5

Comparing prices at a local store, I calculated prices 0.36 USD/GB for SSD NAND flash memory and 5.41 USD/GB for DRAM memory. The difference is 15 times. Why so big difference?

Both are semiconductor devices. Memory cells of both types occupy approximately the same die area: 4 F^2 for NAND flash memory, 6 F^2 for DRAM. (I can't give a good source; this information is scattered over the internet.) MLC of flash memory gives 2 times improvement. (BTW, is the same technology possible for DRAM?) I can't blame market forces because both markets seem similar to me in terms of competition, maturity, and size. For example, DRAM revenue 35.74e9 USD, SSD revenue 10.9e9 USD in 2013.

beroal
  • 211
  • 3
  • 9
  • 1
    For DRAM,they run at higher clock speed like 666MHz and more.But a flash memory doesn't have that much speed.And moreover,the data transfer speed from a DDR3 ram from and to the processor is about 3500 Mbit/sec.This much higher speed than the USB 3.0. – Aadarsh Jan 10 '16 at 19:45
  • 1
    Basically boils down to DRAM read/write speeds, and DRAM rewrite levels. NAND memory has much more limited rewrite, and needs wear leveling, while DRAM does not suffer from that. That's what I chock it up to. – Passerby Jan 10 '16 at 20:41
  • @Aadarsh: So you are saying that faster transistors cost more and DRAM contains faster transistors? – beroal Jan 12 '16 at 09:34
  • @Aadarsh: “And moreover,the data transfer speed from a DDR3 ram from and to the processor is about 3500 Mbit/sec.This much higher speed than the USB 3.0.” I admit that SSDs are way slow than DRAM. SSD speed increased dramatically though: the sequential read speed is 2 GB/s and the price is 0.95 USD/GB for Samsung SM951 M.2 PCIe AHCI SSD (256GB) http://www.pcworld.com/article/2977024/storage/m2-ssd-roundup-tiny-drives-deliver-huge-performance.html . – beroal Jan 12 '16 at 09:54
  • You told about sequential read speed is about 2Gb/s.But think whether you always read in sequential mode,when you use in a PC?.Defragmentation will help to make it somehow sequential,but most of the time,it wont be 100% sequential.And also,the memory bandwidth is about 550Mb/s (SATA3).For a graphics card or RAM the memory bandwidth is about 113Gb/s. or more. – Aadarsh Jan 12 '16 at 10:23
  • You should not use the SSD for the substitute for RAM.Because,it'll damage the SSD very soon,if you use it at it's maximum speed always.Likewise,Flash memory as substitute for RAM will damage the memory due to injection of electron to floating gate. – Aadarsh Jan 12 '16 at 10:28
  • @Aadarsh: “And also,the memory bandwidth is about 550Mb/s (SATA3).” The SSD I brought to your attention uses PCIe. – beroal Jan 15 '16 at 03:57
  • @Aadarsh: “You should not use the SSD for the substitute for RAM.Because,it'll damage the SSD very soon,if you use it at it's maximum speed always.” This was true long time ago. Now it's a superstition. – beroal Jan 15 '16 at 04:00
  • I think this link will give you valuable information. http://superuser.com/questions/617864/why-not-use-ssd-space-as-ram – Aadarsh Jan 15 '16 at 05:56

4 Answers4

4

The key element in the difference is redundancy. A DRAM has to be perfect (that is, it has no redundancy). This is because, in order to get the speeds required of DRAM, there is no buffer level between the address inputs and the memory cells except for address decoding.

Flash, on the other hand, is not expected to run nearly as fast as DRAM, and this allows the insertion of various logic into the address logic. Specifically, data is organized into blocks (equivalent to pages), and there are a fair number of spare blocks included on the chip. When an access is attempted the address essentially goes into a lookup table which redirects access from a known bad block to a good (spare) one.

The result is that manufacturing can be much cheaper for flash than DRAM, since it is much harder to produce a bad (unusable) chip for flash.

WhatRoughBeast
  • 59,978
  • 2
  • 37
  • 97
  • 3
    Although I cannot find a source (after less than a minute of searching), I *believe* that DRAM chips do use redundancy to improve yield, probably at the mat level. (They do not use ECC by default to allow smaller cell sizes or lower refresh rates.) Even on-chip processor caches (with higher speed requirements, though also faster logic gates) use redundancy to improve yield. –  Jan 11 '16 at 13:47
2
  1. Supply and demand. There is much bigger supply of NAND chips out there. Also, there are many manufacturers across many different process nodes. There are only four (as of 2016) makers of DRAM on the latest process node.
  2. 5$/GB is a per GB DRAM on a module. 8 to 36 chips on a module depends on the configuration. Cheap SSDs can have as little as two chips + passives. (M.2 PCIe controller and NAND flash). From the above you can see that on average, DRAM has a larger failure potential than an SSD, therefore, requires higher price.
Lior Bilia
  • 7,282
  • 1
  • 20
  • 30
  • 1
    I don't think that (1), i. e. market reasons, are valid, as I already wrote. Both markets seem mature. – beroal Jan 12 '16 at 09:25
  • Could you, please, clarify (2)? I suppose it is about the probability of defects during manufacturing. Could you provide formulas involving probabilities? – beroal Jan 12 '16 at 09:27
2

From the IC designer perspective, FLASH is much more expensive to make than DRAMs because I use more masks, and they take more area. I don't know where you got your area numbers, but I can believe that if you had a FLASH only process, I could tighten it up a bit. Of course, that would be expensive. You need to have a comparison that is uniform. I get a single bit of FLASH in the space that I get 8-bits of DRAM trench caps on a modern FinFET process. I've attached a picture of the DRAM layout for a 8-bit DRAM cell.

FLASH needs to not leak, and DRAM can be leaky so I don't need to worry about my oxide thickness or quality so much. Also, every time you program FLASH, you use hot electron injection that can cause oxide damage statically. The greater the VDS, the faster you program because of more hot-electrons, but the more risky you get a chance of having a hot-hole make an intermediate state, so you want thicker oxide.

Growing ticker oxide costs time for deposition, and therefore money, but just on your question: If you have a comparable process, I would expect the cost between FLASH and DRAM to be 8x for bit equivalence just on area. I expect that the discrepancy that you see between DRAM and FLASH is just process cost.

Trench Caps

b degnan
  • 3,167
  • 2
  • 17
  • 36
  • 3
    You seem to be telling us why flash should cost more, when the situation is the reverse. – William Brodie-Tyrrell Jan 11 '16 at 02:36
  • 1
    I have never used a FLASH-specific process; however, FLASH does cost more as far as Si wafer space/mask on the processes that I have used. Consumer pricing is a magical thing, and that's why I left it open ended. If you look at die costs at 0.05USD per 1mm^2, that's about the actual cost in silicon after mask setup. – b degnan Jan 11 '16 at 14:39
2

I think the answer is that the different technologies have different yields.

Flash can work and be sold with bad bits scattered all over the chip and the controller will just hide the faults.

DRAM must operate at much higher frequencies, and it must be perfect. Any single bit error in a DRAM die could cause that die to be discarded (unless there is redundancy in its design), and the dies are not small.

See this previous question: Yields in DRAM and other Massively Redundant Processes

Community
  • 1
William Brodie-Tyrrell
  • 2,262
  • 10
  • 14