Monday, August 7, 2017

Resistive RAM (ReRam) is Here in 2017

Resistive RAM (ReRam) is still filling up Journal and Conferences with Papers. But you can actually order it and see demonstrations now. 

2017 is a transition year.

Storage Class Memory applications have gained a lot of publicity with Intel/Micron Announcement of 3d Xpoint/Optane. Resistive RAM from Crossbar, Inc and WDC and 4DS and others has similar cost structure and potentially better physics, scalability, and performance. 

We look at how ReRam technologies compare to SLC NAND/Fast NAND, DRAM, 3D XPoint, and how the costs scale over time with new geometries. Key Takeaways:

  1. 3d Xpoint and Optane are creating new markets for everyone. Customers want choices
  2. All technologies are cheaper than DRAM with any reasonable maturity
  3. Fast NAND/ZNAND is available today and while not new or as fast, it is based on SLC 3D NAND which is quite mature
  4. Crosspoint ReRam is cheaper than 3D Xpoint once scaled to 20/14nm and has ability to scale lower
  5. ReRam technologies can be integrated into foundries with relatively minor changes. Intel Micron has reports "100 new materials in 3D Xpoint"
  6. Crossbar, Inc ReRam is available at foundry today. 4DS showed data with performance and yield on demonstration vehicles. WDC has said ReRam is SCM of choice
  7. Fast speed, 1M cycles, competitive cost is achievable today from ReRam. 
See our presentation at Flash Memory Summit, Wednesday 8/9 at 3:20 PM. Session 203A to find out how competitive ReRam is. 

Mark Webb

Top 5 NVM Myths Going into Flash Memory Summit

Top 5 Myths going into 2017 Flash Memory Summit. NAND, Xpoint, China, HDDs, TLC. We will see if I need to correct myself later!

  1. 3D NAND isn't working and this is causing shortages. 3D NAND is yielding. 3D NAND is working. It is shipping TLC. It is not perfect and various output metrics are still improving. But bit shipments in 2017 will be higher than predicted in beginning of 2016. Shortages are due to increased demand and long lead times to add capacity. 64L will address shortage. It works!
  2. 3D XPoint will replace DRAM or NAND. I hear people quote Intel or Micron saying this. I can't find a time where they actually said that. It could augment NAND or DRAM but the fact that it is 10x more expensive than TLC and 5x slower than DRAM will make any replacements trivial.
  3. SSDs have replaced HDDs. An oldie but a goodie. SSDs are great. Yet even today, most PCs ship with HDDs. HDDs out sell SSDs 3:1 in units and >10:1 in GBs. And the price differential between SSDs and HDDs has gotten worse since FMS 2016. SSDs are better... but we still have some work to do!
  4. TLC is not reliable enough for datacenter/Enterprise. NAND companies and SSD Companies know how to manage TLC specifics, DWPD, etc. TLC can work well, meet all needs, and is extremely cost effective. TLC needs to grow in enterprise!
  5. Chinese will build (enter huge number) dollar Fab and take over NAND/DRAM. Not in 2018, 2019, 2020, 2021. At some point Chinese investment will become dominant. But unless they have a JV deal with established company, it will be 5-10 years before they are 5% of Market. NAND isn't easy, even if you hire great engineers and executives. Someday... but today is not that day.

Mark Webb

Friday, August 4, 2017

3D Xpoint: Technology Review, Markets, Plans

3D Xpoint Technology, Markets, and Plans for the Future. Intel released Optane products based on 3D Xpoint. We reviewed the technology, processing, product performance and cost.

Flash Memory Summit will have two sessions covering aspects of 3D Xpoint plus many more on persistent memory and new memory technologies. Topics covered:

  • What do the technology teardowns show about 3D Xpoint
  • What type of memory is it
  • What does it cost
  • What are the best applications for it
  • When do we get to see 1000x performance in DIMMS
  • Is NAND Dead? is DRAM Dead?
  • How much revenue will companies make on 3D Xpoint in 2017-2020
Just when they say it's not PCM.... guess what.... it's PCM!
The chip has lots of opportunity for cost reduction and optimization
The speed in current applications with NVMe is 4-7x faster than NAND
It's costs less than half of DRAM per bit and 2x SLC NAND per bit.... If optimized
The fastest Cache SSD and the Fastest NVMe SSDs in the world are both 3D XPoint.
It has lots of error correction/redundancy and firmware to manage it.
Revenue won't be what I predicted last year.... I have the new numbers!

Tuesday Aug 8, 2017 3:40-6:00PM Flash Memory Summit  . Multiple great panelists!

Mark Webb

NAND Cost and Roadmap to 128L and QLC

What is the NAND cost reduction as we go from  48L to 64L to 96L TLC. What about QLC? What about 128L

The numbers are in: NAND costs are coming down, even as prices increase.

64L TLC is a major cost reduction and will lead to widespread adoption and volume ramp of 3D NAND.

96L will follow one year later, with string stacking but still provide cost and density improvements.

96 QLC will ramp in parallel with TLC for applications that demand low cost and can tolerate performance.

We will show cost comparisons for each technology for 2017-2019 at Flash Memory Summit 2017 Wednesday Aug 9, session M-22. Also will show NAND vs ReRAM vs 3D XPoint costs in other sessions

Call to set up meeting to discuss which company is leading in cost and performance

Mark Webb

Friday, July 7, 2017

What is the cost reduction from 64L 3D NAND

We just updated our models for 3D NAND, (and DRAM, 3D Xpoint). The results show major impact from 64L shipments

  • Includes wafer cost for capacity added in new or existing factory
  • Includes test cost and packaging cost
  • Includes up to date yields
  • Includes cost as it is ramping (not fully loaded) and at maturity
  • Includes impact of string stacking and CMOS under the array


  • 64L provides 30%+ cost reduction at mature yields. 
  • 3D overall and 64L in particular is a major cost reduction. We have the margins, Costs, prices to back this up. Don't believe the people saying its too expensive. The data is clear
  • The 3D wafer cost is much higher than 2D NAND. Some people have said it doesn't use advanced lithography so it is cheaper. They are unclear on the details of output per tool and per fab. But the bit output and scaling more than compensates for increased wafer cost!
  • Differences between companies on CUA, string stacking, Charge trap, and die size are all taken into account.
  • 64L is where all companies go ALL IN with 3D NAND. this will provide significant bit output increase and cost reduction.

See our presentations at Flash Memory Summit or call us for More information

Wednesday, March 29, 2017

Optane 3D Xpoint for Enterprise and Consumer

Optane 3D Xpoint for Enterprise and Consumer

Intel announced more updates on status of their Optane products. I wanted to say launched but then I realized that actually no product is yet available to the public... soon it will be I hope.

1) Enterprise P4800x SSD: Its arguably the fastest SSD in the world depending on which metric you choose. Intel's focus is on low latency... which is important since transactions involving a queue depth of  32 are quite rare. Its also expensive and only shipping in one density and limited quantities to limited customers. It also supports a virtual memory activity where it looks like memory to the server. Still waiting on third party benchmarks but at this point it is fast and shipping somewhere for revenue.

On the downside, the internet is exploding with "it was supposed to be 1000x faster with 1000x endurance!!!!" and "I waiting 18 month for this?!!!!" .

OK its 8-10x faster and the cells last 2-3x as long as NAND. Its still a great drive for certain Enterprise Applications.

2) Intel Optane memory: This was announced to press at same time... but embargoed for a few more days. 16/32GB cache device announced also at CES.

As I mentioned when this came out earlier at CES... caching is a great idea. 95% of the performance at much lower cost... when it works. Intel showed data that when combined with a HDD is improved boot and program loading performance. There is some concern that these metrics were cherry picked but we will see when third party benchmarks comes out (THG???? Anandtech???)

The price was good $44-$77, it makes a difference in performance, and Intel is hyping the heck out of it. If PC OEMS push it... it could be a big deal. I am not sure what percentage of chipsets accept it today, but I will try it out ASAP when it comes out (See... Intel even tricked me into buying a Kaby Lake PC with 200 Chipset!).

It it better than more DRAM? or a 256GB SSD? who knows for now.

This just in... Optane, as a NVMe SSD/Cache is not 1000x faster and is not going to change the world like the steam engine, the internet, or the cell phone (SORRY) ... but it appears to be a option for fast performance in some applications ... if you can buy it somewhere.

More to come when it actually appears!

Mark Webb

Wednesday, January 18, 2017

Is Optane + HDD a good thing? A great thing?

Intel (and PC OEMs) recently announced that sometime in Q2 we can use a 16GB or 32GB Optane NVMe Module with a HDD to achieve SSD like results. Optane module is connected by M.2 to PCIe bus. It is not a DIMM and it does not act like DRAM memory.

Is this a good thing? Is this a great thing? A Couple questions have been raised

1) This is caching again. In a perfect world, caching would make it so that 95% of storage access would be the cache speed. As a person who runs SSDs, hybrid HDDs, HDDs with a SSD cache, I can say this about caching: "its is real fast some of the time" and "it is HDD slow more of the time than I would expect".  Will this version work better? If so, we will have the blazing SSD like performance 90% of the time. What will the benchmarks and boot time say?

2) What will it cost? I have detailed models for how much that M.2 Module will cost including controllers, PCB, BOM components and the 3DXP memory.... we will save that exact cost for later (Call me!)... but what will the price to consumers be? There are two possible options:

    a) It will be marked up a ton to account for it being new, rare, and desire to differentiate. I will go with a $160 adder prediction for the 32GB version option.
    b) It will be subsidized with a price to push people to convert to the new architecture and help  Intel sell processors. I will go with a $50 adder for 32GB option

3) Am I better just buying more DRAM? Today, the PC OEM is charging >$130 to go from 8GB to 16GB of memory. This is about 4x what you should pay... oh well. But if it were reasonably priced, 8G to 16G for $35 or 16G to 32G for $75 is going to make a big difference in performance. If subsidized price, go with Optane.... else go with more DRAM. More DRAM is the gift that never goes out of style.

5) When can I but this option? According to Intel, it is Optane ready today (but without Optane???).

For actual working, non vaporware Optane, Lenovo says April ... others say June.... others say "later". If Intel doesn't delay it again, and it sells in April, I will be buying that Lenovo PC with 32GB just to say I have been Optaned (or is it Optanized?).

Optane Cache .... April 2017.... Finally !!

Mark Webb