3 ms·

On ISA-bus DMA: it was horrendously slow. They kept the 8237 DMA controller speed pretty much unchanged even after faster processor speeds appeared (maybe for

On ISA-bus DMA: it was horrendously slow. They kept the 8237 DMA controller speed pretty much unchanged even after faster processor speeds appeared (maybe for backwards compatibility- the one thing that the PC centralized DMA controller must work for is the floppy controller). I designed a card that used DMA before I understood this. PIO was much faster. So for example "UDMA" for the ATA-3 IDE drives was dependent on PCI-bus or on-motherboard ports.

ajross 6y ago
The 8237 is an ISA device and operates at the bus speed (with some overhead, ymmv, yada yada). The bus didn't get faster so the devices it automated didn't get faster. But PC DMA certainly did get faster. The PCI bus has bus mastering DMA (i.e. none of that fly-by nonsense implemented by the 8237) as a core feature, and PCI network cards used it. But they were expensive for quite a few years, so we were still buying ISA cards well into the 90's. Which were slow. Because ISA.
- kiwidrew 6y ago
  > The 8237 is an ISA device and operates at the bus speed (with some overhead, ymmv, yada yada). It's even worse than that -- the 8237 DMA controller is clocked at 4.77Mhz for backwards compatability. The normal ISA bus clock is ~8.25Mhz. So ISA DMA is actually half the speed of ISA memory or PIO access. https://wiki.osdev.org/ISA_DMA https://wiki.osdev.org/ISA_DMA is a good reference to the ugly gory details of DMA...
  - ajross 6y ago
    That's correct, but sort of missing context. The original PC bus was synchronous, and everything lived on it. So the idea was you'd fetch an instruction byte from the bus, then another, then another, then spend a cycle interpreting the instruction, then execute the load or whatever it was on the bus, then go back to do it over again. In that world a DMA controller that can do a transfer ever other cycle is much faster than the CPU copy it would be replacing. Modern readers, weened on PCI, see something like "The ISA bus is 16 bits wide at 8 MHz" and think they should be able to stream 16MB/s "across" it "into" main memory. But that's not how it worked with ISA. It's not how it ever worked with ISA. The memory was on the same bus! And the CPU didn't even have a L1 cache!
    - kiwidrew 6y ago
      Indeed. Running the DMA at 4.77Mhz made a lot of sense in the context of the original PC and XT, especially since the 8237 was only rated for 5Mhz anyways... [Not to mention that the 8088 took 4 cycles to read a byte from memory and 5 cycles to read a byte from I/O.] The true power of the 8237 was that it never touched the data byte itself: it just manipulated the bus lines, transferring a byte directly from I/O device to memory in just 4 cycles, much faster than the 8088 could ever manage.
    - andrewf 6y ago
      https://www.lo-tech.co.uk/xt-cfv3-dma-transfer-mode/ https://www.lo-tech.co.uk/xt-cfv3-dma-transfer-mode/ has some benchmarks for Compact Flash cards on an XT bus. On a V20+/186+, "rep insb" saves a lot of instruction fetch traffic. (The page also talks about breaking up DMA transfers, so a Soundblaster on the same bus can do its own DMA often enough)
- jhallenworld 6y ago
  Yeah, I saw DMA going much slower than PIO on the ISA bus. I remember scope shots showing the strobes being longer than anything else happening on the bus. Right, PCI bus master DMA was fast (probably true of EISA and MCA also.. but don't know first hand).

ajross 6y ago

The 8237 is an ISA device and operates at the bus speed (with some overhead, ymmv, yada yada). The bus didn't get faster so the devices it automated didn't get faster. But PC DMA certainly did get faster. The PCI bus has bus mastering DMA (i.e. none of that fly-by nonsense implemented by the 8237) as a core feature, and PCI network cards used it. But they were expensive for quite a few years, so we were still buying ISA cards well into the 90's. Which were slow. Because ISA.

kiwidrew 6y ago
> The 8237 is an ISA device and operates at the bus speed (with some overhead, ymmv, yada yada). It's even worse than that -- the 8237 DMA controller is clocked at 4.77Mhz for backwards compatability. The normal ISA bus clock is ~8.25Mhz. So ISA DMA is actually half the speed of ISA memory or PIO access. https://wiki.osdev.org/ISA_DMA https://wiki.osdev.org/ISA_DMA is a good reference to the ugly gory details of DMA...
- ajross 6y ago
  That's correct, but sort of missing context. The original PC bus was synchronous, and everything lived on it. So the idea was you'd fetch an instruction byte from the bus, then another, then another, then spend a cycle interpreting the instruction, then execute the load or whatever it was on the bus, then go back to do it over again. In that world a DMA controller that can do a transfer ever other cycle is much faster than the CPU copy it would be replacing. Modern readers, weened on PCI, see something like "The ISA bus is 16 bits wide at 8 MHz" and think they should be able to stream 16MB/s "across" it "into" main memory. But that's not how it worked with ISA. It's not how it ever worked with ISA. The memory was on the same bus! And the CPU didn't even have a L1 cache!
  - kiwidrew 6y ago
    Indeed. Running the DMA at 4.77Mhz made a lot of sense in the context of the original PC and XT, especially since the 8237 was only rated for 5Mhz anyways... [Not to mention that the 8088 took 4 cycles to read a byte from memory and 5 cycles to read a byte from I/O.] The true power of the 8237 was that it never touched the data byte itself: it just manipulated the bus lines, transferring a byte directly from I/O device to memory in just 4 cycles, much faster than the 8088 could ever manage.
  - andrewf 6y ago
    https://www.lo-tech.co.uk/xt-cfv3-dma-transfer-mode/ https://www.lo-tech.co.uk/xt-cfv3-dma-transfer-mode/ has some benchmarks for Compact Flash cards on an XT bus. On a V20+/186+, "rep insb" saves a lot of instruction fetch traffic. (The page also talks about breaking up DMA transfers, so a Soundblaster on the same bus can do its own DMA often enough)
jhallenworld 6y ago
Yeah, I saw DMA going much slower than PIO on the ISA bus. I remember scope shots showing the strobes being longer than anything else happening on the bus. Right, PCI bus master DMA was fast (probably true of EISA and MCA also.. but don't know first hand).