Minimum pin-out for 64bit CPU?

[Rudy Toody]

I've had too much time on my hands and have been wondering it one could create a hot-pluggable 64bit cpu. Something you could plug into an add-in board whenever you needed its capability.

I figure that one would replace all parallel busses with serial, etc.

Anybody have a SWAG?

--Fred

 

[BrownTown]

Well I mean it some far out theory all you would need is 1 pin for all the data and just put every thing in serial, but thats not really practicle at all. Although the entire premise isn't really all that sensical either since when would you ever just want to randomly plug in another processor? And if you did want that I'm pretty sure you can just get a 2 processor board and only by one CPU and then some time latter buy a second one and put that in but don't quote me on that. Also, I googled and PCIE has like 164 pins so that would be possible to use. A shit ton of the pins on a normal CPU and power and ground and with the add in board you would only need 2 wires for this and then all the pins would be wired together on the board.

 

[Rudy Toody]

Brown Town--

I was thinking about a mezzanine board with 64 of these sockets for 64 of these chips to create a tiny Beowulf class computer. The chip would run at a frequency that wouldn't require an elaborate heat sink or fan. Just plug it in like you would a thumb drive.

Distributed Computing is the application I had in mind. Each processor would handle one work unit and if the chip ran at 750MHz it would require 8 chips to equal one dual core processor at 3GHz.

The benefit to something easily inserted/removed is that you can use specialty processors. For example: use a normal cpu for simple math needs and a gpu processor for more sophisticated math needs. Mix and match on the board. Also, by using an adaptor on the socket, one could also use thumb drives if extra memory is required for the application.

All of this would be controlled by a cluster controller on another board.

--Fred

 

[degibson]

Originally posted by: Rudy Toody
I've had too much time on my hands and have been wondering it one could create a hot-pluggable 64bit cpu. Something you could plug into an add-in board whenever you needed its capability.

I figure that one would replace all parallel busses with serial, etc.

Anybody have a SWAG?

There's no reason that hot-pluggable CPUs need to be serial, by the way. They would only need to be serial if they were USB-able, which would be a lousy way to do it anyway, since the CPU would have no direct access to memory.

Sun has actually had hot-pluggable CPU boards for quite some time, as well as most other higher-end server vendors. The challenge with bringing this technology to the general purpose desktop platform is twofold:

1) You'd need a reason -- quad core already stretches most users' capability to exhaust their computation resources.

2) You'd need a fast way to access main memory, and a coherence mechanism something like the multi-chip machines of 5-10 years ago. Hopefully, this wouldn't slow down the main processor(s) a the point beyond the performance gain of the hot-plug CPU(s)

 

[Peter]

It doesn't get any simpler (currently) than with AMD's HyperTransport. And yes, hot pluggable HT CPU slots have already been done.

 

[Modelworks]

This is done all the time in the embedded field.
The method for passing data from micro to micro is generally I2c or SPI.

You start with a micro that acts as the master, controlling the data to all the slaves.
Slaves can come and go as needed.
Each slave is given a unique id so that it can tell the master "Its me, I'm here, and this is what I do !"


Its really no different how large render farms are run.
There is a box that acts as master, and then all the slaves run under it, often several hundred CPU.
The slaves can be added or subtracted as the program runs.

 

[degibson]

Peter is quite right about HyperTransport -- its the cats meow! Now, if we could find a way to run coherent HyperTransport out to the USB controller, this could get interesting! As an aside, is the CHT protocol still under wraps? Do you happen to have a reference, Peter?