Ok, we're going to go a bit deep on technical data for a consumer part, but I feel it's necessary to convey just what makes Optane Memory able to be such a damn good storage accelerator, and to do so, I'm going to compare it against the fastest enterprise SSDs we've ever tested – Intel's own P4800X. I'm also going to compare against the fastest NVMe NAMD SSD we've ever tested, the Micron 9100 MAX. Consumer SSDs don't hold a candle to these enterprise parts, but how does a tiny Optane Memory 32GB module with only two dies stack up against these monsters?
Above is a Quality of Service chart. We are looking at Queue Depths 1-4 (where Queue Depth is how many simultaneous requests the SSD is working on). The P4800X was amazing at these lower QD's, nearly making fun of the next fastest (NAND) SSD, but look at that Optane Memory module! It's keeping right up with it, and it is doing so for a very high percentage of the time (noted as the plot lines climb vertically).
This is another technical chart typically reserved for enterprise reviews. The extremely stringent specifications for the P4800X have been left marked by the appropriate X's, and here we see that little 32GB module nearly meeting the 99.999% QD=1 consistency specification of a $1,520 enterprise SSD! Note the 50% (average) point on the QD=1 plot. That's 9.4 microseconds – actually faster than the P4800X. Here's some better detail:
Note that the first 90% of the requests are serviced *faster* by the 32GB Optane Memory module. The NAND SSD (gold line) doesn't even register when zoomed in this far!
Taking a look at writes, we see that the Optane Memory module isn't as fast here, but that appears to be in the interest of keeping power consumption lower for this particular part. Also, write latencies are not as important for this caching SSD to do its job in a typical manner.
To wrap this up, here is where Optane Memory parts sit in relation to other storage methods:
I've updated the 'Bridging the Gap' chart to reflect the even lower latency seen with the 32GB Optane Memory module, which is why the IOPS are slightly higher and the latency is slightly lower (now noticeably below 10 us).
Before wrapping the in-depth stuff up, here are the relevant performance profiles of the Optane Memory 32GB part. Remember, we are treating it as an enterprise drive for this profiling. These are not your typical client tests, though you can get client figures from them:
Random performance saturates at QD=8 and above, so we stopped at QD=32.
Sequential performance nearly saturates at QD=1 here. Now back to a couple of rounds of comparisons based on this saturated performance data: