exxos's Atari, Amiga & retro forum

I've just changed the flags on Gembench 6 to use fast RAM and the speed boost is enormous, including ROM and normal RAM which really shouldn't be affected if the benchmark was purely measuring ST_RAM speed and not have its own bottleneck.

But, wow! P.S. D-oh! I forgot that the ROM was an image in fast RAM. Stuuupid! :-)

stephen_usher wrote: 30 Oct 2020 11:40 I've just changed the flags on Gembench 6 to use fast RAM and the speed boost is enormous, including ROM and normal RAM which really shouldn't be affected if the benchmark was purely measuring ST_RAM speed and not have its own bottleneck.

It is because the routines themselves are now running in fast ram as opposed ST ram, so you get a apparent speed boost in ram speed. ST-RAM itself isn't running faster, but the routines are because they are now running in fast-ram.

Well, I swapped out my board and fitted a stock mc68000 into the motherboard socket and ran Gembench...

100% scores all around.

Not only this but demos don't glitch.

So, I have a clock problem, unless it's due to the noise levels still being too high.

Tomorrow I'll do an experiment with the Mk.I board to see what that does.

I'm thinking more and more that my problems are being caused by decoupling caps which are an order of magnitude too small and hence not filtering the power spikes. So, another order to RS has been placed for 0805s as recommended in another thread.

The closest caps I had in stock were 1uF tantallums, so I hacked these on as decoupling caps and... no difference. The ground is still spikey even at the CPU end of a thick wire which comes directly from the GND pin on the connector, so that should be as low impedence as it gets.

Well, the signal clock looks OK:
The ground looks awfully spiky though, as this is directly at the far end of the thick (24AWG?) wire directly from the ground pin.
You can see the glitches this is producing in the following video, I think it's causing the CPU to miss individual clock pulses when there's a spike.

https://youtu.be/4FCRZda4OBo

The new capacitors haven't made any difference to the GB6 scores.

However, it did allow the floppy of Carrier Command to load its splash image correctly and load the music, which it wasn't yesterday. The glitches in the demo are still apparent.

Putting a 'scope on the CPU and the clock source show no measurable skew. There's a (very) tiny voltage drop in the amplitude but otherwise they're identical.

When the new caps arrive on Tuesday I'll build a new board up. The current one is a bit too hacked up now and I need to find out what's endemic to the board and what's being caused by my hacks. I can also see if the noise on the board is less due to better decoupling of the other components, which I can see are generating significant noise on the ground.

Have you tried moving the power supply out the way or shielding the accelerator from it? I wonder if EMI isn't a factor now that the TF card is close to the transformer...

dalek wrote: 01 Nov 2020 11:26 Have you tried moving the power supply out the way or shielding the accelerator from it? I wonder if EMI isn't a factor now that the TF card is close to the transformer...

On my H4 I have the TF536 right next to the psu without any problems.

dalek wrote: 01 Nov 2020 11:26 Have you tried moving the power supply out the way or shielding the accelerator from it? I wonder if EMI isn't a factor now that the TF card is close to the transformer...

Without the CPU running (press reset button) the ground is as flat as a pancake. All the noise is being generated by the CPU itself (with a small amount from the auxiliary chips).

Similarly, if the system crashes the ground goes quiet.