Interesting old GPU architecture

[MrSpiggott]

Not sure if this is the correct forum, but I was having a clear out today and found a card that collectors of old and unusual GPU technologies might find interesting.




It's a PCHolly5 made by Quantel from a design originating in 2000 and was used for multi-resolution video editing.
This particular model is fitted with two picstore modules which were large video data processors. What differentiates it from modern GPUs is that it didn't work in polygons, but rather worked on the whole image, whatever the resolution happened to be, in one go.
Each module had 8 maths processors (for sizing, rotating, scaling etc), 2 matrix processors (for colour correction, keying etc) and 2GB RAM connected over a 256bit data bus.
The main PCHolly5 card had a fibre connection to six SCSI video disks (5+parity) and an audio disk.
It required a dual socket Xeon server Windows system with a PCI-X interface and a powerful PSU, I think I've got a picture somewhere.

The product line was gradually phased out as consumer GPUs became powerful enough to do the heavy lifting at a fraction of the cost, but remains an interesting bit of history.

As a sad reminder of how Moore's law used to work. Here's a functionally identical, single PicstoreE card, complete with its required 4 daughter modules, from just 3 years earlier.

 

[Albuquerque]

Wow, amazing sample! And yeah, PCI-X was the hot stuff back then -- 64 bits, 66MHz of PCI awesomeness, those were the interface cards of champions and they were quite few and far between, especially for anything that wasn't specifically an I/O card (storage primarily, some really top-end network cards too.)

What OS flavors were supported? What kind of software did it support, or did it come with its own boutique software to match?

 

[MrSpiggott]

Wow, amazing sample! And yeah, PCI-X was the hot stuff back then -- 64 bits, 66MHz of PCI awesomeness, those were the interface cards of champions and they were quite few and far between, especially for anything that wasn't specifically an I/O card (storage primarily, some really top-end network cards too.)

What OS flavors were supported? What kind of software did it support, or did it come with its own boutique software to match?

It originally required windowsNT, but I think the last versions went out with windows 8. It definitely wasn't directX compliant and required dedicated software to run.
Unfortunately even PCI-X wasn't fast enough to handle the amount of data needed which is why there was a dedicated ebod disk interface on the card.

Found a photo of an old system. Ignore the hole, that wasn't standard.

 

[Davros]

I certainly remember the days when to merely watch a dvd you needed a separate card

 

[MrSpiggott]

I certainly remember the days when to merely watch a dvd you needed a separate card

The Creative Labs PC-DVD card? I had one of them hooked up to a Trinitron SD TV over SVideo. Pity it didn't do RGB, but was still a huge improvement over VHS and was the only way to play imported/different region DVDs with a little software hack. I think the second revision of the card put a stop to that.

On a similar note, this is what was need to edit SD DVCPro 50Mb content in real time without too much generation loss back in the early days.

A lot of the national broadcasters and sports channels used them for editing DVCPro encoded material from video servers.
It's much simpler now.

 

[Albuquerque]

Are these all from your own collection? If so, it's quite an amazing collection indeed! Thanks for posting them up!

 

[homerdog]

What do all those FPGAs do? On the Picstore card I think I counted 30 of them

On the PCHolly did they make an ASIC or is the newer FPGA just that much better?

BTW none of this stuff has any presence at all on google so it's much appreciated that you've posted this.

 

[Davros]

The Creative Labs PC-DVD card

Yeah that was certainly one of them, the minimum spec for dvd was a pentium 2 300mhz and when people had that or greater mpeg cards became not needed anymore.
heres another

 

[Albuquerque]

Yes! I do remember the Reel Magic decoder card from one of my various PC magazines "Back In The Day" Look at that ISA-powered bad boy, it's probably a foot long lol...

 

[MrSpiggott]

Are these all from your own collection? If so, it's quite an amazing collection indeed! Thanks for posting them up!

They’re all cards I’ve worked on over the years. That’s one of my test crates. I cut a hole in the side so I could use a scope probe and logic analyser. I couldn’t bear to thrown them in a scrap bin
They are important pieces of history that could so easily be lost. Large broadcasters tend to decommission and throw in the bin or auction for scrap once they’ve served their purpose.

 

[MrSpiggott]

What do all those FPGAs do? On the Picstore card I think I counted 30 of them

On the PCHolly did they make an ASIC or is the newer FPGA just that much better?

BTW none of this stuff has any presence at all on google so it's much appreciated that you've posted this.

FPGAs are fantastic. They’re programmable logic that give you amazing hardware flexibility. You can add features and change functionality without having to replace hardware. These were premium systems, so features and flexibility trumped component costs.
I think at one point, Quantel was the biggest Xilinx customer in the country.

 

[Davros]

FPGAs are fantastic

Isnt that what the Chromatic Mpact graphics card had or was that a dsp

 

[MrSpiggott]

Isnt that what the Chromatic Mpact graphics card had or was that a dsp

I don't think a single FPGA from around 97 would contain enough logic cells to make a GPU. Wasn't it programmable DSP based? Almost a software approach, a bit like what Nuon did on their aborted console/DVD player.

 

[MrSpiggott]

Edit. Found an article...

https://vintage3d.org/chromatic.php#sthash.55vMBZUD.dpbs

 

[Arun]

Edit. Found an article...

https://vintage3d.org/chromatic.php#sthash.55vMBZUD.dpbs

Really interesting article (as is the rest of this thread!) but that looks like a complete change in direction from version 1 to 2... the "6th ALU pipeline" sounds like possibly like an entire 1997-era fixed-function GPU to me! What does the rest of the DSP even do for 3D workloads then? Coordinate memory accesses? Triangle setup and emulating missing features? Or are there more fundamental things left to the DSP that they're not talking about?

Overall it sadly sounds like a very "NVIDIA NV1" strategy of doing everything and the kitchen sink in a single product (although with a DSP rather than fixed-function logic) and then realising there isn't really a market for it and you're better off only being best-of-class in 2D/3D.

If there was a right era for this kind of programmable architecture, it was probably Larrabee, which I still think is sad that Intel screwed up so bad - it would never have been the best 3D accelerator, but with more engineering resources resulting in an incrementally better architecture and especially 32nm rather than 45nm, I think it'd have been a decent first step. In the 1996/1997-era though, using a DSP for a 2D/3D GPU is absolutely pointless, and the key benefit was sharing functionality with video/audio/etc...

This all reminds me of all the focus in the late 2000s in the mobile market to try to do video acceleration on mobile GPUs like PowerVR SGX and ARM/Falanx Mali etc... They were not great at it (worse than DSPs, much worse than fixed-function) but much better than CPUs so the argument was it could be a cost saving for low-end/mid-range mobile SoCs - unfortunately, "better than CPUs" still wasn't good enough usually, and I could be wrong but I'm not aware of any SoC/phone vendor actually shipping that in production.

EDIT: Trying to do video acceleration on mobile GPUs for cost reasons is kind of the opposite of those massive FPGA-based video processors in this thread, where the focus is clearly feature set & performance rather than cost! But in both cases, it highlights the struggle between programmability & fixed-function in video decode/encode/processing accelerators... For what it's worth, PowerVR's VXD/VXE in the 2000s/2010s were a great example of fixed-function video decode/encode accelerators sharing functionality across formats. But fixed-function (and ASICs vs FPGAs) takes longer to implement than using mostly/partly DSP-based solutions, so PowerVR was late to HEVC/H.265 and the video business never recovered from that. Of course, fixed-function vs programmability is far from the only reason PowerVR struggled with execution and focus in the 2010s, but I digress! All this to say that this kind of niche hardware using mostly FPGAs (or DSPs in other cases?) makes perfect sense for that kind of market where volumes are small but time-to-market is key.

 

[swaaye]

Wow that is a lot of FPGAs!

The '90s had a bunch of programmable media processors. MPACT, Verite, MWave, N64... And how about those USR Courier modems that were software upgradable.

 

[MrSpiggott]

Wow that is a lot of FPGAs!

The '90s had a bunch of programmable media processors. MPACT, Verite, MWave, N64... And how about those USR Courier modems that were software upgradable.

This isn't PC related, so a little off the topic, but since we're talking about FPGA powered video processors you might find this interesting. It's an Image card from 94 and has the largest number of FPGAs I've come across. I think it's 100 of the really early Altera Flex devices, but my eyes cross when I start counting. That's a Transputer CPU bottom left, an integer divider middle left and the 3 ceramic devices on the right are digital image filters.
Acknowledgements. I don't have one of these cards any more, so photo taken from https://dexterslab2013.blogspot.com/2018/03/quantel-hal-video-compositing-system.html

 

[Ailuros]

Very interesting thread; only in cases like this it's easier to see how much technology has advanced in the last couple (or more) decades. In the above past they needed almost an entire "motherboard" for something that takes today probably only part of a square millimeter in an SoC.