![Rapid row activations (yellow rows) may change the values of bits stored in victim row (purple row).](https://hackaday.com/wp-content/uploads/2015/03/rowhammer.png?w=310)
As DDR SDRAM increases in density and speed, so too do new challenges and opportunities appear. In the recent DDR5 update by JEDEC – as reported by Anandtech – we see not only a big speed increase from the previous maximum of 6800 Mbps to 8800 Mbps, but also the deprecation of Partial Array Self Refresh (PASR) due to security concerns, and the introduction of Per-Row Activation Counting (PRAC), which should help with row hammer-related (security) implications.
Increasing transfer speeds is primarily a matter of timings within the limits set by the overall design of DDR5, while the changes to features like PASR and PRAC are more fundamental. PASR is mostly a power-saving feature, but can apparently be abused for nefarious means, which is why it’s now gone. As for PRAC, this directly addresses the issue of row hammer attacks. Back in the 2014-era of DDR3, row hammer was mostly regarded as a way to corrupt data in RAM, but later it was found to be also a way to compromise security and effect exploits like privilege escalation.
The way PRAC seeks to prevent this is by keeping track of how often a row is being accessed, with a certain limit after which neighboring memory cells get a chance to recover from the bleed-over that is at the core of row hammer attacks. All of which means that theoretically new DDR5 RAM and memory controllers should be even faster and more secure, which is good news all around.
The future of RAM is static.
And so is the past. It’s called cache. Every computer has it.
The reason computers have different types of memory is because of tradeoffs. SSDs are fast and cheap per GiB and also non-volatile, DRAM is faster (higher throughput and lower latency), but less cheap and volatile. Cache doesn’t need to be refreshed and has the lowest latency as it is inside the CPU, in the same package or on top of the die(3d cache), but it requires more transistors per bit and has restrictions on its maximum size. Cache cannot be upgraded unless you replace your CPU. Unless you can make SRAM as cheap as DRAM both are here to stay. We are always going to have different memory types working together.
Remember when some company tried to sell i-Ram using DDR RAM with battery pack to act like a SSD? https://en.wikipedia.org/wiki/I-RAM
SRAM is many more transistors per bit, and to make it worse it is no longer shrinking in absolute area per bit
Parity must be an essential requirement for all memories