Zorro III speed (asynchronous vs synchrous etc..) 50mb/s ?!

[Karlos]

Jose wrote:

Cool avatar:-)Is that Beavis on the right side?  

Yes, its beavis and butthead rocking away. The original image was:

[NightShade737]

Just checking, but you definately means Mb and not MB?

[Karlos]

NightShade737 wrote:
Just checking, but you definately means Mb and not MB?

I'm pretty sure people here mean 2**10 bytes :-)

[patrik]

If anyone is interested in the details of the ZorroIII bus, the ZorroIII specification covers pretty much all of it.

Have fun!


/Patrik

[Cass]

I've first read about the benchmark program here. My results are the following:

640*400@24 bits

Read RAM : 47274.53 K/sec
Write RAM : 46382.56 K/sec
RAM->RAM : 22630.56 K/sec
RAM->RAM[16] : 26444.66 K/sec
Read VRAM : 3220.04 K/sec
Write VRAM : 7284.17 K/sec
RAM->VRAM : 7065.75 K/sec
RAM->VRAM[16] : 6811.73 K/sec
VRAM->RAM : 2841.36 K/sec
VRAM->RAM[16] : 3157.89 K/sec
Conversion : 21386.14 K/sec [output bandwidth]

Conversion attained 302.67% copy speed

1024*768@16 bits

Read RAM : 47571.85 K/sec
Write RAM : 47164.18 K/sec
RAM->RAM : 22111.55 K/sec
RAM->RAM[16] : 26444.66 K/sec
Read VRAM : 2985.07 K/sec
Write VRAM : 7330.83 K/sec
RAM->VRAM : 7024.39 K/sec
RAM->VRAM[16] : 6923.08 K/sec
VRAM->RAM : 2772.64 K/sec
VRAM->RAM[16] : 3027.75 K/sec
Conversion : 7072.69 K/sec [output bandwidth]

Conversion attained 100.69% copy speed

All this started from a real problem: I run Earth 2140 and noted jerky scrolling. I thought that is impossible to have in market s/w that can't be used on the h/w it's supposed to work on.

If I calculated correctly, for a 16bp 800x600 @ 30fps you'll need about 30MB/sec!
________
Vapolution

[Karlos]

Please ignore the conversion speed results for this test - there's clearly a bug in the calculation for 24-bit pixels (forgot to divide by bytes per pixel there, hence 3x expected). Besides, only the read / write / copy to and from VRAM give the actual bus performance estimates.

@Cass,

Well, that figure you calculate for 800x600x16 assumes that you are basically copying every frame from fast ram to video ram, which I strongly doubt. Only movie players would need to do this.

I would expect any well written 2D tile based games to cache background stuff as bitmaps (where transparency is not an issue) and only refresh the areas of the display where stuff happens.

Even scrolling (without using a superbitmap approach) should only amount to a blit of the remaining visible background area and then an update of all the smaller areas that have changed / been uncovered. I've experimented with this basic idea on the bvision and can confirm that its fast, even on 040.

-edit-

Back on the bus speed issue, here are some results for the local bus on a BlizzPPC 060 @ 50MHz + BVisionPPC:

Code: [Select]

[size=x-small]
Surface width: 640, height: 480, modulus: 0

Surface hwWidth: 640, hwHeight: 480
Test data pixel format

Bytes   : 2, endian native
Bits    : A[  0] R[  5] G[  6] B[  5]
Offsets : A[  0] R[ 11] G[  5] B[  0]
Maxima  : A[  0] R[ 31] G[ 63] B[ 31]

Window pixel format

Bytes   : 2, endian native
Bits    : A[  0] R[  5] G[  6] B[  5]
Offsets : A[  0] R[ 11] G[  5] B[  0]
Maxima  : A[  0] R[ 31] G[ 63] B[ 31]

Read RAM      : 38652.13 K/sec
Write RAM     : 37944.66 K/sec
RAM->RAM      : 16699.80 K/sec
RAM->RAM[16]  : 21980.19 K/sec
Read VRAM     : 5708.84 K/sec
Write VRAM    : 17526.77 K/sec
RAM->VRAM     : 12143.51 K/sec
RAM->VRAM[16] : 11964.10 K/sec
VRAM->RAM     : 4606.52 K/sec
Conversion    : 12121.21 K/sec [output bandwidth]

Conversion attained  99.81% copy speed
[/size]

[Cass]

@Karlos
Actually, that's what I thought in the first place, but I had my doubts after the poster's suggestion and the facts (the game itself)... :-?

Anyhow, the 7MB/sec it's pityful and I don't know if it's idiopathic, or there something that can be done to improve it (I've already have a BUSter v11).

________
Pissing fetish

[Crumb]

@Cass:

7MB/s? It sounds as if you were using a CV64/3D... real CV64s are faster than that (CV64/3D is a cheap design done quickly, just compare the number of components)... I got more than 9MB/s with my Picasso4 and it has a slow z3 bus...

It's a pity that no one has released an updated (and fixed) buster... it would make normal Z3 cards and PCI solutions much faster...

[darksun9210]

i also seem to remember is that the A4000 is based on the A3000 design which in turn is based on supporting 030 chips and bus bursts etc. to from zorro space.

the reason you get good bench results with an 030, is that the bus interface was designed for an 030.
the commodore 040 card didn't support bursting modes, or various other things, so the worst CPU card to bench with is the A3640. the main changes in the rev11 buster over the rev9 were to support the crappy bus signal from the A3640 card.

phase 5 built their cards based on this design and had some interface logic between the 040/060/PPC chip and 030 bus to properly support the various burst modes etc.
from what i remember, this was the major update between the cyberstorm1 and cyberstorm2 series logic, and continued into the cyberstorm3/PPC range.

not sure about any other accelerator manufacturers, but this can easily be found in the speed the CPU can access chip ram.
with my EC030, cybII040, and cybIII060/PPC it was around 7Mb/s, but with my Warpengine040 (28Mhz overclocked to 40Mhz! yehaa! big fat heatsink & fan, attached with heat xfer compound and superglue - FPU thought it didn't exist tho it still worked(?) i was looking at about 2.5Mb/s
(buster rev9 and rev11)

so from my expirience, the best cards to bench with are either 030's or phase5 designed cards on A4000's

[Jose]

Do you think it would be expensive to get someone to do a new proper buster chip? Or anyone left that could do it, for that matter..  :-(

[Orjan]

Buster rev.12... That would be cool...