Update – Temperature Effects

After the initial posting of this article, further details have come to light. BrainSplatter got back to me with some data on temperature effecting the slowing effect on old file read speeds. The same test was run first at 32C (90f) and then at 52C (126F), here were his results:

Newest: 510 -> 507 MB/s = – 0.8%
2 weeks old: 477 -> 427 MB/s = – 10.5%
3 weeks old: 278 -> 211 MB/s = – 24%
4 weeks old: 289 -> 177 MB/s = – 39%

The temperature difference he used was a bit on the extreme side, but it did certainly show a speed delta that amplified with file age. This supports the theory that the EVO's controller is busy crunching ECC data, and as a result is running up against its Dynamic Thermal Guard Protection, which in this case is proportionally clocking the MCX controller / CPU slower in hotter environments.

We replicated this experiment with our 500GB EVO, but with a twist. I started out the first half of an HDTach read pass with the EVO sitting in close proximity to the chipset heatsink of our open testbed. This brought the shell temperature to ~105F (reasonable ambient temperature of low airflow portions of a system under load). At the 50% progress mark, I moved the EVO directly under an open-air fan, bringing the temp down to ~85F. Given that we had a 10 month old >400GB file spanning across the 50% mark of this drive, we should clearly see a speed change:

…and we do. Given this knowledge and looking at the start of this result, it also seems that the controller *could* be reading and error correcting old data at 150-200MB/sec, but it quickly throttles itself back due to spiking temperature under load.

We can draw a few conclusions here:

  • The flash itself is not degrading under higher temperatures, as the speed bounces back once the drive is cool.
  • The controller was likely not meant to run at this sort of continuous ECC load, which further supports that a firmware fix might correct the issue (programmatically compensating for cell drift over time can mitigate the need for heavy ECC calculations and would theoretically restore speeds to near full for old files).

More to follow as we continue tracking this issue. Once Samsung issues a firmware update, we will re-test on our stale data samples and report back ASAP.

PCPer forum users – there's a thread here where folks are reporting their experiences. Feel free to chime in!

« PreviousNext »