Shouldn't mem transfers for Mediators (orother brands) for A4000 be faster the equivalent A1200's??

[Karlos]

Hi,

The information I got back indicated that CPU driven transfers from fast ram to the gfx memory were not especially faster in the Mediator4000 than the Mediator1200.

The test program in question copied data to the gfx card using a loop 16x unrolled 32-bit copy loop on cache aligned source data (16-byte) so in theory should have be best case copy.

My miggy is down just now so I'm fishing out info from the replies I got in my email

As a rough guide (all tests at 1024x768 16-bit)

My machine:

A1200T
BlizzPPC with 040@25MHz
64Mb 60ns EDO ram
BVision / CGX4.2
result : 8874 K/s

Now, having fished through them I found a typical (comparable) 4000/1200 mediator pair (same gfx card, same rtg, same mem speed)

Mediator systems (both using 060 + voodoo 3 3000)

Amiga 4000 (Mediator 4000)
Cyberstorm MkIII 68060@50Mhz
80Mb Cyberstorm RAM (60ns)
8Mb motherboard fastram
Voodoo 3 3000, Avenger, 16Mb / Picasso96 v2.1
result : 7947 K/s

A1200T (Mediator 1200)
Blizz 060@66MHz
128Mb 60ns EDO ram
Voodoo 3 3000 16Mb / Picasso96 v2.1
result : 8290 K/s

Ok, the 1200 system is 16MHz faster CPU but the results are similar for all three systems.

I guess there are many factors but the results I got were pretty consistent with those above.

Probably the A4K at 66MHz would win here : looking at the figures - a crude estimate based on the 16Mhz increase clockspeed would be in the order of 10.5 M/s.

What I discovered is that the G-Rex systems had considerably faster bus speed results, averaging 20M/s. One system (A4k 060@66 IIRC) managed close to 40M/s in that test...

[Karlos]

8Mb limit for A1200? Never heard of that one before :-) Mind you I might not have paid attention.

True or not, I'd assume the amount of accessible gfx memory would have no meaningful impact on the transfer speed - thats purely down to the bus.

Also note, in relation to my tests,  we are talking how quickly we can copy data from fast ram into the gfx card's memory, using the CPU to do the copy.

There was no indication in the results that I got from the volunteers (cheers folks) that the Mediator4000 was much faster than the Mediator1200 in this respect.

As for the G-Rex, I agree. It outpaced the mediator in this area consistently.

As for actual graphics peformance overall, the impact of all of this is relatively minor.

For 2D and 3D alike most operations happen locally on the card. The only time the bus speed comes into it is when data has to be written to the card (texture map uploading or geometry data) or retrieved from it (reading back pixels etc).

[Karlos]

Rumour or not, one lucky G-Rex 4000 owner managed it - he got about 30M/s for that test.

-edit-

Aha - found it. He got 33 M/s copy and 39M/s set (which is basically CPU register -> memory rather than memory -> memory)

-end edit-

He had :

CSPPC with 060 @ 66MHz, GRex 4000 and a Voodoo 3 (IIRC)

For the record, most of the G-Rex results I got back were typically 13-15 M/s. That's faster than the Mediator.

However, unless you own a BPPC/CSPPC, the G-Rex not an option.

Jose's original question was about Mediator1200 v Mediator4000 - I didn't observe a signifigant difference in performance between the two.

[Karlos]

tjaoz wrote:
Mediator 1200 bustest is from 8 to 18MB/s. It depends on the turbo card.

I'm not saying youre wrong - I'm sure that what you say is correct. However, were talking specifically about gfx cards operating under PCI here, not just the pure bus speed, but how fast can we shove data to the card

IIRC, Jose was thinking about some sort of movie player code and copy bandwidth issues a while ago - I'm guessing his original question here may be something to do with that.

Quite a lot of brave folks ran my test programs and for a wide spread of Mediator based systems I found that the speed of CPU access to  video ram on a PCI gfx card (usually voodoo) was not very different accross the range 040@25 to 060 @ 66 for both the Mediator1200 and 4000 versions.

The slowest was about 6.5M/s, the fastest was about 10M/s.

I don't know about BusTest, but the copy code I used was my own asm routines that are typically faster than CopyMemQuick(), using 16x unrolled 32-bit, aligned transfers.

[Karlos]

@Jose

I realise all this might have put you off so I'll give you a better perspective :-)

Dont let the bandwidth issue desuade you. The AGA bandwidth (max 7M/s) drops dramatically with the depth/size of the screen. With my BVision I still get 8-10M/s even at 1600x1200 @ 75Hz - I imagine PCI cards on the mediator have the same performance.

Just to reiterate about graphics performance :

Karlos wrote:

As for actual graphics peformance overall, the impact of all of this is relatively minor.

For 2D and 3D alike most operations happen locally on the card. The only time the bus speed comes into it is when data has to be written to the card (texture map uploading or geometry data) or retrieved from it (reading back pixels etc).

In short a good graphics card slaughters AGA in basic performance - even if the raw bus speed doesnt look many times higher.

Refresh rates, colour depth, blitting, rasterizing are all accelerated far beyond AGAs abilities.