Some Thoughts on Media Co-Processors

[Jason Dunn]

I receive an anonymous message from someone named "John" who had some interesting thoughts on media co-processors in Pocket PCs. Take it away John...<br /><br />"First of all, I am a hardware engineer. Recently NeoMagic annouced a new chip with APA(Associative Processor Array) architecture. This technology is absolutely fasinating. For PDAs without dedicated MPEG4 decoding hardware, CPU handles the job. It is slow even at 400Mhz! High clock speed burns a lot of power. The HP 5xxx PDAs has the MediaQ processor in it which has dedicated MPEG4 hardware decoder, so performance is better, and power is also reduced. However, APA can do much better. APA chips can run at 20Mhz for MPEG4 decoding and better performance. That saves a lot of power!<br /><br />What does APA architecture look like? It is a expandable(scalable) 2-dimensional array of processing elements. Each processing elements contain memory as well as computing hardware! To process an image, just send it to the APA arrays, and simultaneously instruct every processing element process every bit of image data. Compare that to a CPU or a dedicated hardware datapath, you would need to send blocks and blocks of image data through the CPU/hardware datapath. To get good performance you need to run it at high clock speed. APA architecture gives you massively parallel processing. <br /><br />What's more. NeoMagic's APA chip already has MPEG4 encoding algorithm implemented as well. (MediaQ does have MPEG4 encoding hardware yet). You can also implement other algorithms such as speech recognition on APA. That's how flexible APA is. No additional dedicated hardware is required. APA is perfect for PDA, Cellphone devices. We will soon be able to have real-time 30fps TWO way video communications on our cellphones/PDAs with this technology. Hopefully this will give your reader some ideas what's to come in future portable devices.<br /><br />You can learn more about the APA technology <a href="http://www.neomagic.com/product/APA_version3_1.pdf">in this white paper</a>."

[ozen]

sounds cool, but when can we see this kind of thing in everyday devices like cell phones and pda's? I would like to hear more about this and everything that it can do.

[rosettaZ]

TI OMAP 1510 has a very impressive DSP with beyond elegant memory balancing. But it's almost impossible to program with normal tool. In the case of Palm it ended p to be a total waste of silicon.

Xscale has some media acceleration code in the chip too, but hardly anybody uses it because the tool is not easily accessible.

First generation ATI chip are dubbed video decelerator because of extremely bad driver and memory caching.

Now we want to talk about even fancier CPU structure. I would like to see how good driver, tool and total implementation first before believing anyting.

So what if you can design and print a nice chip. Will it pump up quake fps? will it spew more .mpeg fps? show me working PDA prototype, instead of white paper.

[felixdd]

First generation ATI chip are dubbed video decelerator because of extremely bad driver and memory caching.


ATI is famous for bad drivers :lol:

[Kati Compton]

The white paper link doesn't seem to be working...

[lurch]

I receive an anonymous message from someone named "John"
Um, if he told you his name, it's not really anonymous is it??

[Enderet]

Maybe we will see this in the new 600 mhz ppcs...... 0X

[Janak Parekh]

I receive an anonymous message from someone named "John"
Um, if he told you his name, it's not really anonymous is it??
Well, John could mean John Doe. John is as anonymous as names get, at least in this country... ;)

--janak

[ctmagnus]

The white paper link doesn't seem to be working...

Replace the %20 with an underscore, like http://www.neomagic.com/product/APA_version3_1.pdf

[Janak Parekh]

Replace the %20 with an underscore, like http://www.neomagic.com/product/APA_version3_1.pdf
Thanks for figuring it out! I've fixed it.

--janak

[Talon]

Two things jump out at me.

1) The products currently avalible are replacements for the xscale not co-processors. Good if they can get their foot in the door but it does make it harder to get a design win.

2) No mention of PPC only WinCE. Since CE work there is probably no technical reason for PPC not working but without MS buy in they are going to have trouble signing up any PDA manufacturers.