DSP chips wondering

[philos]

I'm not programmer, be clear, so it's only speculation.

A friend linked the following paper, saying the TI arquiteture could be (much) more powerful and cheaper than the Analog Devices Tiger Shark line used in AxeFx line:

http://www.ti.com/lit/wp/sprabn8a/sprabn8a.pdf

I'm not trying to "teach the priest" (as we say in Brazil), just wondering if Cliff and the developing team are considering some change in that way and build more powerful and cheaper products.

Thank you in advance for any response.

 

[Rotti]

Search bar
http://forum.fractalaudio.com/lounge/46281-new-ti-multicore-dsps.html

 

[iaresee]

Also: TI wants to sell you their chips, not Analog Devices so...grain of salt and all that when you read their material.

 

[philos]

Thank you for the responses. I forgot to search, and is very interesting the discussion dates two years old!

Someone linked a paper where the TI DSP has 3x the processing power for the same price in a benchmark...(in spite of the manufacturer biasing, 300% is a lot...). The argument a member used that Analog Devices are "audio specific" are not true. There's audio specific products in both companies (as premium operational amplifiers), but that's not the case with these DSP's.

Maybe the point is that, in the latter dual DSP model, the AxeFx hardware development became far less important than the software counterpart (the evolution along the AxeFX firmwares and the great Lexicon and Eventide algorithms of the past, even with 80's hardware limitations, adds to this). The hardware can do more and more.

If Eventide's actual DSP (in most audio products, in different quantities, if I understood correctly) is "4 times more powerful than H3000's one" as a developer said in H9 forum, it is faaaaar less powerful than these ones we're talking about here. The algorithms are developed along decades, starting with old chips, so they become very efficient. There's no reason to think Fractal future hardware and software will not follow a similar path (firmwares upgrades confirm this).

My point focus the price question, that is a key decision factor for many when (re)building a rig, including myself. With a single chip being more powerful than 2 Tiger Sharks at the same price, there's nothing to loose, and significantly cheaper (supposing the portability is possible).

All this, of course, if the TI marketing is not lying too much...

 

[AlbertA]

This is supposed to be an independent benchmark (I no longer work for TI BTW )

http://www.bdti.com/MyBDTI/bdtimark/chip_float_scores.pdf

I think realistically we are probably taking about 2x increase for 32-bit floating point performance (per each core). We need to remember the TigerSHARC design is SEVERAL years old.

Of course, this would be a complete redesign of the hardware board for Fractal.

 

[loopie]

This is supposed to be an independent benchmark (I no longer work for TI BTW )

http://www.bdti.com/MyBDTI/bdtimark/chip_float_scores.pdf

I think realistically we are probably taking about 2x increase for 32-bit floating point performance (per each core). We need to remember the TigerSHARC design is SEVERAL years old.

Of course, this would be a complete redesign of the hardware board for Fractal.

I would take this benchmark and throw it to the bottom of the trash pile.
Why would any "independent" create a bench mark using an benchmark application that is 14 year old, comparing a new
specific purpose processor from TI with a Intel Pentium III?
Why didn't they include some new GPU processors in the comparison? Those things blow everything out of the water.


If alarm bells aren't going off, then something is wrong.
A big reason why the Intel processors are not WAY faster is because they have done everything they can to keep backward compatibility.
The TI's are not backward compatible with anything.
The initial cost to build your "device" is way more expensive than maintaining it.
Do you see any future plans from TI on how they are going to maintain, upgrade the processor?
Starting from scratch could be a multiple year process and suck all the money out of your company.
Take a look at what happened with "Marshall Amplifiers". Someone sucked all the money out and killed the company trying to develop something else.

We have a much better chance of Analog Devices coming out with a faster processor that is instruction set compatible.

 

[philos]

Thank you again for the responses.

I'm no engineer (although have some basic knowledge of analog electronics), but, looking the IT market since early 90's and knowing the difference between the compatibility path and dedicated path in processors architecture (in the old days a IBM RISC used to smash the Intel processors with much higher clock), I understand the choice restrictions for real time processing, and the optimization level in the code (that difiicults the portability).

What I don't knew the TI DSP's were completely new and the compatibility with Sharks is approaching zero...

But the "wandering" remains in the sense Fractal was reach by a bottleneck with the "static" Tiger Shark development (not accompanied of prices cut). I remember Cliff talking about the superiority of these kind of DSP's for the task over PC CPU's (when people asked for a PC plugin), but don't know if a GPU could be used (would be almost "endless power", as you well observed, btw).

In the actual state, fine tune the code appears to be the proved best bet, but what I told about Eventide is, with the company growing, new products, etc. a dedicated chip (even from AD, perhaps) could be an option, because of lower cost (after R&D stage) and better scalability and code reuse independent of standard product lines (again, like Eventide and Lexicon have doing for decades with great success, as everybody knows).

Again. I'm justing speculating. It's not a criticize by any means. I used to prefer the AD275 Opamp over the TI/Burr Brown "better specs" equivalent, by the way . But the digital domain is totally different and the DSP doesn't have a hardware "tone" or harmonic distortion "signature" by it's own. It's all in the code and all the value of the chip is in the power/price/implementation equation.

 

[AlbertA]

I would take this benchmark and throw it to the bottom of the trash pile.
Why would any "independent" create a bench mark using an benchmark application that is 14 year old, comparing a new
specific purpose processor from TI with a Intel Pentium III?

What according to you is wrong with their benchmark? Just it's age? Have you read what the benchmarks actually are? They are typical operations anyone doing DSP would use and why it's useful for comparison.

Why didn't they include some new GPU processors in the comparison? Those things blow everything out of the water.

Depends on the application. It wouldn't blow them out of the water on TDP and latency both of which are criteria for a number of products. A GPU is not an architecture I would choose for something like the Axe-FX. About the only thing you could use all that massively parallel hardware would be convolution reverb with long impulses.

Also BDTI is a business, offering bench-marking services among other things. I'm assuming you can probably contract them to run GPU benchmarks if you need them.

If alarm bells aren't going off, then something is wrong.

huh?

A big reason why the Intel processors are not WAY faster is because they have done everything they can to keep backward compatibility.

Well sure they designed that ISA and the general consumer market demands backward compatiblity (see past Xeon flops). But the other reason is their architecture is driven by "general compute" requirements not simply digital signal processing requirements. Though their vector engines are good now (see AVX2) and their TDP is starting to come closer to that of DSPs (Haswell-ULT at 15W Broadwell-Y around 5W but yet to be released).

The TI's are not backward compatible with anything.

Entirely different ISA of course. You can say the same about Analog Devices, Motorola DSPs, CEVIA, etc, etc... Just like ARM is also not compatible with x86. Or how SHARC is not compatible with TigerSHARC.

Within an ISA, at least with TI there is binary backward compatibility and some if migrating to a newer family and there's definitely source compatibility.

The initial cost to build your "device" is way more expensive than maintaining it.

I would agree.

Do you see any future plans from TI on how they are going to maintain, upgrade the processor?

Those are discussions you have with your sales representative. But given their track record so far (last family has had a run of more 15+ years and still going) maintenance has not been an issue. I don't see how this would be any different than dealing with any supplier.

Starting from scratch could be a multiple year process and suck all the money out of your company.
Take a look at what happened with "Marshall Amplifiers". Someone sucked all the money out and killed the company trying to develop something else.

Sure. However I've seen small companies that have switched platforms very smoothly and others that haven't for a variety of reasons. But yes it's a significant effort.

We have a much better chance of Analog Devices coming out with a faster processor that is instruction set compatible.

Much better chance than what? Fractal switching to another DSP platform? I would have no idea - but given Analog Devices financial statements for the last few years there's no indication of investment in high end DSP I wouldn't' bet on that.