HDDs and NAND Flash will be Around for Some Time
by Hu Yoshida on Feb 19, 2013
I have written several times about the declining rate of areal bit densities in hard disk drives (HDDs). Below is a chart that was published in IEEE Transactions on Magnetics, which shows that the decline has been going on since 2010. This means that the price erosion of hard disk capacity, which has been about 30% for the past 50 years will slow down to less than 20% if we depend on bit densities. This reduction in price erosion will have a dramatic impact on the cost of storage unless we find other ways to reduce storage costs.
This chart shows that NAND flash bit densities are following the same technology curve as HDDs. However, the growing acceptance of flash in the consumer and server space at the expense of HDDs will drive up the volume so that the price of NAND will be lower than HDDs by 2019 as shown in the chart below.
The durability of flash technology is nowhere near the durability of HDDs, so there will still be a continuing need for HDDs for long-term random access retention. But, HDD bit densities are not increasing at the rate they have in the past, so we need to look at other ways to increase densities.
As I noted in my previous blogs, HDDs have struggled to increase performance while increasing capacities, and users have tried to compensate for the lack of performance through wide striping and short stroking HDDs. With the introduction of flash technology and dynamic page-level tiering, we have the ability to create a dynamic pool of flash modules and HDDs where pages are written first to the high-performance flash layer and remain in that layer while they are active. This frees the disk from having to worry about performance and we can explore other ways to increase the capacity of HDDs that will augment the bit densities, and provide the type of cost erosion that we have enjoyed in the past.
Below are a few suggestions on how to increase densities if performance is not a factor:
- Spin the disks slower and put more bits per track
- Use more of the recording surface
- Add more disks to the disk enclosure (Western Digital)
- Seal the enclosure in inert gases to add even more disks (Western Digital)
- Shingled recording where tracks are overlapped to increase track densities
- Combinations of bit pattern media and heat-assisted magnetic recording
These approaches may extend the life of HDDs, but sooner or later we will need to replace HDDs with solid state technologies, which are more durable. There are several technologies that could replace NAND flash with higher performance and greater endurance.
Phase change RAM, which uses heat to change crystalline and amorphous states, has access times in nanoseconds and endurance cycles of 108 compared to 103 for MLC flash. Another solid state technology is Spin Torque Transfer – Magneto RAM, which uses spin polarized current to transfer angular momentum to magnetic material and has access times in nanoseconds and endurance cycles of 1015, which is approaching the endurance cycle of HDDs. In order for these technologies to be viable they must be able to capture the volumes that are needed to drive costs down. The consumer market currently drives volumes, and the consumer will not pay extra for nanosecond access times or 108 endurance cycles. NAND flash is good enough for the consumer.
Unless there are some major breakthroughs in technology or user demand, we will continue using a combined pool of NAND flash for performance and HDDs for capacity in enterprise storage for the next 5 to 10 years.
Comments (2 )
Hello, Hu. It is very interesting and I fully agree with your ideas. some additions:
The Nearline disk will continue to develop by adding platters and using sealed, gas filled HDD. For example Helium filled , 3.5” 5.6 Tbyte end this year (HSTG WD).
The NAND flash technology is reaching its limits and can grow using 3 dimensions or 20nm. Tt is very challenging to maintain same performance, write endurance, and retention specs beyond 20 nm, therefore most likely the 3D cell NAND technology may win. For example Hybrid Memory Cube (Micron & Intel)
I predict that in the mid-decade other semiconductor technologies will emerged – so called Storage Class Memory (SCM)*. The SCMs will deliver better performance than the Flash NAND and much better endurance at comparable price. The SCM design targets are better performance than the NAND based memories but slightly slower (x3) than the biPolar DRAMs.
There are two types of SCMs on the drawing boards Memory type and Storage type:
1. Memory Type – <200 nsec Read/Write/Erase time, 10**12 write/erase cycles
2. Storage Type – <2 µsec Read/Write/Erase time, 10**8 write/erase cycles
Both types will require 1/10 power of enterprise type HDD . The storage type planned price is below than 4-6x the cost of enterprise HDD.
The first pilots of new technologies will be delivered in 2013 mainly for hand held devices. industry's first Spin-Torque Magnetoresistive RAM (ST-MRAM) chip targeted for smart phones is 50 times more expensive than NAND flash. Another technology shipped today for mobile devices is phase-change memory (PCM).
*Some examples of future enterprise SCMs: Phase change SSD memories (Micronic,IBM GA 2016), Memristor (hp, Hynix E2014?), Racetrack memory (IBM GA 2018).
Josh, thank you for the update on solid state technologies. There will be many interesting possibilities in the future. Please keep us informed on the progress of these technologies. I believe the challenge will be to capture market share which will be driven by volumes and price.