5 ms·

Tinycpm – CP/M ON A TINY 2040

61 points by bubblehack3r 4y ago

rbanffy 4y ago
Worth noting a 2040 is a supercomputer by the standards from when CP/M was first released.
- myself248 4y ago
  I was just pondering this recently. A lot of today's microcontrollers have specs comparable to PCs during the heyday of the demoscene.
  - pjmlp 4y ago
    That is what I find so interesting about the ESP32, although I have zero uses for it. Compare it with an PC 1512.
- guenthert 4y ago
  Well, kinda, sorta, but as mighty as the RP2040 is, it still pales to the Cray-1 of '76 with its 8MiB of RAM and 160MFLOPS. The RP2040 beats the Cray for sure in the MIPS/$ category though :)
  - mk_stjames 4y ago
    Yea, sadly the two things I wish the RP2040 could have- more SRAM (it only has 264kb) and onboard floating point in hardware. Other than that I love them and I am using Picos for experimentation now. The Cray M1 may have only been clocked at 80mhz but the FPU could do a floating point mult + add in parallel making the "160 MFLOPS". It's similar to the performance of an ESP32, I think, which is a great FLOPS/$ comparison. A Cray-1 adjusted for inflation would be $35,000,000. A ESP32-S3 dev board is $6.
    - GeorgeTirebiter 4y ago
      Most embedded chips have much less RAM in the M0+ space; they tend to have more Flash. The interesting choices for the 2040 are: 1. More RAM than usual, using ext SPI Flash to load RAM 2. Fancy programmable peripherals 3. TWO cores 4. Very flexible IO mapping. Others do IO mapping, but it's really nice on 2040 5. "Pretty fast" at 133 MHz and the Total Win (for me) 6. Integrated into Arduino IDE As for Floating Point, there are some great packages out there that are optimized for Cortex-M https://www.quinapalus.com/qfplib-m0-tiny.html https://www.quinapalus.com/qfplib-m0-tiny.html There also Q-notation, which uses integer HW to do a sort-of fixed-point 'float'. We used this extensively in the 2016 Model X body controls, because the falcon wing door was totally controlled without a floating point processor! That was hard to do... https://en.wikipedia.org/wiki/Q_(number_format) https://en.wikipedia.org/wiki/Q_(number_format) Motorola, err, I mean, Freescale provided an excellent library for their automotive-qual PowerPC MPUs.
guenthert 4y ago
Perhaps understood, but since it's not explicitly stated, it might be worth clarifying that this entails a Z80 emulator to run CP/M for i8080. It's not a new CP/M for ARM abomination ;-}
- benj111 4y ago
  Why would that be an abomination. The current setup sounds like more of an abomination.
pjmlp 4y ago
Of course Turbo Pascal had to be part of the testing. :)
reaperducer 4y ago
I like that he used Zork to test his creation. I use the same benchmark. However, in my experience, Zork starting up isn't actually proof of very much. I've had any number of CP/M installations go bad once I got a hundred or so moves into the game. If he can complete Zork on the Tiny 2040, then great. But just starting it is only half the battle.
kragen 4y ago
running cp/m on a 32-bit arm running at 133 megahertz with 264 k of sram, 2-8 megabytes of xip qspi flash, and 4 gigabytes of flash on sd card, is not challenging, though it might be useful i routinely ran cp/m in 48k or 64k on a 4-megahertz z80, which is about 0.6 dhrystone mips, on one or two 100-kilobyte floppies, with transfer rates of about 1 kilobyte per second. i think cp/m will run in 16k this arm is as fast as 250 of those machines, it has as much ram as four of them, as much offboard xip program memory as 64-256 of them, and 40000 sssd floppy disks worth of storage, which furthermore can probably transfer data at 10 megabytes per second, ten thousand times as fast as the floppy also the z80 didn't have a pio programmable i/o pin driver, it had to handle i/o interactions itself instead of outsourcing them to a channel program in pioasm, which slowed down computation yes, if you're emulating a z80 on arm that entails some emulation slowdown, but it certainly isn't 250x. it might be 8x. if you do dynamic machine code translation it might be less than 2x instead of calling it 'tinycpm' he should call it 'giantcpm'
- ideonode 4y ago
  Why not just celebrate it for being neat, rather than berate it for not living up to your expectations? I thougbt it was interesting, and it has inspired me to try something similar on my pico. Especially after rereading Code recently. Oh, and I assume that tinycpm is because it runs on a Tiny 2040.
  - kragen 4y ago
    it looks like a fine project; it just has a deceptive name

61 points by bubblehack3r 4y ago

rbanffy 4y ago

Worth noting a 2040 is a supercomputer by the standards from when CP/M was first released.

myself248 4y ago
I was just pondering this recently. A lot of today's microcontrollers have specs comparable to PCs during the heyday of the demoscene.
- pjmlp 4y ago
  That is what I find so interesting about the ESP32, although I have zero uses for it. Compare it with an PC 1512.
guenthert 4y ago
Well, kinda, sorta, but as mighty as the RP2040 is, it still pales to the Cray-1 of '76 with its 8MiB of RAM and 160MFLOPS. The RP2040 beats the Cray for sure in the MIPS/$ category though :)
- mk_stjames 4y ago
  Yea, sadly the two things I wish the RP2040 could have- more SRAM (it only has 264kb) and onboard floating point in hardware. Other than that I love them and I am using Picos for experimentation now. The Cray M1 may have only been clocked at 80mhz but the FPU could do a floating point mult + add in parallel making the "160 MFLOPS". It's similar to the performance of an ESP32, I think, which is a great FLOPS/$ comparison. A Cray-1 adjusted for inflation would be $35,000,000. A ESP32-S3 dev board is $6.
  - GeorgeTirebiter 4y ago
    Most embedded chips have much less RAM in the M0+ space; they tend to have more Flash. The interesting choices for the 2040 are: 1. More RAM than usual, using ext SPI Flash to load RAM 2. Fancy programmable peripherals 3. TWO cores 4. Very flexible IO mapping. Others do IO mapping, but it's really nice on 2040 5. "Pretty fast" at 133 MHz and the Total Win (for me) 6. Integrated into Arduino IDE As for Floating Point, there are some great packages out there that are optimized for Cortex-M https://www.quinapalus.com/qfplib-m0-tiny.html https://www.quinapalus.com/qfplib-m0-tiny.html There also Q-notation, which uses integer HW to do a sort-of fixed-point 'float'. We used this extensively in the 2016 Model X body controls, because the falcon wing door was totally controlled without a floating point processor! That was hard to do... https://en.wikipedia.org/wiki/Q_(number_format) https://en.wikipedia.org/wiki/Q_(number_format) Motorola, err, I mean, Freescale provided an excellent library for their automotive-qual PowerPC MPUs.

guenthert 4y ago

Perhaps understood, but since it's not explicitly stated, it might be worth clarifying that this entails a Z80 emulator to run CP/M for i8080. It's not a new CP/M for ARM abomination ;-}

benj111 4y ago
Why would that be an abomination. The current setup sounds like more of an abomination.

pjmlp 4y ago

Of course Turbo Pascal had to be part of the testing. :)

reaperducer 4y ago

I like that he used Zork to test his creation. I use the same benchmark. However, in my experience, Zork starting up isn't actually proof of very much. I've had any number of CP/M installations go bad once I got a hundred or so moves into the game. If he can complete Zork on the Tiny 2040, then great. But just starting it is only half the battle.

kragen 4y ago

running cp/m on a 32-bit arm running at 133 megahertz with 264 k of sram, 2-8 megabytes of xip qspi flash, and 4 gigabytes of flash on sd card, is not challenging, though it might be useful i routinely ran cp/m in 48k or 64k on a 4-megahertz z80, which is about 0.6 dhrystone mips, on one or two 100-kilobyte floppies, with transfer rates of about 1 kilobyte per second. i think cp/m will run in 16k this arm is as fast as 250 of those machines, it has as much ram as four of them, as much offboard xip program memory as 64-256 of them, and 40000 sssd floppy disks worth of storage, which furthermore can probably transfer data at 10 megabytes per second, ten thousand times as fast as the floppy also the z80 didn't have a pio programmable i/o pin driver, it had to handle i/o interactions itself instead of outsourcing them to a channel program in pioasm, which slowed down computation yes, if you're emulating a z80 on arm that entails some emulation slowdown, but it certainly isn't 250x. it might be 8x. if you do dynamic machine code translation it might be less than 2x instead of calling it 'tinycpm' he should call it 'giantcpm'

ideonode 4y ago
Why not just celebrate it for being neat, rather than berate it for not living up to your expectations? I thougbt it was interesting, and it has inspired me to try something similar on my pico. Especially after rereading Code recently. Oh, and I assume that tinycpm is because it runs on a Tiny 2040.
- kragen 4y ago
  it looks like a fine project; it just has a deceptive name