Xref: utzoo comp.sys.amiga:24202 comp.sys.amiga.tech:2180 Path: utzoo!attcan!uunet!cbmvax!daveh From: daveh@cbmvax.UUCP (Dave Haynie) Newsgroups: comp.sys.amiga,comp.sys.amiga.tech Subject: Re: Hard Disk Performance tests, comments invited Message-ID: <5034@cbmvax.UUCP> Date: 19 Oct 88 17:23:34 GMT References: <10150@cup.portal.com> Distribution: na Organization: Commodore Technology, West Chester, PA Lines: 41 in article <10150@cup.portal.com>, thad@cup.portal.com (Thad Thad Floryan) says: > Xref: cbmvax comp.sys.amiga:26234 comp.sys.amiga.tech:2750 > Comments are invited. I can relay messages to/from the original poster. OK, here are some... > DMA boards cannot access any 32bit memory, so they must write/read > to/from chip ram, or 16bit ram, were it can then be copied by the > operating system into 32bit ram. This is implementation dependent. With Ronin's 32 bit memory, no DMA to 32 bit memory is possible. With the 32 bit memory on a Commodore A2620, DMA from 16 bit devices works just fine. The 16->32 bit copy mentioned is because you're using FFS, and have 32 bit memory in front of 16 bit memory. Certainly, in this specific case, a well designed CPU driven hard disk would be expected to run faster. For example the GVP board requires the CPU to copy from it's static RAM into some 32 bit RAM, and that's the only slowdown on the system. With any DMA driven device and the Ronin boards, you'd have real DMA into some 16 bit memory, and then a CPU copy into 32 bit RAM. That's essentially the GVP time, PLUS the DMA time. You loose. Now consider the same setup using DMAable 32 bit memory on the A2620. The GVP still has 16 bit memory. The most efficient transfer loop would use a long to long copy (no tricks are really needed, since the 68020 or 68030 will cache all the instructions in the copy loop). It'll take two 16 bit/7.16 MHz memory cycles (8 - 7.16MHz clocks) to transfer a longword to the 68020, one 32 bit/14.3 MHz memory cycle (2 - 7.16MHz clocks, the Ronin board appears to have one wait state, so the 68020 will run a 4 clock cycle). With DMA directly to 32 bit RAM, you need two 16 bit/7.16 MHz cycles (8 - 7.16MHz clocks) to transfer the longword. Period. So you go faster than non-DMA would. However, the difference between the DMA and non-DMA device is much less than in a complete 16 bit system, since the access to 32 bit memory is fast. -- Dave Haynie "The 32 Bit Guy" Commodore-Amiga "The Crew That Never Rests" {uunet|pyramid|rutgers}!cbmvax!daveh PLINK: D-DAVE H BIX: hazy "I can't relax, 'cause I'm a Boinger!"