defining the main clock frequency of AT91 boards

[Björn König]

Bernd Walter wrote:
> On Mon, Mar 17, 2008 at 06:48:05PM +0100, Björn König wrote:
>> I think users will notice very quickly if they run their board with the
>> wrong frequency, because they won't get output on the serial console -
>> maybe already too late. ;-)
>
> Console speed is about MCK.
> Inside the kernel the xtal Clock is only used to setup PLLB for USB,
> so it is just USB, which is not working.

The PLLA clock is generated from the main clock. PLLA is the source for
PCK and MCK. Therefore a different quartz have influence over the console
speed.

> I was thinking about network performance.
> Core can only be faster with higher PCK if running from cache, otherwise
> it is slowed down by MCK based things.
> I was hoping to get more speed for routing, but maybe the network code
> is working too much inside caches.
>
> What kind of application did you try with the higher MCK?

Everything but network I/O. :-P - I'll check this later again.

> What kind of RAM settings are you talking about?
> Almost all SDRAM on such boards is 133MHz (worst case I would expect
> 100MHz), so no need to slow memory settings down for just 80MHz.
> And I don't see any reason to even add further waitstates, since 133MHz
> SDRAM should be good for at least 100MHz without additional waitsates.
> The only parameter you can tune is refresh, since the refresh frequency
> doesn't need to increase as well, but I doubt that this would make more
> than an academic difference in speed.

I'm talking about all parameters that you can tune in the SDRAMC
configuration register, i.e. RAS, RCD, RP and so on. A higher MCK means an
decreasing clock cycle length, so I should increase the delay parameters.
I make an example: the active to precharge delay (RAS) requires to be at
least 44 ns for my memory module. The cycle length of a 60 MHz MCK tick is
16.66 ns. That means the delay has to be at least 3 clock cycles. With a
80 MHz MCK I have to increase it to 4 clock cycles to use the memory
modules within the specifications. With a theoretical MCK of 133 MHz it
has to be at least 6 cycles.

Björn