** Audio effect processing on your video card **

[VitaminD]

it doesnt change things.

[Alan]

Care to enlighten me?

The UAD card was revealed long ago to be a video card called Mpact2 and UA themselves have owned up to that fact.I saved that thread long ago so I'll paste Universal Audio - Joe's(Joe Bryan
VP Engineering & Technology
Universal Audio)
reply to the matter(you'll have to dig up the thread yourself).

The DSP is indeed the Mpact2, as anyone with SiSoft's Sandra or a PCI device analyizer already knows. The http://www.haas.com/doc/mpactbkg.htm link posted here describes the Mpact1 processor, not the Mpact2 which is roughly twice as powerful.

Comparing the Mpact2 to a general purpose (GP) processor like the Pentium or PowerPC is difficult because they have profoundly different architectures and design strategies. The closest comparisons are the SSE2 or AltiVec engines in these GP CPUs which are actually derived from design advances orginally used by Chromatic based on super-computer architectures.

These engines are commonly referred to as "vector" units because they process data in blocks rather than one datum at at time, which allows them to save instruction decode cycles. Standard scaler DSP (e.g. 56K) and GP processors must perpetually decode the same instructions over and over when doing the same sequence of operations, and this wastes cycles.

The fundamental differences between the Mpact2 and a GP CPU (even it's vector units) is the amount of parallelism and the available memory bandwidth. The Mpact2 uses an internal 11-way 792-bit bus for moving data around between the DMA engine, PCI bus interface, the instruction cache, the data cache, the 5 execution units, the various peripheral interfaces, and the main dual-spline 1200 MByte/s RDRAM interface. This provides 11GBytes/s of bandwidth. In addition to this ultra-high speed internal data bandwidth, the multiple execution units are parallel instruction SIMD units, which can process up to 8 independent data sets with multiple instructions per clock.

In audio DSP terms, the internal clock speed is indeed 1GHz, and can exceed this speed for certain sub-operations. For example, a floating point multiply/add instruction consists of 2 FLOPs (floating-point operations). The Mpact2 typically provides 2GOPS of execution unit perfomance (and up to 6GOPS for some specialized units), hence the 1GHz processor speed equivalent. However the comparison isn't quite accurate because the Mpact2 can do additional non-FPU operations in parallel (like calculating addresses, dithering, moving data, etc.). In addition, the Mpact2 doesn't have a deep pipeline like the GP CPUs. The Pentium-4 has a 14-clock pipeline that's also data dependent. This means coding audio algorithms like recursive filters (which have high data dependencies) on the Pentium requires lots of waiting around for the pipeline delay.

In practice, the GP CPUs overcome the pipeline delay by allowing new operations to be started while the others are in process, but this has limited applications in audio processing. It's much better suited to calculating spreadsheets, or decoding an MP3 stream or MPEG movie.

Another advantage the Mpact2 has over GP CPUs is the multiple dedicated execution units and the cache arcitecture. The Pentium actually shares the same physical silicon hardware multiplier between the FPU, ALU and vector units, so all multiply instructions must fight over the same hardware. And the over-hyped Hyper-threadding CPUs share this same silicon between multiple thread contexts which pretend to be multiple CPUs, but they're not. You still only have one multiplier unit.

The Mpact2 uses an independent direct-mapped instruction cache, and a separate ld/st data cache with read-ahead, write-behind access. This means the Mpact2's instruction fetch unit doesn't cause random cache-line misses like the GP CPU's multi-way set-associative code/data cache. The Mpact2 compilers also provide control over instruction cache use which avoids cache misses at runtime. The Mpact2 data cache appears as a huge, 9-way ported register set to the execution and bus interface units which eliminates all conflicts between multiple execution units and the "key-holing" effect in GP CPUs' caused by accessing the cache through a tiny register set.

The Mpact2 has a 1200 MByte/s RDRAM interface that has independent, full-speed access to the instruction and data caches, plus the PCI bus at all times. This means nothing needs to wait for anything else in the Mpact2 to access a resource. Everything runs at full speed, all the time. This *never* happens in a GP CPU, becuase they're designed around a statistical model of resource sharing, while the Mpact2 is designed around a deterministic model.

That's one of the main advantages DSPs have over GP CPUs when processing real-time audio. Audio samples come in on a rigid time schedule and can't wait to be processed like video or pretty much any other type of compute intensive operation.

The Mpact2 is not going away anytime soon. We have large inventories, and the chip is available for the customers who are still using it. Incidently, there are DVD decoders that use this chip that are still being made.

The advantages this processor have brought to everyone involved are clear: we get an extraordinary advantage in the marketplace, and our customers get the benefit of this power with the best processing algorithms available.

The transition to other DSP platforms is not a difficult one, as we have shown with our TDM plugs. In some cases, the main difficulty has been optimising the algorithms to fit, but these optimizations can be applied to the UAD versions as well. The forthcoming 3.1 versions of the 1176LN and LA2A are not only better sounding than before, they're more efficient as well.

Regarding third-party support, the DSP coding aspect of plugin development is a small portion of the process, while the business aspects are very important. We would rather spend our company resources developing new algorithms for our customers than supporting a development platform for other companies. That said, there are still opportunities for 3rd party plugin developers, only not at the code porting level right now.

The evolution of Universal Audio has included the incorporation of both Kind of Loud, and Hyperactive Audio Systems, which I started several years ago. We think this is a unique combination of analog, DSP and digital hardware expertise, and we hope you agree.

-Joe

Joe Bryan
VP Engineering & Technology
Universal Audio
Santa Cruz

Hope this is helpful.

[jens]

Notice the similarity in the two cards below?

Not really. Care to enlighten me?

it's virtually the same....

[xg2]

Notice the similarity in the two cards below?

Not really. Care to enlighten me?

They're practically the same card?

[xg2]

If someone writes 7-8 good plugins that use a NVIDIA GPU and sell them all bundled for $100 they will sell alot of them and make some money. If they sell them for $200-800 a piece, a few professionals might buy them, but who else. Sell five hundred at $300 or sell thousands for $75-100.

I totally agree. I just hope that Bionicfx see sense and don't pander to the 'pro' market, which then means that the majority of musicians will never benefit from this technology, or will simply encourage illegal copying.

NVIDIA would make out if they bought out the company and included them with the high-end cards. Some of these companies do not understand the market very well. There are hundreds of thousands of home musicians, who have a tough time forking out 200-300$ for Sonar or Cubase. If my core recording system is $300, what fraction of that is reasonable to spend on effects?

Exactly. They need to get the price right - there is a huge market of home musicians out there that can't afford to pay ludicrous prices for software.

[whyterabbyt]

Considering the number of gamers willing to fork out £400 for the newest-generation graphics card, I have to say that, for the graphics card manufacturers, licensing/developing and bundling audio-processing software for a minority market probably wouldnt make financial sense....

[Amberience]

I don't think thats the idea. It doesn't make cultural sense to give gamers free music software.. let alone financial sense.

The idea is us muso's go out and buy a card able to do the things the technology says it can do, and then we buy the software I suppose.

Then again, BionicFX might end up selling bundles - a card and software for £300 say????

[BlueJay]

Exactly. They need to get the price right - there is a huge market of home musicians out there that can't afford to pay ludicrous prices for software.

From what I gathered on their front page...

Why is this necessary? If you spent $25,000 on a proprietary studio DSP solution, you are probably happy as a clam and don't care about this technology. For everyone else trying to make music on a PC with more than a few tracks and effect plug-ins, then you are well aware of the frustration of never having enough processing power.

...I doubt they are going to put it on the top shelf of the music stores. I would be quite happy with a $100-150(USD) price, and would not think twice about the purchase.

[ObsoleteAcc99]

They're really dropping the ball to try to market a $200-800 card. More intelligent would have been to develop the algo and some plugins so that existing nVidia owners could use it by buying just the software.

Greg

isn't that what is actualy going on? you just buy the sotware and use it on your existing graphics card? (as long as it is nvidia fx series or 6800)...if not then i have misunderstood this whole thread

[ObsoleteAcc99]

A whole PC is WAAYYY cheaper than a UAD1, or a Winfast S800 graphics card from 1997...

You must be joking? oh, you mean pentum 1 100mhz with 16mb sdram, 1 gig hdd, 8xcdROM and 15inch curved crt.

[whyterabbyt]

ttoz quoth

You must be joking? oh, you mean pentum 1 100mhz with 16mb sdram, 1 gig hdd, 8xcdROM and 15inch curved crt.

UAD-1 here : £399

The Athlon XP2500 based system, I just specced with 512Mb memory, and an 80Gb drive : £310

As is normal, its the software, not the hardware, which determines which would be more useful to a given muso...

[mauseoleum]

these things are very new-ish ... uad has been around some time (well obviously)

[ObsoleteAcc99]

ttoz quoth

You must be joking? oh, you mean pentum 1 100mhz with 16mb sdram, 1 gig hdd, 8xcdROM and 15inch curved crt.

UAD-1 here : £399

The Athlon XP2500 based system, I just specced with 512Mb memory, and an 80Gb drive : £310

As is normal, its the software, not the hardware, which determines which would be more useful to a given muso...

pc's are much cheaper your side of the world then that's for sure. although that is a rather basic setup with no frills.

[whyterabbyt]

True, but as a 'horsepower-expander' a la FXTeleport or whatever, it wouldnt need any frills. That figure includes a WinXP license at £80, BTW.

[Lunch Money]

They're really dropping the ball to try to market a $200-800 card. More intelligent would have been to develop the algo and some plugins so that existing nVidia owners could use it by buying just the software.

Greg

isn't that what is actualy going on? you just buy the sotware and use it on your existing graphics card? (as long as it is nvidia fx series or 6800)...if not then i have misunderstood this whole thread

That's what I was thinking all along, too. But then in recent pages, people have been talking about marketing a card bundle a la UAD-1.

f**ked if I know what the story is anymore.



Greg