sf  on February 1st, 2010

Hi Kevin,

Congratulations to the launch of your new 25nm NAND process technology. Since SSDs are one of the target applications for these new class of NAND parts what endurance can we expect out of 25nm products? Will you be able to reach the same levels of endurance as your 34nm Enterprise NAND? What should we expect of non Enterprise parts that would go into Consumer SSD applications (e.g. RealSSD drives).

Best Regards,
Sf

Kevin Kilbuck  on February 2nd, 2010

Thanks for the post, SF.

The current endurance capability of our 25nm MLC products is sufficient for a wide variety of applications, including SD cards, USB thumb drives, and portable media players. As new NAND process technologies ramp into mass production and mature, endurance naturally increases over time. We expect 25nm MLC products to be used in client SSD’s as the production volume increases and the process matures.

In addition, we will introduce 25nm SLC and E-NAND products that have higher endurance levels than standard MLC. At this time, we don’t know where our endurance levels will be specified, but rest assured they will be significantly higher than standard MLC, enabling these products to be utilized in enterprise applications.

Also, keep in mind our 8GB product enables a doubling of capacity within a given form factor. By doubling capacity in the same application (same usage conditions), endurance requirements are cut in half.

If anyone wants to a little more detail on how reliability and performance increase as NAND matures, see my posts on Enterprise NAND and 3-bit-per-cell NAND

aidle  on February 3rd, 2010

Greetings,

Congratulations to the launch of your new 25nm NAND. With this technology, even the CD technology will be replace with SSD in near future. By the way how much power saving we will expect from 34nm to 25nm technology?

Best regards,
aidle

Kevin Kilbuck  on February 4th, 2010

Aidle,
One of the key benefits of moving to 25nm is the ability to reduce the number of devices used in the system by a factor of two. For example, to build a product with 256GB of storage would require 64 34nm devices (32Gb), but only 32 25nm devices (64Gb). This roughly cuts the NAND power consumption in half.

GullLars  on May 25th, 2010

Hi kevin,

I saw your presentation after you posted it in february, and have recently come back to some stuff regarding it. I hope i’m not too late here to get a response.

I have been looking for performance numbers for NAND used in SSDs, more specifficaly ONFI 2.x NAND. The only thing i could find with usefull information was this document from 2008-2009: http://onfi.org/wp-content/uploads/2009/02/onfi_2_breaks_io_bottleneck.pdf
I am currently discussing SSD performance with other power users on the xtremesystems forum, and i would love any hints to ballpark figures for performance of the 25nm flash if you are able to share them.
I would also like to know if they are ONFI 2.1 compliant.
One of the beefs i have with SSD manufacturers today is the low performance scaling with capacity, and i like how you made more performance avalible with 6Gbps interface, but if the numbers in the document i referenced above are anywhere near correct, there is a lot of unused (raw NAND) read performance in most SSDs out there. It seems to me most SSD manufacturers are hung up on sequential write performance, and won’t even consider an unbalanced sequential read:write ratio, even though there are some major usage patterns (like boot drive) that would benefit from it. The numbers in the document referenced above sugggests a raw read:write performance ratio of about 10:1 for MLC, is this anywhere near your upcomming 25nm products?

Kevin Kilbuck  on June 3rd, 2010

Hi GullLars,

You raise some good questions about flash performance and how it can be maximized in SSD applications.

Today, most client SSDs utilize the SATA 3Gbps interface, which NAND can outperform, even with the legacy asynchronous interface. Our RealSSD C300 is capable of achieving SATA 6Gbps, which makes the NAND interface the bottleneck, and Micron was able to establish performance leadership by using the ONFI-2 synchronous NAND interface, which is capable of achieving 200Mbps on a single (x8) NAND channel. The latest ONFI spec revision is 2.2, which Micron’s 25nm NAND products will adhere to.

You are correct about most computing applications being read intensive, with the exception being certain enterprise applications that have closer to a 1:1 read:write ratio, but this is the exception, not the rule. For writes, the NAND array is the bottleneck, however much of this latency can be hidden in multi-channel/multi-die applications, and the ONFI 2.X synchronous interface can provide a performance boost on writes in SSD applications also.

Thanks,
Kevin

GullLars  on June 3rd, 2010

Hi, and thanks for the answer Kevin.

You got me confused on a couple of points here with your answer:
“the ONFI-2 synchronous NAND interface, which is capable of achieving 200Mbps on a single (x8) NAND channel”
200 Mbps = 25MB/s (given 8 bits pr byte). Did you mean 200 MBps (200 MB/s)?
Is the x8 for 8 dies on the channel, or the channel being 8-bit?

IF it is 200 MB/s from a single sync ONFI-2 channel, then why does your C300 top out around 350 MB/s with 8 channels? Is the SSD controller the bottleneck for sequential reads?

I’m familiar with the parallell architecture of SSDs, and how you use interleaving and channel striping to write to all dies/LUNs. What i find inspireing is how you use the block abstraction to stream logical random writes as physical sequential writes to pre-erased pages/blocks, and just remapping the abstraction table, giving small random writes (with a queue) almost the same speed as large sequential writes.

When you posted your demo of the ONFI-2 PCIe cards back in 08′ doing >100K random read IOPS and >800MB/s, i had great hopes for comming generations of SSDs, but it seems to me the consumer/prosumer SSD manufacturers have been dragging their feet. This fall or possibly early next year i will be upgrading my old Mtron array, should i get my hopes up for a 6Gbps or PCIe drive from Micron with great read speeds and “good enough” write in a 2-5:1 sequential speed R:W ratio?

With 200 MB/s from each ONFI-2 channel, a bootable 8/10-channel PCIe 2.0 x4 card would be at the top of my wish list for christmas

Are you able to share expected ballpark performance figures for your 25nm 2-bit MLC?

I’m interrested in rough revised numbers i can substitute for the ones in the paper i linked to in my previous post.

If you can max an ONFI-2(.2) channel with 2-3 MLC dies, and they are 8GB each, you could breach 1GB/s read speeds with 6 channels and around 100GB capacity. That should allow keeping the cost fairly low compared to current 1GB/s+ SSD arrays.

Regards,
Lars

Kevin Kilbuck  on June 8th, 2010

Lars,
You are correct, I should have said 200MB/s. Technically speaking, we should refer to the interface speed in terms of mega transfers per second (MT/s). The ONFI-2 interface is capable of 200 MT/s per I/O pin. Our NAND products that support the ONFI-2 synchronous interface have 8 I/O’s (one byte), so the peak bandwidth of a single device is 200 MT/s x 8 bits, which is 1.6 Gb/s or 200 MB/s.

A single NAND channel from the SSD controller is also x8. The above are simple peak bandwidth calculations of the NAND. The bandwidth the host sees is a function of many things, including the NAND bandwidth, the host interface (SATA 3G vs 6G), NAND management overhead (commands, ECC, wear leveling, etc), firmware optimization, read-to-write ratio, sequential vs random data, etc.

kallidoan  on August 5th, 2010

Kevin,

Its been some time since the announcement. Are we any closer to seeing the 25nm NAND in some SSD Hard Drives this calendar year?

Thanks,

Randy

Kevin Kilbuck  on August 6th, 2010

Hi Randy,
I can’t speak to our SSD roadmap, but I do know we’re getting closer to 25nm SSDs every day. True, we’ve had production 25nm NAND for many months now, but we have pretty exacting standards for our SSDs. It’s a fact of NAND development that it takes some time to get peak performance and reliability out of any new process. Our 25nm NAND is performing great in lots of consumer applications, but SSDs demand much more than MP3 players and Flash cards do.

One of my earlier videos on TLC NAND explains a little bit about NAND development cycles and helps clarify this. When we do release a new SSD, we’ll announce it here—keep an eye on our blog for news and updates.

-Kevin