Why is everything IDE/ATA?

[daxb]

I think nowadays the interest isn't so high as back in the days. There must be a reason why nobody produced a DMA controller. All available ones are no DMA and max. PIO-4 when UDMA mode exists for several years already.

[Jope]

Probably the problem was: no m68k / zorro compatible DMA IDE controllers available during the early 90s and designing an ASIC was too expensive for the volumes that would be sold reasonably, SCSI was the preferred workstation storage bus anyway and DMA capable controllers for that were off the shelf items.

[pandy71]

Only few SCSI controllers on Amiga use DMA and then they usually use DMAC from Commodore with DMAC limitations (like 24 bit address bus) - not sure why such approach - perhaps general assumption was that CPU pool mode is sufficiently fast and more important is to have HDD at all, this is even more hilarious as A570 CDROM use DMA for 150KBps transfer(and there is possible to hook WD33c93 to use SCSI but still with lots of DMAC limitations). Perhaps issue was with programming? Personally never saw any documentation for Commodore DMAC and only known to me DMA controllers for MC68k family was those one designed by Motorola (MC68440, MC68450 and one or two less known). Is there any documentation for Commodore DMAC? DMA should not be very difficult on Amiga - bus seem to be straightforward.

[grelbfarlk]

SCSI DMA controllers I can think of off the top of my head, GVP HC+8, Blizzard 2040/60, Tekmagic 2060, Wildfire, WarpEngine, A4091, Fastlane, CSMK-1,2,3,PPC, GVP4060, Quikpak 060, A4000T SCSI, A530. They were pretty common.

IDE DMA controllers, can't think of one.

[Jope]

The only DMA IDE for the Amiga is the Masoboshi 702.

[hth313]

Originally Posted by ajk

I put an ACARD SATA-to-SCSI drive converter in my A3000. It takes any 2.5" SATA drive and interfaces it with the system quite nicely and is fairly future-proof. The only down side is that these things are rather expensive.

It is at least 10x more expensive than the actual drive, wow!

On the good side it does not occupy an expansion slot... and I can just drop by the local office supply shop and get a HD for it...

What kind of performance do you get out of it?

[ajk]

@hth313

The ACARD adapters won't be the bottleneck in any Amiga, in my experience. You'll get pretty much the maximum speed. For the A3000 internal SCSI it's a few megabytes per second. I've got a similar one in the A4000 with a CyberStormPPC, and that goes to something like 15-20 megs per second (if the tests are to be trusted; certainly seems pretty snappy in any case).

It's very much up to debate whether they are worth the money, but this was some years ago and at the time I couldn't find many other solutions.

[Daedalus]

The ACard adaptors are serious money indeed - I picked up a couple of SCSI-IDE ones a few years ago (10 years?), when they weren't nearly as expensive as they are now. On my setup of a Blizzard 1260 + SCSI kit (which also has DMA), I get around 8MB/s out of a theoretical maximum of 10MB/s using an ACard and an old ~100GB IDE hard drive. As ajk says, the bottleneck won't be the adaptor, it will be the Amiga itself.

[daxb]

But where is the bottleneck? Is it limited by RAM speed? What has the CPU to do when using DMA controller + connected harddrive? Ok, if the data needs to go through the motherboard then it would be slow but usually that isn't used because of it.

[Hewitson]

The GVP SCSI controllers are among the very best for the Amiga as far as I'm concerned. DMA capable, bulletproof, and "FAAAAAAAST"

[pandy71]

Originally Posted by grelbfarlk

SCSI DMA controllers I can think of off the top of my head, GVP HC+8, Blizzard 2040/60, Tekmagic 2060, Wildfire, WarpEngine, A4091, Fastlane, CSMK-1,2,3,PPC, GVP4060, Quikpak 060, A4000T SCSI, A530. They were pretty common.

Common but only amongst few main vendors capable to pay for ASIC or CPLD. There was no cheap DMA controller on MC68k market - 68440 or 68450 was quite expensive and uncommon out of VME world. Some of those SCSI controllers offered DMA only within 24 bit address space.

Originally Posted by daxb

But where is the bottleneck? Is it limited by RAM speed? What has the CPU to do when using DMA controller + connected harddrive? Ok, if the data needs to go through the motherboard then it would be slow but usually that isn't used because of it.

Count cycles - DMA transfer from IDE to CHIP obviously need to follow Agnus timing, transfer to FAST on CPU bus is less problematic but common only in large systems (i mean A3000 and higher) - at some point IDE PIO was capable to provide acceptable performance at a cost of 2x 74245 and additional glue logic.

[Toni Wilen]

Most boards listed above used 53C710+ SCSI chip that has SCSI and full DMA controller built-in.

Only Commodore and GVP used ASIC DMA controller.

[Daedalus]

Originally Posted by daxb

But where is the bottleneck? Is it limited by RAM speed? What has the CPU to do when using DMA controller + connected harddrive? Ok, if the data needs to go through the motherboard then it would be slow but usually that isn't used because of it.

It depends on the setup. On Zorro-II SCSI controllers, the Zorro bus is likely to be the bottleneck. On accelerator-mounted SCSI setups however, the Zorro bus isn't involved which helps a lot. With my Blizzard SCSI for example, the measured speed is relatively close to the theoretical maximum, so in that case the interface type is the bottleneck as it's only designed for 10MB/s. The Cyberstorm SCSI seems to manage twice that, again not that far off the theoretical maximum of its SCSI implementation, so in those cases the bottleneck is SCSI itself.

[daxb]

Yes, that is what I was thinking about the bottlenecks. So, if you put DMA controller on a accelerator card you could theoretical get UDMA speed 33-133MB/s. But nobody did it.

[grelbfarlk]

Cyberstorm MK3/PPC have an NCR53C770 which can do up to 40MB/s, 29MB/s is more typical with an IDE->SCSI bridge. Some people with SCSI-3 drives do 35MB/s with proper termination/cables etc. If you're getting less than ~29MB/s either your drive is the bottleneck or improper termination or cables are to blame.

Quikpak and Falcon 040 had an NCR53C720 which could do 20MB/s but I don't recall ever seeing benchmarks showing this.

[pandy71]

Originally Posted by Toni Wilen

Most boards listed above used 53C710+ SCSI chip that has SCSI and full DMA controller built-in.

Only Commodore and GVP used ASIC DMA controller.

Yes, indeed, NCR53C710 has implemented MC68K bus timing and logic. So this may explain a lot - if there is DMA on SCSI then or it will use DMAC or SuperDMAC from Commodore, or GVP or NCR chip with embedded DMA logic.
Thx Toni!

But is there any DMA on card with for example NCR53C94 or WD33C93 using DMAC not from Commodore or GVP?

btw this discussion remains me some discussion while ago about somebody tried to use regular uC (cheap, fast one) to implement virtual DMA IDE, accordingly to him it was no beneficial from Amiga perspective (but no further details was released).

[Toni Wilen]

Originally Posted by pandy71

But is there any DMA on card with for example NCR53C94 or WD33C93 using DMAC not from Commodore or GVP?

Blizzard SCSI boards use NCR5394/FAS216 or similar chips and most likely DMA controller is in MACH CPLD chips. (Can't see any obvious ASICs in SCSI Kit IV, for example)

Masoboshi uses NCR5394 and MC68440 (with custom markings on chip label! It was only found when emulation was added and it was noticed register spacing and locations looked suspiciously similar..)

HardFrame uses Adaptec AIC-6250 and MC68430.

Very rare SupraDrive 2000 (not the common PIO wordsync model) uses 5380 and MC68430

I think thats all..

[pandy71]

Thx Toni, so eventually list (of DMAC's used in Amiga designs) will be quite short - MC68440* (and similar DMAC from Motorola e.g. 68450, AFAIR it was also 68850 DMAC [not MMU 851] but seem it is forgotten as 68430 - can't find any datasheet), DMAC from Commodore (SDMAC on Zorro III compliant adapters?), NCR53X7x0 chips with embedded MC68K complaint DMAC, few companies implemented own DMAC (as ASIC or CPLD).

*Where 68440 based are 16 bit and limited to 24 bit address space (can be extended by auxiliary page unit?) and based on data-sheet they are unable to deliver more than 3.5MBps in Amiga (maximum 5MBps at 10MHz system clock, no WS) - this may explain why PIO is comparable to DMA speed if DMAC 68440 like was used.