5 ms·
High-Throughput Random Number Generation Using Quad Row Activation in DRAM Chips [pdf]
- lisper 5y agoThe utility of this turns on a question whose answer is not readily apparent from skimming the paper: is this something that can be implemented purely in software? Or does it require a hardware mod? Because if it's the former then it could be a game-changer. If the latter, not so much.
- readams 5y agoThe paper uses a custom memory controller based on SoftMC implemented on an FPGA. The DRAM modules themselves are off-the-shelf. It appears you need precise control over DDR4 command timing that would not be possible with a normal memory controller. So it's hardware-only, though you could imagine adding this feature to memory controllers in the future.
- lisper 5y agoSure. but there's not much difference between that and just adding a regular old TRNG. A hardware design change is a hardware design change :-(
- failwhaleshark 5y agoAs an alternative for high-throughput TRNG entropy sourcing: Why not just sample thermal noise like the Simtec Electronics Entropy Key# did? Doesn't it just need a fast-enough ADC? (How expensive is a ~ 10 Gb/s ADC?) # I'm kinda annoyed they didn't open-source the design before shutting down. :(
- mastax 5y ago> How expensive is a ~ 10 Gb/s ADC? $1000. Much cheaper to use many slower ADCs in parallel. Or put your entropy directly into a Schmidt trigger, etc. since you don't care about any of the specs that make ADCs so complicated.
- cycomanic 5y agoWhere do you get that price from? We have been trying to source high speed ADCs for some experiments and it's virtually impossible to get anything above 5 GS/s and prices are more than an order of magnitude higher than what you state.
- tails4e 5y agoRFSoC is an fpga with high speed ADCs and DACs integrated, 8 5Gsps ADC, and 16 10 GSPS DACs in a single device, and less than 1000 bucks IIRC, though not sure on volume (i.e. If you can buy single parts)
- mastax 5y agohttps://www.analog.com/en/parametricsearch/10826#/sort=1746,desc https://www.analog.com/en/parametricsearch/10826#/sort=1746,... That is the price for just the IC, at 1000 quantity. Depending on the experiment that might not be a relevant number.
- nine_k 5y agoYou don't need an ADC, you need a fast threshold circuit with regular sampling that will turn the noise into a bitstream, and a shift + latch register to accumulate large enough words from these bits. The biggest problem of such a design might be correlation with the variations of the power supply voltage.
- failwhaleshark 5y agoAh-ha, hack enginerding FTW. Overenginerding costs $$$. I wonder about metastable sequential logic (D flip-flop sampling 3 NOT gates in a loop) XORed thermal noise sampling from a resistor XORed with the D flip-flop's clock signal. Violating setup-and-hold times of digital logic would also probably help generate some fun randomness.