The post on on notebook flash drives [see Notebook flash SSD market: fantasy or mirage?] generated many comments.
Part of what makes it hard to discuss flash is the dearth of information about how it works. My investigation of flash issues has been helped along by hints and tips from insiders and the occasional paper that sheds light on FTL design issues [see Flash chance, based on a paper from Microsoft Research].
Thus I was pleased to get a 2500 word email from a polite and knowledgeable SSD engineer cum marketing guy commenting at length. I asked him if I could publish his comments and he said yes – if I preserved his anonymity and removed the names of the companies he’s worked for.
Seemed reasonable. Since it’s long I’m breaking it up into 2 parts.
In the editing I’ve removed some info, abridged some comments, added the bold face headers and broken some long paragraphs into 2 or 3 shorter ones for online readability. At all times I’ve sought to preserve the author’s meaning.
Begin SSD guest comment
First up, great post. I agree with most of what you said. I haven’t used an SSD drive myself, but the reviews I’ve seen make me wonder if I ever will – way too expensive, for way too little benefit.
The lay of the land
Quick background comment on flash memory. There are two main kinds of flash memory: NOR & NAND. NOR is similar to SDRAM, NAND similar to HDD. NOR can be accessed randomly, is faster (at least for reads) than NAND, but the chips are smaller and cost a lot more per GB.
NAND can only be accessed in blocks like a HDD, the chips are larger, and the cost per GB is less than NOR. NOR is commonly used for firmware (e.g. the BIOS in your PC), NAND is commonly used for bulk storage. In the discussions about SSDs, we’re always talking about NAND, so I’m going to say “NAND” rather than “flash” in the rest of this email.
NAND flash has a ~10x worse $/GB than HDD, but it has about a ~10x better $/IOPS than HDD.
Your tour guide
I’ve been in the semiconductor business for ~20 years, first as an engineer, then gradually transitioning in the management & marketing. In my last job I developed relationships with all the NAND market players. When I first started looking at NAND chips, 4MB chips were still around, now we’re working with 4GB chips – wow!
I think that the SSD drive makers can do a MUCH better job than they’ve done so far, and that the raw technology is capable of doing much better. I think eventually the SSD products will get better, and we’ll see SSD drives (or their successors) used almost everywhere.
1st, the numbers
A state of the art MLC NAND chip today is 4GB, so a 64 SSD drive has at least 16 NAND die inside. The peak write speed should be ~5MB/sec/die, so the SSD should be capable of ~80MB/sec sequential write. Peak read speeds should be ~30MB/sec/die, so the SSD should be bottlenecked by the SATA interface.
These are MLC numbers. SLC performance will be even higher, about 8x better for write speeds for the datasheets I compared. True, these are best case raw performance numbers, and in the real world there are complications that will keep you well away from these numbers, but it should be possible to do waaayyyy better than we’re seeing now.
Responding to StorageMojo
[He goes on to quote and respond to some points from the StorageMojo post. I’ve put those in quotes.]
Flash has a place in one notebook niche: below the $40-$50 minimum cost of a disk. As we’re already seeing with the Asus Eee, replacing $50 of disk with $10 of flash makes a big price difference.
I agree 100% with this – if I can build a system using either $10 of NAND, or $50 of HDD, and the $10 of NAND is enough storage, then NAND wins. It doesn’t matter that the HDD has higher $/GB, or that it will have loads of spare GB – it costs $40 more, and it’s out.
$10 of NAND storage will buy a rapidly increasing amount of storage, so the cut-over point where NAND wins based on entry cost along is rising rapidly. I think that the $/GB number is halving every 12-18 months, so in 2-3 years we’ll get 4x more NAND for the same cost.
Given the multi-billion dollar cost of semiconductor fabs, getting the notebook SSD market wrong would make Toshiba’s $250 million HD-DVD loss look cheap.
Actually, while the size of the $$$ at stake are probably pretty large (inventories, controller chips design efforts, etc), they are not as large as a fab. A modern day, state of the art wafer fab costs several billion dollars, but that investment won’t be completely at the mercy of SSDs succeeding, for two main reasons.
One, these fabs are built to make both SDRAM & NAND. Both markets are very sensitive to the balance of supply/demand, and therefore both markets exhibit wild price swings. By building the fabs to support both types of (very high volume) products, they can switch from one to the other based on the supply/demand balance in both markets.
Two, there are other huge markets for NAND, primarily memory cards (SD, MMC, xD, memory stick, CompactFlash, and variants), & consumer electronics devices (phones, especially SmartPhones, GPSs). One of the biggest customers on the planet for NAND is Apple (iPods, iPhones).
It is true that Toshiba is playing a billion dollar poker game with (mainly) Samsung as to fab capacity (if there is overcapacity, both companies suffer, but if one under-invests and the other over-invests in capacity, then the over-investor wins), but SSDs succeeding or not will happen slowly enough that the capacity differences can be absorbed by speeding up or slowing down the bringing on of new fab capacity.
. . . today’s spot market MLC $2500/TB . . .
That spot market price is about right. This implies that the 64GB SSD in the Macbook Air should be about a $300 upgrade, not a $1,300 upgrade. True, you do get a slightly faster CPU in the deal, but I think that we’re looking at way high early adopter prices right now.
And if the market doesn’t appear, a billion dollar write off.
I’m guessing that they are betting $10M to $20M on a project to build a SSD controller design chip. They can’t afford not to have the controller, in case the SSD market results in a significant proportion of their volume, and they can’t assume that they will be able to buy the controller from an outside company (or even more risky, a competitor).
Power: no SSD notebook has gained more than 10 minutes battery life over disks. Since flash is already power-efficient that won’t change. Disks have multiple opportunities to improve power use – and with over a $1 billion a year in R&D behind them – they will.
The primary users of power in a note book are (in order)
- The display back light
- The CPU
- Everything else
The HDD is lumped in with everything else. Flash should have a significantly better power consumption than HDD, but since both are operating in the power shadow of the display & CPU, it doesn’t make a lot of difference.
Despite what a commenter said, spinning the HDD platter doesn’t take a lot of energy. Spinning them up to speed from idle does take a lot of energy, but only for a few seconds. Keeping them spinning once they are started only consumed enough energy to overcome the bearing friction, and that friction is pretty low. Most of the power spent in accessing a HDD is in moving the read/write heads, and in the read channel electronics.
One other think you didn’t mention is that after ~30 years of development, Windows (Linux, OS-X) is pretty well optimized to the characteristics of HDDs. Have you ever heard of the Windows XP Prefetcher? Wow!
Now, if we can do something about the power consumption of the display back light and CPU, then SSD vs. HDD might make a difference, but by then we’re talking about cell phone like battery life so it probably won’t matter.
End of part 1
Next up: flash financials; 4 level flash durability; data protection and more in the conclusion to NAND – an engineer’s perspective
Comments welcome, of course. Did you notice that he actually disagreed with much of what I said? But he was nice about it.