Milestones achieved and future works

[jj1bdx]

TODO

• Operational test with the hardware assembled in a chassis
• Connecting to a 3.3V-driven digital interface

  Milestones achieved

  Raw output test of v2rev1 shield 26-OCT-2015:

• Speed: ~9400bytes/sec of SHA512-filtered output (76kbytes/sec of raw output)
• No filterting on AVR
• Achieved nearly the same quality of being filtered on AVR (FIPS 140-2 test, TestU01)

  Completed as v2rev1 shield 25-SEP-2015:

• Speed is now ~10800bytes/sec
• A production-level RNG shield design: CMOS-level output voltage from each noise generator unit
• Higher output voltage from the noise diode circuits by using 2N3904
• Higher gain of output amplifiers (74HCU04 linear amplifiers biased to Vdd/2)

  Completed as v1amp shield 22-SEP-2015:

• Completed as v1amp shield 22-SEP-2015: making a new RNG shield, by adding an amplifier transistor for each noise generator output: see https://code.google.com/p/avr-hardware-random-number-generation/downloads/list

  Completed as v1rev2 in 2009:

• 1000bytes/sec output with 250kHz/4us timer sampling rate with 16MHz system clock

• Output smoothing: von-Neumann 2-bit smoothing, and 2-byte xor to 1-byte looks sufficient for maintaining the output quality.
• MD5 of 32 bytes (+ a previous block output of 16 bytes) into 16 bytes output is overkill (too small gain for a too complex software).

  NO-TODO

• Trying to obtain entropy from ATmega168/328P ADC outputs is impractical. ADC clock is slow (50kHz ~ 200kHz) comparing to STM32F103's (~12MHz), and the resolution is coarse (10bit). Only the LSB (1 bit) change was observed when measuring Vref (1.1V) under the full-scale (0 ~ 5V) setting.

[jj1bdx]

Some more future ideas:

• Implementing on ATtiny2313
• Implementing with surface-mounted parts

[jj1bdx]

v1amp shield built on 22-SEP-2015 and under testing. Thoughts:

• Amplification of the noise diode output is essential for easy comparing; achieved 3500bytes/sec.
• Calibrating two fixed bias amplifiers within the same DC output voltage is hard, and not for the mass production, since tuning the values of the resistors attached to collectors of output transistors is required.

[jj1bdx]

v2rev1 shield design fixed and prototype build on 24-SEP-2015. Thoughts:

• Using 74HCU04 consumes considerable power (15mA per each inverter when being used in the linear amplifying region at 5.0V), but it is worth investing for because no need to tune the bias points any more.
• Sampling should be performed on an equal interval. Free-running sampling degrades the statistical quality of the generated random numbers.
• ATmega168 is not that fast to be capable to handle everything with interrupts when the event rate exceeds 50kHz or so. Removing the interrupt and introducing a spin-lock-like synchronization made the code run much faster.

[jj1bdx]

v2rev1 2nd hardware built and under testing since 2-OCT-2015. No significant difference observed between the two v2rev1 hardware units.

[jj1bdx]

avrhwrng-rawtest software has been under testing since 26-OCT-2015. Features: 1Mbps raw output of concatenating PD7/PD6 input, repeated at the rate of 400kHz. No filtering, so SHA512 whitening must be required to obtain statistically high quality random numbers. Actual measured speed: 76kbytes/sec.

Fundamental question: _If no filtering is needed on the RNG hardware, does the hardware have to be a microcontroller? Isn't an FTDI FT232RL/FT232H/FT245RL parallel I/O enough?_

[jj1bdx]

v2rev1 3rd hardware built and tested on 25-OCT-2015. No significant difference observed between the three v2rev1 hardware units.

[jj1bdx]

Tag stable-rawtest-20151111: avrhwrng-rawtest.hex has been uploaded. No further advanced development for the current design; further development activities will be performed on a new hardware other than 8bit AVR Arduinos. This issue is closed.

[jj1bdx]

IPSJ IOTS2015 27-NOV-2015

• PDF Paper in Japanese published
• Slides in English published
• Junction temperature (of CPU) could be used as a source of randomness

FSIJ monthly meeting

• NeuG binary file can be freely released abiding the all license terms (e.g., GPLv3 for chopstx, etc.)