motherboard architecture overhaul

[mooshi]

i think it is loooong overdue for a real overhaul on motherboard architecture.

what do you all think of the following proposal?

No more north/south bridge
No more PCI
No more AGP
No more IDE
No more floppy drives, serial connections, parallel connections or even SCSI.
No more Front-side Bus

where does this leave us? well, why don't we simply have an N-way switch. perhaps based on IEEE 1394b, perhaps HyperTransport, perhaps something else. it doesn't matter. just make it serial, and point-to-point.

now, suppose we go with 1394b. this allows for up to an 800MHz clock speed, with up to 1.6Gb/sec transfer (up to 3.2 over optical lines sometime soon).

now, all we have for the "motherboard" is, say, an 8-way switch with 8 general sockets. each of these sockets is identical and just a serial interface (in this case 1394b).

and into each of these sockets can be inserted an expansion card. and on these expansion cards we put the interface/hardware we want, working with only a 1394b interface. now, for an N-way system, we'll need N-1 individual 1394b controller/transceivers on each card for dedicated one-hop point-to-point connections between cards on the "motherboard".

expansion cards are:
cpu with memory
video
legacy IDE, SCSI, etc.
LAN
IEEE 1394b connection to another "motherboard"

this way one could utilize each expansion bay to whatever was most appropriate--hardcore computing: where one has most bays filled with cpus and memory; IO for interfacing with disk arrays for databases; etc.

a computer becomes a more abstracted resource, much more easily configurable to what needs one has. i guess i heard INTEL showed off 3GIO or something recently that is something like this, but i don't really know.

WHAT DO YOU ALL THINK? this seems totally cool to me. and smart. yeah, harder for legacy support, but it's not like that can't be put onto an expansion card. and the software for such a thing seems straightforward.

i guess i don't understand why we don't have this now.

COMMENTs?

 

[zsouthboy]

Sounds nice... but.... how would we overclock then? 😉

Also, how are you going to make an N-way switch that would still be able to be addressed by software, e.g. to tell it to give more time to graphics, less to peripherals..... you see what I'm saying.......

zs

 

[mooshi]

<< Also, how are you going to make an N-way switch that would still be able to be addressed by software, e.g. to tell it to give more time to graphics, less to peripherals..... you see what I'm saying....... >>



see, that's the thing--the motherboard wouldn't need control from software--it would work much more like an ethernet switch where all it should have to care to look up is location information (cpu0 wants to talk to cpu1--lookup cpu1 and direct to it). since each expansion card has like N-1 controllers (like 1394b as i'd said) for N-way connectivity, it would be determined in the hardware logic just before it goes to one of those controllers. and in 1394b routing is done within the protocol.

i don't know that one would even need a BIOS, except maybe something minimal on expansions with the cpu/memory.

i guess the this all just abstracts a computer up one level to something more of a distributed system over ethernet--each expansion card is like a specialized computer: some are for number crunching, some are for data archiving, some are for connectivity to other networks, etc.



<< Sounds nice... but.... how would we overclock then? >>



hehe. that'd have to be done on individual cpu expansion cards. it's all still there. it's all good! 😉

and since it could all be hot-swap, if you crash your new hammer from OCing it, just pop it out and reinsert--no reboot (assuming you've already got another cpu running the os).

 

[spazntwich1]

I've thought of something similar to this myself, and I agree with you in that it would be VERY cool if it was well implemented.

Thing is, we might need to wait for buses that can transfer at speeds greater than 1.6 gbytes/sec before this could be pulled off and still have good performance.

 

[zsouthboy]

<< I've thought of something similar to this myself, and I agree with you in that it would be VERY cool if it was well implemented.

Thing is, we might need to wait for buses that can transfer at speeds greater than 1.6 gbytes/sec before this could be pulled off and still have good performance. >>



Yeah, that's the problem. We'd need like 20 gigs... Think about, even a geforce4 TI 4600 sometimes chokes on 11 gigs...

I wonder what pm thinks of the idea...

Zs

 

[Smags87]

the idea is good, but i think the cpu MUST be on board as well as the northbridge and possibly southbridge

i mean do we really want to put all that load on bios with even slower transfer rates?? i think not...