Closer Look at Intel X25-M

Intel’s SSD line-up includes three series right now. The X25-M and X18-M series are based on MLC memory, have a capacity of 80GB and differ in size: they are designed in 2.5-inch and 1.8-inch form-factors, respectively. The X25-E series is based on SLC flash memory and has a capacity of 32 gigabytes. Intel is going to extend all the three series at the end of the year by adding models with twice higher capacities, i.e. 160GB and 64GB for the mass and extreme series, respectively.

The X25-M looks like any other SSD. It has the same rectangular case of 2.5” form-factor, the same size and fastening, and the same position of the connectors. The case contains twenty 4GB chips manufactured on 50nm tech process and marked as 29F32G08 (we couldn’t find information about such chips but there are Micron chips that have a similar marking and specs; as we have noted in the Introduction, Micron and Intel have a joint manufacture). Besides the flash memory, there is a PC29AS21AA controller and a 16MB module of 166MHz SDRAM (Samsung K4S281632I). So, we didn’t have to wait long for SSDs to acquire buffer memory. We don’t know if it is used for the drive’s internal purposes (we’ll explain below what purposes these could be) or as a buffer for organizing a queue. Anyway, it is good that there is buffer memory here.

The SSD owes its high speed to the ten-channel controller (one channel for each two memory chips). Thus, we have a kind of RAID0 built out of ten flash memory chips, the controller distributing the load uniformly among all the chips. The resulting speed of the whole SSD is roughly equal to the speed of one chip multiplied by the number of channels. This is the explanation of how Intel managed to achieve such a high speed of linear operations out of rather slow MLC chips. That’s a logical solution. If simple USB flash drives come with dual-channel controllers now, an SSD just demands a multi-channel controller. Coupled with buffer memory, this should solve the problem of low sequential speeds of flash memory that was the main drawback of SSDs which we noted in our previous review. You will see in this review what speeds this new solution can offer.

The declared specifications promise superb speeds of writing and reading. The write speed is lower than that of last-generation 3.5” hard disk drives that have become faster than 100MBps, yet it is very good for flash memory, especially for multi-level cells. The read speed of 250MBps is impressively high. The operating temperature range is good, too. Flash memory used to work at temperatures not higher than 55°C but the X25-M is rated for temperatures up to 70°C. Such a high operating temperature used to be declared for industrial flash memory that differed considerably from the ordinary type in terms of pricing.

The service life parameter is most important, too. It is usually declared for flash memory as the number of rewrite cycles. For modern MLC chips it is declared to be 10,000 cycles (this parameter equals 100,000 cycles for SLC chips, which is the second most important difference between these two flash memory types, besides write speed). This is somewhat inconvenient because you can’t compare the number with the service life of hard disk drives. And it is overall unclear for how long your SSD is going to work. Of course, you can multiply the number of cells (or, to be exact, the number of pages because the minimum changeable unit of information with flash memory is a page; the X25-M has a page size of 4 kilobytes) by the capacity of each and then by the number of rewrite cycles and get the maximum amount of information you can write to the drive. This value is 800 terabytes for the X25-M. It seems to be a big number, but there are two problems about it.

First, this number includes all possible write operations including auxiliary operations. PC enthusiasts have already learned that the share of auxiliary operations is high when the OS is installed on a flash drive, and the service life of this drive proves to be surprisingly short. Recording the date of accessing a file is a write operation, for example, and affects the drive’s service life. There are quite a lot of such operations that may not be obvious for an inexperienced user because modern OSes and programs keep various log files, work with the file allocation table (located on the drive) and change their internal files (the page-file alone is being changed constantly). All of this involves write operations. As a result, 800 terabytes is not the total of files the user can write to the drive during the latter’s lifetime, removing old and writing new ones. The real number is far lower even if you try to reduce the amount of such unobvious writes by turning off the page-file and setting your OS up accordingly. We wonder if notebooks with SSDs which are becoming popular now undergo such optimization.

The second problem roots even deeper, in the very algorithm of operation of flash drives and their interaction with OSes. A page is the minimum size of a data chunk at writing whereas the smallest data chunk for erasing is a block consisting of several pages. This is determined by the mechanism of erasing data in flash memory. For the X25-M this block is as large as 128 pages or 512 kilobytes or half a megabyte. As a result, if there is a request to erase (or rewrite) one page, the drive has to erase 128 pages. If there is useful information in these pages (which is a likely situation), this information has to be first copied into the cache and then written back. As a result, the number of write operations is increased manifold, with a proportional reduction of the drive’s service life.

Another aspect of the problem is that the load is distributed unevenly among the flash memory cells. Some cells store the same information for very long without any changes because some files are only read but not modified. Other cells, on the contrary, are often rewritten. Thus, the drive degenerates unevenly. Some of its cells may still undergo a number of rewrite cycles when other cells have already reached the limit.

And the third aspect of the problem is the interaction of SSDs with the file system. To remind you, the OS uses the file system to access the drive by means of the LBA addresses of data blocks it wants to request. LBA addressing is natural for HDDs but flash memory doesn’t have rotating platters divided into tracks and sectors. So, flash memory is equipped with an address translator that knows which physical cell corresponds to which LBA address. What may be a problem here? The problem is that the modern protocols do not erase data from all the LBA addresses when removing a file. They just change the part of the allocation table that refers to the removed file. Coupled with the address translator, it means that data written into a memory cell can be considered present even when from the file system’s point of view the file containing this data is removed. And recalling the above-described peculiarities of erasing data on the physical level, it means that nearly every request to delete data necessitates the writing back of even those data that belong to the already removed file.

This may be somewhat hard to grasp, so let’s put it in a different way: the LBA address translator that is responsible for removing data from cells corresponding to specific LBA addresses does not know anything about information stored in the file system. When deleting a file, the file system does not send a request to remove data from all the LBA addresses corresponding to the file. As a result, information in memory cells remains effective for the SSD (from the SSD’s point of view, it must be stored still) even when those cells belong to files that are already removed from the OS’s point of view.

The job of reducing the overhead (i.e. the amount of rewrites for each cell) associated with the described problem must be done by the address translator. In fact, the table of correspondence between LBA addresses and physical cells is not fixed and constant. It is dynamic. At different moments different cells may correspond to the same LBA address. So, the translator’s algorithms try to change the table in order to distribute the load uniformly among all cells, to achieve the same level of wear for all of them. This is where the SDRAM module installed on the SSD’s card comes in handy.

The translator is also responsible for reducing the wear from the erasure of cells in blocks. As you may guess, the efficiency of these algorithms affects the SSD’s performance as well as service life. Particularly, the drive takes some time to adapt to a change of load (for example when switching from sequential to random-address reading).

But let’s emerge from the technical depths onto the end-user’s level. What does the service life of the drive mean to us? Intel has suggested an elegant solution. The service life of an SSD can be determined basing on the average amount of written data while the above-described technical problems that increase the effective amount of written data is accounted for with two weights. The write amplification weight reflects the above-described problem with the increased amount of rewritten data. The wear leveling weight makes up for the reduction of the service life due to the fact that some cells wear off faster than others.

Intel evaluates the efficiency of its drive’s algorithms quite high. The company uses a write amplification weight of 20 and a wear leveling weight of 3 for the competitors’ MLC-based products. For its own products the company sets both weights at 1.1. Of course, we can’t check this out, so we have to trust Intel. Intel guarantees that if you write 100 gigabytes of data onto its SSD, the latter will serve you for 5 years (but the company provides a 3-year warranty for this product). It is, however, unclear how an ordinary user can measure the amounts of writing happening on his computer. It seems like we have to rely on statistical data and on the manufacturers’ claims.

But what does Intel mean by the MTBF of 1.2 million hours (as long as 137 years)? What does this impressive number mean and how was it arrived at? Obviously, it is calculated basing on statistical data and means the probability of the device’s failure after a certain period of time. But we don’t know what loads this MTBF is calculated for. Perhaps we’ll find this out in the future.

What we can check out right now are the speed characteristics of the X25-M. We took the drives discussed in our previous SSD-related review as its opponents. These HDDs still remain among the best in their market sectors. Here they are:

• Samsung SSD, 64GB, 2.5-inch form-factor
• Samsung SpinPoint F1, 3.5-inch form-factor, 7200rpm spindle rotation speed, 32MB of buffer memory, SATA interface, 1000GB capacity
• Fujitsu MBA3300RC, 3.5-inch form-factor, 15,000rpm spindle rotation speed, 16MB of buffer memory, Serial Attached SCSI interface, 300GB capacity

We did not include 2.5-inch hard disk drives into this review because they would just clutter the tables and diagrams. Such HDDs wouldn’t have had a chance among such opponents.