Welcome, Guest. Please login or register.

Author Topic: New type of accelerator design?  (Read 15159 times)

Description:

0 Members and 1 Guest are viewing this topic.

Offline Karlos

  • Sockologist
  • Global Moderator
  • Hero Member
  • *****
  • Join Date: Nov 2002
  • Posts: 16879
  • Country: gb
  • Thanked: 5 times
    • Show all replies
Re: New type of accelerator design?
« on: December 14, 2009, 08:18:18 PM »
A 68020 compatible core (possibly with a few 040 user mode extensions such as move16) implemented in an FPGA would seem the best bet to me. I don't know how feasible that is, but if it could be done would be an ideal accelerator base maybe. If it could be implemented, along with a basic memory controller that allows generic DDR memory (simply based on availability) that'd be nice.

Something like that would certainly have me interested :)
int p; // A
 

Offline Karlos

  • Sockologist
  • Global Moderator
  • Hero Member
  • *****
  • Join Date: Nov 2002
  • Posts: 16879
  • Country: gb
  • Thanked: 5 times
    • Show all replies
Re: New type of accelerator design?
« Reply #1 on: December 14, 2009, 08:25:47 PM »
Regarding the original suggestion, it's hard to see how using a "fast and cheap" x86 makes sense when you can just run UAE to the same end and have less latency from the rest of the system.
int p; // A
 

Offline Karlos

  • Sockologist
  • Global Moderator
  • Hero Member
  • *****
  • Join Date: Nov 2002
  • Posts: 16879
  • Country: gb
  • Thanked: 5 times
    • Show all replies
Re: New type of accelerator design?
« Reply #2 on: December 30, 2009, 07:54:58 PM »
Quote from: Tron2k2;535456
That would be HOT. I'd love to plug such a thing into my A500.  I always wondered why someone didn't just make a board which went into the 1200's accelerator slot, had a USB2 cable at the other end, and essentially made the 1200 a slave to the modern machine on the other end.  USB2's bandwidth exceeds anything happening inside the 1200, it could just see the other box as a really fast accelerator.

USB2 has a theoretical peak bandwidth of up to 480Mb/s (megabits/sec). In actual use you'll not see that sort of speed. My USB2.0 hard disk can pump data at around 30MB/s, which is certainly not faster than (for example) local memory speed on a good accelerator card.

Of course, the main problem with USB is not so much the theoretical bandwidth but the latency. USB is a serial, packet based protocol, pretty much designed for streaming data between devices. Latency there is not that big an issue, a few milliseconds point to point lag doesn't matter as long as the packets keep transferring at a continuous rate. That's not true for your A1200 trapdoor where you have dozens of independent (and interdependent) signals to worry about. Hardware interrupts and the like don't need to be held up by being encapsulated into a USB packet, sent down a wire to a host, which in turn will be interrupted to deal with it and so on (yes I realise USB has interrupt and control transfers but I don't think their timing requirements are necessarily suitable for dealing with stuff happening inside your miggy).
« Last Edit: December 30, 2009, 07:58:26 PM by Karlos »
int p; // A