22 Apr
2024
22 Apr
'24
1:42 p.m.
On Apr 22, 2024, at 4:21 PM, Chuck Guzis via cctalk
<cctalk(a)classiccmp.org> wrote:
On 4/22/24 13:02, Wayne S wrote:
Not so much "to arrive at the same time" but rather "to arrive at the
correct time". And not so much to reduce chip wait times, because for the most part
that machine doesn't wait for things. Instead, it relies on predictable timing, so
that an action set in motion is known to deliver its result at a specific later time, and
when that signal arrives there will be some element accepting it right then.
A nice example is the exchange jump instruction processing, which fires off a bunch of
memory read/restore operations and sends off current register values across the various
memory buses. The memory read completes and sends off the result, then 100 ns or so later
the register value shows up and is inserted into the write data path of the memory to
complete the core memory full cycle. (So it isn't a read/restore actually, but raher
a "read/replace".)
Another example is the PPU "barrel" which books like Thornton's show as a
passive thing except at the "slot" where the arithmetic lives. In reality,
about 6 positions before the slot the current memory address (PC or current operand) is
handed off to the memory so that just before that PP rotates to the slot the read data
will be available to it. And then the output of the slot becomes the restore, or update,
data for the write part of the memory full cycle.
I read somewhere that the cable lengths were
expressly engineered to provide that signals arrived to chips at nearly the same time so
as to reduce chip “wait” times and provide more speed.
That certainly was true for the 6600. My unit manager, fresh out of
UofMinn had his first job with CDC, measuring wire loops on the first
6600 to which Seymour had attached tags that said "tune". But then, take a gander at a modern notherboard and
the lengths (sic) to
which the designers have routed the traces so that timing works.
Indeed, and with multi-Gb/s interfaces this stuff really matters. Enough so that high end
processors document the wire lengths inside the package, so that "match interconnect
lengths" doesn't mean "match etch lengths" but rather "match etch
plus in-package lengths".
The mind boggles at the high end.... FPGAs with dozens of interfaces running at data rates
up to 58 Gb/s.
paul