Fractal Talk

They are sub optimal for sure but the development rate of arm cores is so fast they are leaving everything else in the dust from a price to performance point of view. You still have to rebuild the libraries of algorithms to take advantage of it, but it's a one time cost. It will 100% happen, DSPs were the result of optimizing for limited processing powers but nobody is investing much on that anymore because simply isn't needed now. General purpose processors are way faster now and they keep getting better.
ARMs are great for general-purpose processing because you can distribute the processing over multiple cores.

However, for DSP, CISC will ALWAYS be better. The 'R' in ARM stands for Reduced Instruction Set Computer (RISC). The benefit of RISC is a smaller die size which means less expensive. The downside of RISC is that more instructions are needed to do a given task. This means more memory is required to store the program and the given task takes longer to complete. In the early days of RISC the memory requirements, and additional cost of that memory, outweighed the cost advantage of the processor itself.

As memory has become less expensive RISC processors are now cost-competitive and even beating CISC processors due to the reduced power and thermal management expense.

For real-time Digital Signal Processing, however, nothing beats CISC. This is because you can do more in a single clock cycle and, hence, more throughput for a given latency. For example, to find the maximum of two values is a single instruction in most CISC DSPs. This is a common operation in signal processing. With a RISC processor you need, say, 10 instructions to do the same thing. Furthermore the increased number of instructions required to do the same tasks means cache misses are more frequent thereby slowing things down even further. You end up having to access external memory much more often than with a CISC processor so then you end up being bottlenecked by the DRAM speed and latency.

It is untrue that nobody is investing in DSP anymore. TI just released their C7x architecture which builds on the C6x architecture (used in the Axe-Fx III BTW) and adds single-cycle vector operations as well as new instructions and other enhancements. Analog Devices continues to develop the SHARC core. They also have other things in the pipeline that I'm not at liberty to divulge due to NDA. Furthermore there are DSP IP cores that are available for custom solutions. Cadence has a licensable DSP core that is often mated to an ARM (or other RISC processor) on a custom chip.

ARMs continue to get faster and more powerful but they are doing that by adding instructions and coprocessors which is increasing the die size and obviating the whole reason RISC was created in the first place. These latest "RISC" processors have huge dies and require large heat sinks or even active cooling. Yes, they are still smaller than a CISC die but the line between RISC and CISC is getting blurry.

This isn't to say you can't do DSP on a RISC processor. Several products are using RISC processors: Tone-Master Pro, UA pedals, Headrush. In all those products, however, I measure very high levels of aliasing. The logical conclusion is that these products are using low amounts of oversampling due to the lack of processor throughput.

As with all engineering problems understanding the technology is key to making an informed design decision. There are pros and cons to each. You have to weigh cost vs. performance. If you prioritize performance then a dedicated DSP is the superior solution. If you prioritize cost or cooling requirements then a RISC chip may be the superior solution.
 
boom smile GIF
 

Axe-Fx III Firmware Release Notes

25.00

New “Cygnus X-3” amp modeling:

Improved preamp algorithms.
Improved power amp algorithms.
Improved output transformer loss modeling.
Improved preamp supply voltage modeling.
Nearly all amp models have been “remastered” as a result of new measurement techniques and analytical methods.
As a consequence of the new modeling many default parameter values have been changed. Existing presets are automatically updated to the new parameter values upon recall. The updated parameters are:

Preamp Low Cut Freq
Preamp Hi Cut Freq
Triode1/2 Plate Frequency
All Cathode Follower parameters
Preamp Bias
Preamp Bias Excursion
PI Bias Excursion
Power Tube Bias Excursion
Transformer LF/HF
Transformer Drive
Power Tube Grid Bias
Cathode Resistance
Cathode Time Constant
Negative Feedback
Supply Sag
Added “Sunrise Splendor” Drive model. Existing model renamed “Sunrise Splendor Hi-Cut” to denote that the Hi-Cut switch is engaged.

Added “Class-A 30W Brilliant” Amp model.

Added “Kill Dry” control to Delay, Multitap Delay, Pitch, Plex and Reverb blocks. When set to ON the dry signal is muted. This simplifies routing these blocks in parallel. When using these blocks in parallel the user can turn Kill Dry on and set the wet level using the Mix control, leaving the Level control at 0dB. This also has the advantage that the wet level is controlled by the global Reverb/Effects Mix, if desired.

Improved Input block Noise Gate.

Improved Classic Expander and Modern Expander types in Gate block.

Fixed wrong Miller capacitance value in Deluxe Verb Vibrato, Double Verb Vibrato, Super Verb Vibrato and all Vibrato Verb models.

Fixed wrong FX loop recovery gain in Suhr Badger models.

Fixed global settings not recalled if MIDI data is present on the MIDI In jack during bootup.

Fixed wrong Overdrive Volume taper in ODS-100 HRM, ODS-100 Ford and Bludojai Lead amp models. Existing presets should be auditioned and adjusted as necessary.

Various other fixes and improvements.
 
Kill dry is super handy

Is this to get around the volume differences when putting post-amp blocks series or parallel?

Every few weeks I find myself dicking around with that aspect, figuring out which I prefer and I can never really come to a conclusion after tweaking the Mix/Level/Input Gain, I can get what I want out of both situations, I just don’t dig the volume drops when putting things parallel in comparison to other presets.
 
Is this to get around the volume differences when putting post-amp blocks series or parallel?

Every few weeks I find myself dicking around with that aspect, figuring out which I prefer and I can never really come to a conclusion after tweaking the Mix/Level/Input Gain, I can get what I want out of both situations, I just don’t dig the volume drops when putting things parallel in comparison to other presets.
It just allows you to adjust fx level with the mix knob even when routed in parallel. Without kill dry you need to set mix to 100% and use the fx output level as a "mix" control to not screw up your dry level.

This will also allow you to use global delay/reverb mix controls even when using parallel routing.
 
ARMs are great for general-purpose processing because you can distribute the processing over multiple cores.

However, for DSP, CISC will ALWAYS be better. The 'R' in ARM stands for Reduced Instruction Set Computer (RISC). The benefit of RISC is a smaller die size which means less expensive. The downside of RISC is that more instructions are needed to do a given task. This means more memory is required to store the program and the given task takes longer to complete. In the early days of RISC the memory requirements, and additional cost of that memory, outweighed the cost advantage of the processor itself.

As memory has become less expensive RISC processors are now cost-competitive and even beating CISC processors due to the reduced power and thermal management expense.

For real-time Digital Signal Processing, however, nothing beats CISC. This is because you can do more in a single clock cycle and, hence, more throughput for a given latency. For example, to find the maximum of two values is a single instruction in most CISC DSPs. This is a common operation in signal processing. With a RISC processor you need, say, 10 instructions to do the same thing. Furthermore the increased number of instructions required to do the same tasks means cache misses are more frequent thereby slowing things down even further. You end up having to access external memory much more often than with a CISC processor so then you end up being bottlenecked by the DRAM speed and latency.

It is untrue that nobody is investing in DSP anymore. TI just released their C7x architecture which builds on the C6x architecture (used in the Axe-Fx III BTW) and adds single-cycle vector operations as well as new instructions and other enhancements. Analog Devices continues to develop the SHARC core. They also have other things in the pipeline that I'm not at liberty to divulge due to NDA. Furthermore there are DSP IP cores that are available for custom solutions. Cadence has a licensable DSP core that is often mated to an ARM (or other RISC processor) on a custom chip.

ARMs continue to get faster and more powerful but they are doing that by adding instructions and coprocessors which is increasing the die size and obviating the whole reason RISC was created in the first place. These latest "RISC" processors have huge dies and require large heat sinks or even active cooling. Yes, they are still smaller than a CISC die but the line between RISC and CISC is getting blurry.

This isn't to say you can't do DSP on a RISC processor. Several products are using RISC processors: Tone-Master Pro, UA pedals, Headrush. In all those products, however, I measure very high levels of aliasing. The logical conclusion is that these products are using low amounts of oversampling due to the lack of processor throughput.

As with all engineering problems understanding the technology is key to making an informed design decision. There are pros and cons to each. You have to weigh cost vs. performance. If you prioritize performance then a dedicated DSP is the superior solution. If you prioritize cost or cooling requirements then a RISC chip may be the superior solution.

Class Soldier GIF by U.S. Army
 
ARMs are great for general-purpose processing because you can distribute the processing over multiple cores.

However, for DSP, CISC will ALWAYS be better. The 'R' in ARM stands for Reduced Instruction Set Computer (RISC). The benefit of RISC is a smaller die size which means less expensive. The downside of RISC is that more instructions are needed to do a given task. This means more memory is required to store the program and the given task takes longer to complete. In the early days of RISC the memory requirements, and additional cost of that memory, outweighed the cost advantage of the processor itself.

As memory has become less expensive RISC processors are now cost-competitive and even beating CISC processors due to the reduced power and thermal management expense.

For real-time Digital Signal Processing, however, nothing beats CISC. This is because you can do more in a single clock cycle and, hence, more throughput for a given latency. For example, to find the maximum of two values is a single instruction in most CISC DSPs. This is a common operation in signal processing. With a RISC processor you need, say, 10 instructions to do the same thing. Furthermore the increased number of instructions required to do the same tasks means cache misses are more frequent thereby slowing things down even further. You end up having to access external memory much more often than with a CISC processor so then you end up being bottlenecked by the DRAM speed and latency.

It is untrue that nobody is investing in DSP anymore. TI just released their C7x architecture which builds on the C6x architecture (used in the Axe-Fx III BTW) and adds single-cycle vector operations as well as new instructions and other enhancements. Analog Devices continues to develop the SHARC core. They also have other things in the pipeline that I'm not at liberty to divulge due to NDA. Furthermore there are DSP IP cores that are available for custom solutions. Cadence has a licensable DSP core that is often mated to an ARM (or other RISC processor) on a custom chip.

ARMs continue to get faster and more powerful but they are doing that by adding instructions and coprocessors which is increasing the die size and obviating the whole reason RISC was created in the first place. These latest "RISC" processors have huge dies and require large heat sinks or even active cooling. Yes, they are still smaller than a CISC die but the line between RISC and CISC is getting blurry.

This isn't to say you can't do DSP on a RISC processor. Several products are using RISC processors: Tone-Master Pro, UA pedals, Headrush. In all those products, however, I measure very high levels of aliasing. The logical conclusion is that these products are using low amounts of oversampling due to the lack of processor throughput.

As with all engineering problems understanding the technology is key to making an informed design decision. There are pros and cons to each. You have to weigh cost vs. performance. If you prioritize performance then a dedicated DSP is the superior solution. If you prioritize cost or cooling requirements then a RISC chip may be the superior solution.
What do you think of the developments with RISC-V? Seems like it's super early days but the idea of a royalty free, open architecture is super fascinating to me.
 
Back
Top