> On Jul 14, 2015, at 4:41 PM, Chuck Guzis <[email protected]> wrote:
>
> On 07/14/2015 10:29 AM, Paul Koning wrote:
>
>> The accuracy of the FPGA depends on the approach. If it’s a
>> structural (gate level) model, it is as accurate as the schematics
>> you’re working from. And as I mentioned, that accuracy is quite
>> good; it lets you see obscure details that are not documented and
>> certainly not visible in a software simulator. The example I like to
>> point to is the 6000 property that you can figure out a PPU 0 hard
>> loop by doing a deadstart dump and looking for an unexpected zero in
>> the instruction area: deadstart writes a zero where the P register
>> points at that time. But you won’t find that documented or explained
>> anywhere. The FPGA model derived from the schematics reproduces this
>> behavior, and when you look at how it happens, the explanation
>> becomes blindingly obvious. *This* is why I feel there’s a point in
>> doing this sort of work.
>
> I can agree with some points, but not others. In the 6600, for example,
> clock distribution was a big design issue--not so in an FPGA. You had racks
> of taper-pin mats of wiring between the cordwood moules extending over (by
> today's standards) long distances. Cooling was a huge issue. In those
> respects, an FPGA hardly resembles a "real" 6600.
Certainly, the physical aspects are completely different. And clock
distribution, certainly. Not so much between chassis, interestingly enough,
but throughout the logic within a chassis. And wire delays in chassis to
chassis cabling are very significant. Wire delays within the chassis, in a few
cases, but 99% of the wires are short enough that their delay is not a real
consideration in comparison with the logic circuit stage delay.
The example I gave, and others like it, are properties of the detailed logic
design, they are not dependent on the details of the timing closure, or the
physical design of the machine.
paul
