There’s something really odd about Nimbus Data’s colossal 100 terabyte ExaDrive DC SSDs, and it’s not their sheer capacity (although that’s pretty remarkable by itself!) The strange thing is that they can’t be worn out. It’s physically impossible.
At first glance that may seem wrong-headed. NAND flash wears out, and that was the cause of a lot of concern early in the SSD market. I remember battles in trade shows where an SSD maker would harp about HDDs saying that they were subject to mechanical damage, and overstating the facts to make it sound as if HDDs had to be handled with extreme delicacy or they would suffer catastrophic failure.
Naturally, HDD makers fought back with a very truthful retort: That the NAND flash in an SSD would wear out simply by writing to it too much. They would point to those infrequent times that this occurred and make them out to be just as intolerable of a situation, if not more so.
That was a long time ago, though, and since then SSDs’ internal wear management has become significantly more sophisticated, and host software has removed a lot of unnecessary write activity that was damaging to the SSD. Meanwhile, SSD users, particularly in the data center, have learned how their workloads behave, and how to segregate the high-write traffic from the low-write traffic so that they can now manage SSD wear to accommodate their needs.
In the meantime SSDs have incorporated internal wear measurement and reporting, in the form of SMART attributes, to indicate exactly when they are likely to fail. They also have added a Trim command to help guide the SSD’s internal controller to erase unused blocks of data in a way that improves the SSD’s internal wear management.
So SSDs are less of a worry than they were a decade ago, but they still can wear out — except for the Nimbus 100 terabyte ExaDrive DC SSD. Why is that?
It’s largely due to the SSD’s huge capacity. This SSD is sold in capacities of 16, 32, 50, & 100TB. That largest one is about five times the capacity of today’s highest-capacity HDDs! When such a large capacity is accessed through an I/O channel with a more modest bandwidth, it’s impossible to overwrite the flash enough times to wear it out.
A little math will show you how that works. The maximum write bandwidth of these SSDs through their SAS and SATA ports is 460MB/s. There are 86,400 seconds in a day, so the maximum number of megabytes written in a day would be the product of these two numbers. Divide that by 2^20 and you get terabytes per day, which is 37.9. If you divide this number into the SSD’s terabyte capacity it gives you drive writes per day (DWPD). For the four capacities offered these will be:
Capacity | DWPD |
16TB | 2.37 |
32TB | 1.18 |
50TB | 0.76 |
100TB | 0.38 |
All of these are modest numbers.
When you take into account the fact that these SSDs are guaranteed for a five-year life, then the maximum number of writes to any memory location, without overprovisioning, would be:
Capacity | Writes |
16TB | 4,326 |
32TB | 2,163 |
50TB | 1,384 |
100TB | 692 |
While these numbers might rule out the use of TLC, which generally has endurances in the hundreds, it fits well with MLC flash, and it’s even easier to handle with SSDs that use the technique of first storing writes in bits that are used as SLC, only to later be moved to MLC bits as the data becomes cooler.
Is it difficult to get that much flash into an SSD form factor? Not really. The reasons it’s not done more often are that it’s often counterproductive to put your fastest storage behind a slow interface, and few people are so worried about space that they will pay extra to get the most flash in the smallest form factor.
Back in 2014 The SSD Guy published a post called: “How Big Can an SSD Get?” This was back when everyone wanted to produce a “Hero” SSD. My analysis found that 30TB was then possible with the 128Gb (16GB) chips available at that time. Today’s 176-layer TLC NAND chips have 1Tb (128GB) capacities, or an order of magnitude more, so it would seem that a 300TB 2.5″ SSD could be built today. Since it’s likely that Nimbus uses MLC chips rather than TLC, then the maximum 2.5″ capacity would be half as much, or 150TB. Add to this the fact that the Nimbus SSDs use a 3.5″ form factor and it would be even simpler and cheaper to put 100 TB of capacity into an SSD.
This distinction can be achieved by any SSD with a high enough capacity and a slow enough interface, but it’s intriguing to consider that there are indeed SSDs that cannot be worn out.
This would be a perfect match for the “flash on a reel” form factor, where the effective transfer rate is close to zero!
Steve, You have a wry sense of humor!
Jim
Only a select few can appreciate that. Thanks!
Seriously though, a relatively slow interface with a 50 microsecond latency (as their website claims) might be just right for a hash table that can retrieve records up to about 32k with just one random access.