home *** CD-ROM | disk | FTP | other *** search
-
- > IIRC Motorola claims 6 operations per instruction cycle for the 56k, but
- > then one instruction cycle is two clock cycles, which may well be the case
- > with the TI chip as well.
-
- The 56000/1 Reference claims 7 operations/instruction cycle. This does of course
- depend on how you define an 'operation'. The 56k's main advantage is multiple
- parallel data busses.
-
- Most of the TI DSP / GPU processors I have worked with take either 4 or 2
- clock cycles per instruction cycle - some are even binary compatible with
- the DSP56000. It is of course possible these particular chips are very recent
- and only require a single oscillator cycle per instruction (or less), but more
- details are needed before we can know the answer.
-
- The most common TI DSP is almost identical to the 56k, but each instruction takes
- exactly twice as long to execute. The clock rate is around 50-80Mhz. I can't
- remember the name, but it's something like TMS34010/20.
-
- The reciprocal/sqrt functions sound encouraging - and since it's a DSP specific
- implementation, they are likely to be average/low latency (high for a DSP of
- course!) as these instructions (along with mult/div) normally form the core
- purpose of such chips. Such instructions usually absorb a very large percentage
- of the available silicon in order to produce an effective hardware solution to
- arithmetic and vector calculations. I doubt they will be anywhere near as slow
- as software implementations.
-
- An example is the cascading bitwise root algorithm for the DSP56k, which can
- produce a 24-bit square root in around 24*5 cycles - but a hardware implementation
- can parallelise the 5 intermediate stages and reduce this to 24 cycles. Dual
- execution units can reduce this further to 12 cycles. I imagine these new opcodes
- will be implemented with the same attitude.
-
- > Yes, isn't TI the producer of that monster with multiple DSPs and multiple
- > RISC processors on one chip, capable of executing something like 10^9
- > instructions/second?
-
- 1000 MIPS? I'm sure that will encourage efficient program design. :)
-
- > But how would they connect it?
- > It seems that for many things the bus is the bottleneck even with the
- > relatively slow DSP we have now.
-
- I'd be happy just to see Falcon developers making use of the DSP we already
- have before adding even more. Relatively slow it may be, but it's still much
- more capable than Falcon users seem to think.
-
- It would be great to see people shake this concept of a fancy soundchip and
- start using it to remove bottlenecks. There are a few out there using it for
- a variety of purposes, but most still treat it like an add-on sample player.
-
- > with a Afterburner (the best thing would be to stick this card directly on the
- > Afterburner giving a fast 32-bit bus!!) it will go into warp speed.
-
- > Is there such a bus to use on the AB040?
-
- Yes, there appears to be a 32-bit local data bus in addition to the standard
- expansion slot.
-
- Doug.
-
-