Pi Pico 2 / RP2350

[Nathaniel Lewis]

Hey all,

Didn't see anyone discussing it on the list yet, but Raspberry Pi has launched their second microcontroller, the RP2350, along with the Pi Pico 2.

https://www.raspberrypi.com/products/raspberry-pi-pico-2/

From skimming the datasheet, it seems to address a significant amount of the complaints about the RP2040. Probably the three largest improvements are the addition of a proper low power sleep state, a general reduction in power draw, and that the ADC's non-linearity problems have been fixed. The effective number of bits has also been improved (9.5 versus 8.7)

Other things of note to me:

- The cores have also been upgraded from 133 MHz Cortex-M0+ cores to 150 MHz Cortex-M33 cores. These have an FPU, multiply-accumulate, and integer SIMD capability. You can also switch either or both cores for RISC-V cores (Hazard3, which seems less capable than the ARM cores, but interesting nonetheless).

- Internal SRAM has been increased to 520 KiB (up from 264 KiB)

- PSRAM support, so you can have up to 16 MiB of external RAM (admittedly very slow compared to the internal memory).

- An additional PIO block (3 blocks x 4 state machines for 12 PIO state machines).

- Two package sizes - the large one has 48 GPIO and 8 ADC channels.

- Both of those have the option for 2 MiB of internal QSPI flash ($0.20 more expensive per package) called the "RP2354".

If you make your own boards, Pi is still only pricing it at $0.80/part for the base model, and $1.10/part for the largest one.

- Nathan

[Chris Albertson]

This is good news. With floating point, they might edge out the ESP32.

I don’t understand about switching cores to RISC-V but it could be good for AI-type projects as RISC-V is inherently good at vector instructions and modern AI is nothing but vector math.   But it all depends on the details

BTW there are RP2040 boards on AliExpress with 16MB ROMs.  They are cheap so I thought I might buy some but the new RP2350 seems much better.

[Chris Albertson]

> On Aug 8, 2024, at 8:46 PM, Chris Albertson <alberts...@gmail.com> wrote:
>
> This is good news. With floating point, they might edge out the ESP32.
>
> I don’t understand about switching cores to RISC-V but it could be good for AI-type projects as RISC-V is inherently good at vector instructions and modern AI is nothing but vector math. But it all depends on the details

It seems the cores switch at boot time. But the RISC-V cores lack the floating point unit so they are not so powerful at FP math as the M3 cores and it seems you can not use both at once core selection happens at boot time.

Most users will program the PR2350 using either Micro Python of Arduino IDE and both of those hide details like M3 vs RISC-V core from the user. You would need to be programming in assembly language to notice the difference.

BTW, I really recommend Python over Arduino, you gain a lot of functionality, like easy access to true multitasking and asynchronous IO that is hard to do with Arduino. Python is not slow. It can even be compiled if you like. And Multitasking is built into the Python language. There is no GIL.

[A J]

This looks like a nice upgrade over the M0. The M33 is supposed to be about 20% quicker than the M4.

The M33 is is also supposed to be more energy efficient than the M0. Good code on a M7 is supposed to

run about twice as fast compared to M4.

A large piece of SRAM could be good to AI models. I wonder we could put multiple chips on a board to chain

models.

I am wondering if the Python API will be extended to add more of the chips capabilities.

[Chris Albertson]

One more feature I found.   The cores are individually selectable between M3 and disc-v so in theory you could run one of each.  But it is not clelar why you’d want to use the risk-v core at all, other then for developing risk-v code.  We will have to wait for the chip to ship to see if maybe the risk-v core uses less power or something.

The M33 is is also supposed to be more energy efficient than the M0. Good code on a M7 is supposed to

run about twice as fast compared to M4.

I read in a spec sheet that deep sleep mode can get to below 100 uA.   But there are not many hands-on reviews yet.  The chip is yet to ship

A large piece of SRAM could be good to AI models. I wonder we could put multiple chips on a board to chain

models.

You are never going to run an LMM, even a reduced-size LLM on this chip.  It is orders of magnitude too small.   Even object recognition is not going to run on this.  I think a tiny CNN is about the limit.

I am wondering if the Python API will be extended to add more of the chips capabilities.

I read that there is already a port of Python but some things remain to be finished such as support of the risc-v cores.  So yes the work is ongoing but already usable.

I will spend $10 and buy a couple of them as soon as they are shipping.   In the meantime they do have a 1,300 page document that describes how this chip works in some detail but een in 1,300 pages many things are not covered.

[Nathan Lewis]

Another interesting thing I noticed is that the inputs are 5V tolerant.

Obviously you can't drive anything expecting 5V CMOS voltage levels (TTL inputs would be fine), but it does save you from needing to put level shifters between your uC and all those 5V quadrature encoders.

- Nathan

On Fri, Aug 9, 2024, at 7:42 PM, A J wrote:

It looks like the chips could be out by years end.

I used a TI M4 board with open source arduino code support and it runs fast compared to a uno.

https://www.electronicsweekly.com/news/products/micros/rp2350-available-by-year-end-with-flash-inclusive-option-2024-08/

The M33 does have a single precision FPU and DSP for signal processing. 

--

You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.

To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/hbrobotics/acbcc8df-c531-4b3f-8e52-da8524e4b0acn%40googlegroups.com.

[Chris Albertson]

The other good thing about Pico compared to the old Arduino or even TI’s M4 boards is that Pico costs $4 and is a TINY chip.  $4 is so close to free that there is no practical difference.

[ste...@uppump.com]

I appreciate everyone's comments about the RP2350.  Its been very interesting since we are exploring using it in our ISS Space program at Valley Christian High School in San Jose and nationally/internationally through our educational non-profit The Quest Institute (www.Q4excellence.com).  We've been running this program for over 10 years that allows students to build custom experiments and send them to the ISS.  The program has slowly evolved over the decade and its now time for another upgrade.  My team is looking at using the RP2350 as the MCU that students will use to control their custom experiments.  We've recently been using the Arduino Mega but the RP2350 seems to be leaps and bounds faster with way more memory.  Until the RP2350 ships, we are experimenting with the RP2040.

One other thing we are looking to do is develop experiment templates.  These would be semi-assembled, yet customization, experiment units that would allow those students more interested in the science to focus on the science instead of the engineering. 

I mention this only because I know there a lot of smart people in this group and maybe a few of you would be interested in being apart of this program in some way.  Whether it be helping design a experiment template that gets send to the ISS or mentoring a student team building their own experiment.

Feel free to reach out if you have any interested or questions.

Thanks,

Stephen Huber

Director of R&D, Valley Christian Schools

Program Director, The Quest Institute

shu...@thequestinstitute.com