[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]


On  7 May 98 at 11:05, Philip Ronzone wrote:


> Sorry that you didn't recognize the tongue-in-cheek paragraph.

O sorry, I didn't put in enough dots... ... ... ... After 
sending I thought "it could come over too serious"...

> Intel had proposed somnething like this for a really cheap
> mostly-software modem -- to which I thought "what a great
> idea. For Intel. Dump a $9 part, and use 80% of the
> CPU cycles of a $350 part."

Well that was a little overblown at that time from Mr
Grove and I think it still is. Nowadays there are 2-4
times faster CPU's. Within 2 generations those old 80%
will be less than 10%. Since most people use a modem not
all the time it will be about 1-5% of a $350 part in the
near furture. (Hehe... I observe lots of people that use
40-80% of their CPU time for MP3 decoding and they have no
problems with the performance penalty!)

However... I bet a big part of the 80% you mentioned was 
caused by the ISA bottleneck and maybe no busmastering 
or DMA...


>   > The trick is to get the data as fast as possible into the 
>   > CPU. Without letting the CPU wait for it or slowing it 
>   > down. So a good UART has a databus interface as wide as 
>   > the CPU databus and busmastering.
> The point is that getting the data into the CPU is only PART
> of the problem. If, after getting the data into the machine,
> the software has to, say, BE/LE swap, and align buffers etc.,
> the advantage of high speed device<->cpu data transfer is
> weakened.

Of course there is some weakening, I only try to state
that a 400MHz 32 or 64 bit CPU loses only one or two
400 MHz cycles to do a BE/LE swap or align buffers. That's 
about 1/10 or less of one 33MHz PCI bus cycle.

A perfect design will always be faster but CPU's are so 
incredible fast these days that functionality that was done 
in hardware in the past can be done in software -cheaper-, 
not faster(!).

If you can get a lower-cost BE than LE part and your design 
is LE it might well be that less than 1% of the cycles of a 
$350 CPU might compensate that. A price that halves every 
18 months.


(Always a bit too serious?)