 cycles (when the shift register is reloaded). Sure, memory bandwidth is an upper limit for any graphics architecture. That's why you need wide busses (e.g. 128bit), optimized memory architectures and intelligent RAM (e.g. TMS VRAMs) to get snappy performance. What will be nice to have today is 3D hardware support also the ability for the chip to perform bitblt operations via DMA. Currently, we read and write images to the vga in a tight loop which performs banking as neccessary. Sun's GX has 3D hardware support (that is, it sports a full geometric pipeline for things like PEX). TI's 34020 can support several co-processors at once (forgot the part number-starts with 34, too, and can support external microcode RAM for USER DEFINED and downloadable INSTRUCTIONS!) On DMA: No No No. Not if you do physical DMA (which all you get on x86 platforms). This is what the XGA does and it's, excuse me, bloody awful. DMA doesn't really gain you performance to begin with. Since memory bandwidth is the only limiting factor for graphics performance, DMA can do no better than the CPU. Most important: With DMA THERE IS NO SECURITY! Memory protection ceases to exist. A simple BitBLt will deliver or destroy any data in the physical address range of the DMA processor. I simply wouldn't dare to use such a card. DMA might be ok to use with such slow devices as hard disks, where DMA setup is completely done in kernel context, but X expects the frame buffer to be accessible by processes in user context. While we argue about ATI vs S3 vs who knows what, the amiga/videotoaster folks are making commercials and animation for TV series like Babylon 5. Sorry, the only US network we get here in Germany is THE network (CNN). But I'm sure they don't do those animations (whatever they are) in real-time on the Amiga alone. Hmm, I still don't know what I should buy. Is there any PC based graphics card out there yet that supports a 32-Bit interface on the local bus that comes within 90% of the speed of ordinary memory yet supports accelleration (= or perhaps rather "automatization") of graphics primitives a la Mach32, GX or S3 928? --- Thomas M. Hoberg | Internet: tmh@first.gmd.de 1000 Berlin 41 | tmh@cs.tu-berlin.de Wielandstr. 4 | Germany | BITNET: tmh@tub.bitnet +49-30-851-50-21 |