DIY RISC-V Ultracluster

[iMo]

[brucehoult]

see also https://www.eevblog.com/forum/projects/risc-v-ultrascalar-from-bitluni/

[betocool]

I'd very interested in the software. From what I understand so far each MCU drives an LED and receives SPI command from a master MCU. And there is a master MCU per blade?

Still, very funny and interesting.

Cheers,

Alberto

[brucehoult]

I'd very interested in the software. From what I understand so far each MCU drives an LED and receives SPI command from a master MCU. And there is a master MCU per blade?

With 12KB RAM 256KB flash and a 100 MIPS 32 bit CPU you could do a bit more than just be a dumb SPI slave accepting RGB commands!

[SiliconWizard]

That's a fun project but it has little practical value outside of showing off a "large" project IMHO. The bottleneck is of course communication between the MCUs, and those MCUs are very limited in the kind of interfaces they can use.  I think he used CH570 MCUs? The fastest bus between them will probably be SPI.. And attempting to scale up efficiently any kind of distributed load on this constrained throughput with this many MCUs is almost guaranteed to end before it started. Driving a few LEDs per MCU is probably about it at this point.

I think we discussed this idea before with CH32V003 or similar - kinda same conclusion. I would personally be more interested in this kind of architecture with significantly beefier MCUs that can communicate via a very fast bus.

Staying with WCH, their beefiest (so far) CH32H417 looks like an excellent candidate. It's only about $4, it has 896 KB RAM, runs at 400 MHz (+ a second core), and has an embedded serdes (full duplex link at up to 1.5 Gbps) that could make a pretty fast bidirectional link between MCUs. Just musing about that.

[Xiamatsu]

That's not just blinking the LED 

[Xiamatsu]

N32G430F8Q7 (QFN20)
Cortex-M4F  (DSP,FPU)
128MHz, 64K Flash, 16K RAM

At a price starting from $0.3 (1000 units) is very interesting in this case.
There is also tssop20, qfn28, qfn32 from $0.45

Or  AT32F421  Cortex-M4  (DSP, not FPU)
120MHz,  64K Flash, 16K RAM

tssop20, qfn28  from $0.47  (not  qfn20)

The AT32F422  has been announced and even the QFN32 is already on sale from $0.5
( 180MHz, 128K Flash, 20K RAM )

[ftg]

I think price is not an issue, as he got the MCU's for free from WCH.
The CH570 is an interesting device in general as it has USB, 2.4GHz radio, tons of flash and even an integrated LDO, at a sub-20cent price point in reasonable volume.

[dxl]

Reminds me of the Transputer area end of the 80ies... Which just died because communication between Transputers was slow and programming paradigm was totally different compared to "normal" mcus.

[tszaboo]

While it looks interesting, I don't think there is an application for it.
The bandwidth of the I/O is limited.
If you would place the clusters in series, and you would connect the micros for example to another 4 micro, then maybe there would be an application for it.

If I would need to burn a few reels of these ICs, I would definitely turn it into a phased array antenna. It already has RF an interface. Turn maybe 128 chips into RR transmitters, 128 into receivers, all the rest in between to calculate beamforming angles and do all the signal calculations.

[brucehoult]

I think price is not an issue, as he got the MCU's for free from WCH.

And if not, then paid for by Altium anyway.

[MadTux]

It already has RF an interface. Turn maybe 128 chips into RR transmitters, 128 into receivers, all the rest in between to calculate beamforming angles and do all the signal calculations.

Highly doubtful, you could actually control phase angle on RF signals on these controllers.
Would need a dedicated PLL synthesizer chip for that.

Apart, quite a waste of resources, you could dump quite a few CPUs into a large FPGA and have decent interconnection fabric at decent speed, latency....
And build something actually useful, like GPGPU or just for the fun of getting cache coherency and everything working across a lot of cores.
That thing is like calling my home network a cluster, because everything is LAN'd together and you can ssh between them....

[ftg]

While it looks interesting, I don't think there is an application for it.
The bandwidth of the I/O is limited.
If you would place the clusters in series, and you would connect the micros for example to another 4 micro, then maybe there would be an application for it.

If I would need to burn a few reels of these ICs, I would definitely turn it into a phased array antenna. It already has RF an interface. Turn maybe 128 chips into RR transmitters, 128 into receivers, all the rest in between to calculate beamforming angles and do all the signal calculations.

How would you get them phase coherent?
In the ESP32 based 2.4GHz "RF camera" the channel information from the wifi modem is used to correct the phase offsets between the radios.
The CH570 has a much more primitive radio that only does GFSK with BLE-like characteristics.

[tszaboo]

While it looks interesting, I don't think there is an application for it.
The bandwidth of the I/O is limited.
If you would place the clusters in series, and you would connect the micros for example to another 4 micro, then maybe there would be an application for it.

If I would need to burn a few reels of these ICs, I would definitely turn it into a phased array antenna. It already has RF an interface. Turn maybe 128 chips into RR transmitters, 128 into receivers, all the rest in between to calculate beamforming angles and do all the signal calculations.

How would you get them phase coherent?
In the ESP32 based 2.4GHz "RF camera" the channel information from the wifi modem is used to correct the phase offsets between the radios.
The CH570 has a much more primitive radio that only does GFSK with BLE-like characteristics.

I didn't think about is so much. For getting it coherent, clearly the design would need to have a symmetric/uniform layout. They would need to send the same message, with the same low level/ASM code sending the message. I guess the phase shift would be with varicap diodes driven with different PWM signals, providing variable phase shift.
Maybe it's not possible. But I think figuring out this would be more interesting challenge than just connecting a bunch of these to some SPI bus to drive an LED. Even driving them from the same clock wasn't solved here, while that likely would just need a few logic gate buffers and resistive dividers.