Picasso 5 AMA

Hi,

I have announced a Zorro graphics card during the Amiga40 show. I'm almost ready to order prototype PCBs (maybe 10 more hours of work for component selection), but I haven't decided on a colour yet.

The (currently) only other Z2/Z3 card we have is the X-Surf-100, which is blue. However, the BigRamPlus (256M memory card) that is going to be replaced by this new card was black. So let's open this AMA ("ask me anything") thread with a question from me to all you interested readers - and I'll answer your questions that you may have looking at the rendered panel.

Jens

Quote from robdangerous

Does this also include the functionality of an Indivision AGA MK3 or maybe integrate with it?

It is not a flicker fixer, but it integrates with all our flicker fixers: You can see a lot of connectors and adapters on the board, and they are inputs for signals that shall pass-through to the monitor when RTG isn't active. This might be easy to miss, as 256MBytes of memory dominate the appearance of the card

Direct connectors are:

10-pin: Indivision ECS V1/V2/V3, Indivision AGA MK1

12-pin: Indivision ECS V4

22-pin: Indivision AGA MK3

The two adapters are for Indivision AGA MK2 and Indivision AGA MK2cr.

Digital signals will be routed to the HDMI output, analogue signals will be routed to the VGA output.

The small connector above the VGA output is an audio input: If you don't have Indivision AGA MK3, then audio won't be in the stream and you may need additional equipment to play Amiga-audio. However, HDMI includes Audio, and I've put an AD converter on the board, so the analogue audio output of the Amiga can be routed to the speakers of a HDMI monitor. This will give you audio even if you "only" have Indivision AGA MK2(cr), and it'll also give you audio during RTG sessions.

Quote from robdangerous

Does it support 1920x1080?

Yes, it will support full-HD for sure. There's also enough memory and bandwidth for 1920x1200, but at some point, you'll have to admit that even the Z3 bus (and the Amiga CPU) is not built to handle that much data for a screen. You really have to consider Z3 to be a bottleneck. However, the whole concept of this card is basically "ACA1240/1260 fast GFX card connected to a BigRamPlus", so it's developed with a lot faster bus access in mind. The next iteration of this card will be the the ACA1240/1260 RTG card.

Jens

Quote from robdangerous

PS: The color poll isn't working for me, can't select anything.

Hmm.. I was wondering what it's still 0 results, but it doesn't take my vote either. First time I'm using a poll with this forum software. We might need to switch to just mentioning the colour in the thread itself.

Jens

Quick sunday check of eMails, and the assembly company gave me info that one of their customers has cancelled a contract, so their machines are free in the first week of December. Now I'm really in a hurry getting the boards to Germany, and any other colour than green will delay this (these are 6-layer boards that take extra time for layer lamination/pressing). I've just placed my order for green boards, so I can take advantage of that free time slot at the assembly company.

These are just prototypes anyway, probably won't end up in anyone's regular computer.

Jens

Just checked what it would take to support Warp3D (actually, I've been doing research since I got your message earlier today):

- The main thing appears to be triangle drawing (edge walking + span filling).

- harddware Z-Buffer. Our plan is to use 32 bit memory for a 24-bit pixel, so if 8 bits are enough for a Z-buffer, that might work. Seems low-ish, though - my guess is that a Z-buffer would have to be 16 bits (like on Voodoo), so high-colour would be the max. we could offer with this concept - at least if we want to have all information in a single data word in memory, which is 32 bits.

- Perspective-correct texture mapping - bilinear doesn't appear to be enough, it really must be 1/Z division per pixel to approximate correct perspective.

- Bilinear texture filtering

- Alpha blending for transparency (this involves transformation into gamma space, blend there, transform back to RGB space)

- Gouraud shading and linear fog appear to be optional. I have no idea how this can be simplified for hardware, or if it somehow mixes nicely with Alpha blending, for example. More reading required - I'm not a 3D expert at all.

Memory-wise, I guess we're good: In theory, we can make all of the 256MB accessible to the GFX core, but since this card shall have a sibling for ACA1240/1260 with just 64MB physical memory, we should put a cap there. Given that the old Voodoo cards had a lot less (mostly 8MB), that's no limit, though. However, the 10k LE FPGA appears a bit small to implement all that. I fear the number of multipliers aren't enough for real-time alpha blending, and especially that 1/Z perspective correction. Only that will cost 800-900LEs, and that's just a single pixel pipeline. The old Voodoo cards implemented 4 pixel pipelines, so we're talking roughly 3500 LEs or more than a third of the chip, just for the pixel pipelines, no alpha blending, no texture filtering and no triangle drawing implemented yet.

The choice of a 10k LE FPGA iss mainly for cost reasons: I would like to be able to offer this card for around 200,- EUR including 19% German VAT, as I believe that high-price items won't really crack critical mass to be accepted by the wider Amiga community.

Unfortunately, we can't just put a larger FPGA into the design, as the larger ones of the same package need more power pins - so this comes at the cost of fewer IO pins. The 256-ball fine line BGA package gives us 176 IO-pins for the 10k LE chip, but only 162 for the 16k LE chip. The Picasso5 design really uses each and every pin of the 176 available, and I've already used tricks like connecting some signals to other chips that provide the information serially to the FPGA.

Then there's the problem of writing a Warp3D driver - who is capable and willing to do that for a "new chip"?

TL;DR: I see a lot of roadblocks, so don't hold your breath

Jens

Like I wrote: It's a stretch to implement this on such a small FPGA. I can totally see a modification of the blitter to also support drawing rectangles and copying a texture to it, but when it comes to perspective correction, Z-buffer and alpha blending, we're quickly exhausting the available FPGA resources.

Warp3D being co-owned by Hyperion and the Friedens is not a problem at all - I have a good relationship with all of them, so getting permission to write a driver with their API is not an issue at all. I might even convince the Friedens to write that driver - if the FPGA core would fit, that is. I'm giving this a less-than-1% chance, and personally, I don't really see a path to success: Even if we could implement some 3D acceleration and Warp3D support, it would likely be 6 man-months of extra work, which easily translates to 44k EUR extra cost, not counting associated employer outlay, taxes and capital cost. If we're aiming high and expect to sell 300 of these cards, this means that the price increases by almost 150 EUR per card - plus VAT, associated employer outlay and capital cost, that is.

With that kind of price increase, we'd lose more buyers than we'd gain from the 3D support. So one more nail in the coffin of this idea: It would completely destroy the idea of bringing RTG "to the masses", and the Picasso5 card would be the same niche product that most RTG cards were back in the days.

The only reason why I can aim for such a low price is that the bulk of the work is already done for Indivision ECS V4, which has a much wider audience. Yes, we need to expand the blitter to use 32-bit memory (thus doubling it's speed), but a lot of the work, especially for testing edge cases, is already done and can be ported from Indivision ECS V4. This would not be the case for anything 3D.

Jens

Quote from NicoPPC

3D support. Yes 🤠.

Please do read closely, as the essence of what I wrote is: I see no way to reach the target price of 200,- EUR for this board if 3D is part of the product. More like twice that, and that's why I think the idea is not worth implementing.

Jens

Quote from NicoPPC

As for the driver, Bigfoot is the best for that but I am not sure it would be interested 🤔.

Just out of curiosity: Is Mark Olsen still active? I mean, it took him about two years for the TinyGL bounty, and there appears to be radio silence since 2024.

Jens

Maybe - if I come across some extra money. Or someone sells me land that's right next to an Autobahn where I can build another solar field (easy permissions for that kind of land!) - that would bring in extra money

Jens