SD Card Write Problem

I have an Atmega328p chip running at 5V internal 8 MHz. During my tests powered via USB 5V from a USB-TTLSerial adapter. Code is compiled for 8MHz and works well, regarding speed of millis().
I have an SD-Card Adapter with a 3.3V regulator and do signal level shifting from 5V to 3.3V for MOSI, CLK and CS.
They are connected on a breadboard, max 2-3" (<10cm) wires. The 5V power has 47µF + 100nF caps.

I'm working with small FAT16 formatted SD Cards.
Reading works well, but when writing, the cards behave different: One works fine, another one does not: Write always fails.
After running into a write error, I have problems getting that card to be readable again. ( A reset helps usually )
Without writing, reading ia always fine.
I tried with low level access (directly reading/writing empty sectors) using Hardware SPI ( HALF-SPEED ). There's nothing else connected to the SPI lines.

All the cards are doing fine in my computer.

Question:
What's more likely : speed issues, power supply issues, signal issues ?

It's perhaps those three all together: speed issues, power supply issues, signal issues.

Do you use a breadboard for the signal level conversion : bad connections.
What is the brand of the card, does it use more power than the others ?

What actual speed are you running the SPI at? How are you doing the level shifting?

@Krodal:
Brand : TakeMS 2GB ( some years old, though )
Power consumption: not sure how to measure the actually required power spikes during write. I don't have a real oscilloscope.
Is analogRead fast enough ? If yes, should I use a second arduino to measure, or dig into how to do this interrupt driven.
Bad breadboard connections: I understand both reading and writing uses all SPI signals, and reading works well...
@MarkT:
Speed: I'm testing based on the Arduino SD library and tried both
card.init(SPI_HALF_SPEED, chipSelect); // as used in CardInfo.ino example
and SPI_QUARTER_SPEED. Is it possible to be too slow with my 8MHz board ?
LevelShifting : I understand both reading and writing uses all SPI signals, and reading works well...
Is there power drawn from the MOSI or CLK line ?
The adapter provides 10k pullup resistors between its 3.3V regulator and the signal lines,
so I just added small reverse diodes to my 5V atmega signals, providing 0.6 .. 3.3 V signal level to the SD card.

Thanks to both for your feedback so far.
More hints ( especially regarding live power measuring, and whether this makes sense ) are very welcome.

My next step will be to cross over the available components
A ) Arduino Uno with Adafruit DataLogger shield ( working )
B ) Standalone 8MHz with LCstudio SD card module ( read only )
and see if problem is rather the 8MHz no crystal or the SD card module side

Instead of analogRead() with a second Arduino, I sometimes use a fast optocoupler (with a resistor for the internal led) to the mic input of my computer to measure the noise of a power supply.

The 47uF could be bad. You could place a second capactitor of about 100uF ... 470uF. It might seem odd, but I've seen the craziest unexplainable things with a bad capacitor.
I found a post that is simular: Arduino Forum

What about this : http://www.adafruit.com/forums/viewtopic.php?f=31&p=146648
It says: "I have heard of issues with accessing SD cards on 8MHz variants"
So the 8MHz might be a problem. But I can't find a good explanation what the problem is with 8MHz.

Thanks, seems an interesting link:

A quote from fa16lib himself:

I find that most older cards work at 8 MHz and almost all older cards work at half speed, 4 MHz.

Many newer cards fail at all speeds with resistor level shifters. These cards expect very fast SPI signal rise times. The SD spec requires a rise time of less than 10 ns for all clock speeds under 25 MHz. It is just not possible to get sharp edges at 8 MHz into the typical capacitance with resistor level shifters.

hi...

I was wondering if you were running a +3.3v logic / 8MHz internal clock Arduino circuit...

could you just -not- use the (resistor based) voltage divider and 'direct connect' (so to speak)?

I am trying, and I am not having success.

I have a minimal/breadboard 'Arduino' running @ +3.3v / 8Mhz internal clock (no crystal/caps)..

successfully flashed bootloader....sketch uploading works every time as well.

With no resistors to stop a 'sharp edge'/fast SPI signal..... is there another problem?

thanks...

Thanks for this late response.
Currently my atmega328 runs at 5V, but this is not necessary with 8MHz.

An interesting alternative approach. Thanks !

I just wonder if the 5V->3.3V regulator found on my SD-Card module would be sufficient to supply the atmega328 (plus even a small nokia 5110 ---
except for the backlight).

Up to now the balance between "get it to run safely" vs. "minimal build" removed the external crystal and relies on a 5V supply only.

The 3.3V for the nokia vcc currently come from a "pseudo regulator" built by a led and a 1k resistor.
Pulling it from the SD-Module was possible as well, but I have the power indicator led anyway.
I would not dare use that to drive my atmega328, with probably > 20mA, though.