Welcome, Guest. Please login or register.

Author Topic: A new Zorro III memory board  (Read 8023 times)

Description:

0 Members and 1 Guest are viewing this topic.

Offline mboehmer_e3b

  • Sr. Member
  • ****
  • Join Date: Aug 2002
  • Posts: 312
    • Show all replies
    • http://www.e3b.de/usb/
Re: A new Zorro III memory board
« on: December 07, 2007, 08:35:11 AM »
Quote

The Mediator, Prometheus, and G-Rex can take PCI cards with plenty of FAST graphics memory. Even though there would be no drivers for these "newer" gfx cards, it shouldn't be too difficult to find there memory address and add it to the free memory.


All PCI memory "solutions" you list up here have three common problems:

(1) memory on gfx cards is optimized for writing. It is dead slow on reading, as usually you don't read it in normal operation. Reading on PCI may lead to delayed transactions, with penalty in time and the risk of Buster timeouts (can be seen when accessing PCI-PCI bridges or FlashROM on gfx cards).

(2) no gfx mem access without gfx processor initialization. The gfx mem is only accessible via the GPU, and this beast needs to be setup correctly.

(3) Endianess. Some gfx cards change the byte order of the memory when changing screen modes. In case you switch to a new screen, your memory gets scrambled. Bad idea.

Michael
 

Offline mboehmer_e3b

  • Sr. Member
  • ****
  • Join Date: Aug 2002
  • Posts: 312
    • Show all replies
    • http://www.e3b.de/usb/
Re: A new Zorro III memory board
« Reply #1 on: December 07, 2007, 04:17:24 PM »
Quote

Slow is relative. Zorro3 speeds are slow compared to what a gfx card can move around.


Relative is not the problem. If your PCI gfx card delays your PCI read just a few times, you are lost as the Buster will timeout the Zorro III cycle, leading to a bus error.

As you want to go to more memory I guess we are talking about modern PCI gfx cards and not the 4MB S3 Virge models.
Most PCI gfx cards nowadays are using PCI-AGP bridges, and so the problem gets worse. Those integrated bridges will add more delays, and the problem of write / read combining.

The bridge will follow certain rules when different read / write operations are pending, and this may lead to bad situations in real life (like: several writes are pending inside the bridge, followed by a read, so all writes will have to be completed before the read is completing -> long delayed read).

Quote

Do the buster problems occur with a Revision 11 buster?


The timeout is not a "problem". It's intended to break deadlocks when a Zorro III access fails.
Or do you mean the other Buster "features"? Which one, as there are quite a few of them, and they are available in different tastes in all Buster revisions :-(

Michael