Friday, May 15, 2020

When is a NAND stack not a stack?

With some recent technology advances in NAND memory, I am seeing a lot of confusion on the use of the term "stack". I will attempt to give a simplified view on different definitions.

If you are a wafer processing expert, you can ignore this overview blog and we can argue details and pros/cons by Zoom.

  • Die stacks: This is where we stack one NAND die (a silicon chip) on top of another, wire bond it out and put it all in one package. we have been doing this on memory for well over 20 years. 16 die stacks are available from everyone so if our base die is 512 Gbits, we can get a 8Tbit (1Tbyte) package.
    • Pros: more memory in a package 
    • Cons: cost is cost of 2 die plus packaging and test cost (its not a cost reduction)
  • 3D NAND layers: These are not stacks but some people still use the term. NAND Layers are "stacked up" on wafer in a series of depositions and then a channel hole is made through the "stack". we are shipping ~96 Layer NAND today 
    • Pros: as you go from 64 to 96 Layers you get ~50% more bits for an additional process 10% cost adder (oversimplified). so its a big bit cost reduction
    • Cons: etching that hole and filling it and making it all work is not that easy (see next)
  • String Stacking: If you want to go from 64 to say 128 Layers but you cannot etch and fill the hole, you can do it in two steps. Do 64L, dep and etch hole and fill. Then do another 64L dep etch and fill hole. This is string stacking and everyone but Samsung has been forced to do this so far. Samsung will do it in the future.
    • Pros: you get your layers increased and lower cost
    • Cons: two step approach is more expensive than single stack (say 5%)
  • Wafer bonding: Instead of doing die stacks, you can bond (call for details) two wafers face to face, then dice them. This is used for wafer that are not the same technology process. YMTC is now doing this on their Xstacking process for 64L. 
    • Pros: you get logic and array "stacked" on top of one another for small die size and each wafer optimized
    • It is expensive, complex, and requires additional post stack processing
There are processes with interposers like HBM used for DRAM, TSVs, stacking DRAM on NAND and a processor.... etc but I will save those for a different blog. 

I prefer the terms I use, but it is important to not get these confused as the benefits and cost are completely different

I can go through details on cost for each and how they are done and complexities. These provide much of the competitive strengths of each company. Call/email for more info

Mark Webb

Thursday, April 2, 2020

Is memory like water, hand sanitizer, or TP?

While shopping recently, the interesting human behavior got me thinking: "which product matches DRAM and NAND supply and demand?".  Is it toilet paper, hand sanitizer, or water?

A few weeks ago, I went to a Target shopping for basics. I don't shop at Costco and buy a month ahead of time, I get a few times a week and just buy what I am about to run out of (I am a big fan of vendor managed inventory). I went to the paper towel and toilet paper aisle and it was empty. I had no idea why. I called a friend and said "is Target going out of business? whats up?" I was informed that there was a run on products due to COVID-19. Luckily for me, the grocery store down the road had tons of paper products, so a picked up some.  A few days later obviously things got a lot worse.  My reaction was "go to store, get a ton of everything, this is a survival competition.... whoever gets the most stuff wins " ... I am an engineer ... but I am still human. Note: I have literally seen the supply chain manager for a hyperscale company do the exact same thing.

In the 2-3 weeks since, supply and demand forces kicked in and.... from this...  we can learn a lot about commodity markets like NAND and DRAM (Yes, kids, memories are a commodity).

Bottled water was a quick one to go from all  the stores.  "what? there is no water? then I need more!" I was at a Dollar Tree where a customer came it and said "I want to buy all the bottled water in the store" ... the manager spent the next 30 minutes loading it into a truck. We had a shortage for about 2 weeks and then... funny thing happened. My grocery store stacked water to the roof on 3 aisles. No more shortage. No more people buying pallets of water.

SUMMARY: Real demand didn't increase. People are not drinking more water. The fear increased so people want more and bought tons to add to their inventory. The greater the shortage, the more they want. Once the supply increased, the fear was gone, people say "OK I will just buy water when needed."

Toilet paper:
Toilet paper is running short and last time I was at stores, the shelves were still empty.  It is being rationed and people are ordering on Amazon with a deliver time of late May. Unlike water, stores have not been able to increase supply fast enough so the panic continues.

SUMMARY: Real demand didn't increase. People are not using more TP. The fear increased so people want more and bought tons to add to their inventory. The supply did not not increase, the fear is still there. When this is over, what do you think will happen to TP sales?

Hand Sanitizer:
Hand Sanitizer is running short and last time I was at stores, the shelves were still empty.  People are ordering on Amazon with a deliver time of June. People want more and unlike water, stores have not been able to increase supply fast enough so the panic continues

SUMMARY: Real demand  DID increase. People are using more Sanitizer .The fear increased so people bought tons to add to their inventory AND to use. The supply did not not increase fast enough, the fear is still there. When supply is there, people will probably use more hand sanitizer for a year or more.... maybe forever.

These give examples of what happens with memory. Because price gouging is illegal in groceries (but not memories???) price didn't affect most groceries. When there is a shortage, everyone wants more to store for inventory. Sometimes supply can adapt, sometimes it cannot. Sometimes real demand increases.... often it does not. Lots of discussion on what is happening today and next year in DRAM and NAND to follow up on this.

Next we can discuss supply dynamics of Masks, ventilators, gloves and whether they should or should not (considering microeconomics input only) build more factories. Spoiler alert: It is like NAND, NOR and DRAM capacity planning.

Mark Webb

Thursday, February 6, 2020

Top Memory/Storage items from Recent Earnings

We had earnings announcements from Seagate, WDC, Hynix, Samsung, WDC, STX, Intel, AMD and others. Lots of new data and predictions for you. 2020 is not going to be dull.

  • Everyone says memory cycle downturn has ended. But how fast will we recover and how good will it get? Will DRAM prices skyrocket? Should we party like its 2018? 
  • Supplier inventories are “near” target (except for Micron who is stockpiling for NAND overhaul). Now FINALLY we have a fair starting point for pricing discussions and supply agreements. And customers need to plan their inventory level targets.
  • Speaking of customers…. Datacenter Giants (not to be confused with the Datacenter Dinosaurs who also exist) will buy and build big…. Then take a hiatus and digest the infrastructure build much to the dismay of CPU and memory companies… then build again.
  • Everyone’s 2020 bit demand growth is approximately aligned (within 5%) but is the number good or bad for EPS growth?
  • WDC and STX are making more money on HDD and the technology is marching on. At this rate, HDD bit ASP will be ZERO by 2025 (just kidding, I should have used a log graph). But we ARE poised to have the SSD to HDD bit price ratio INCREASE just like in 2017.
  • Optane DC persistent memory (DIMMS, 3D XPoint)  IS  cheaper than DRAM and has great performance….  but is still growing slower than expected despite Intel success in Cascade Lake ramp. When will this change? When will CXL help here? 
  • New DRAM and NAND memory technologies are coming out and how they achieve cost reductions is different for each company. The NAND cost leader will change in 2020
  • China Virus and trade issues will either hurt or help memory manufacturers depending on who you ask. We predict the average of those two scenarios!

Mark Webb

Monday, January 20, 2020

5 Things to Look for in Persistent Memory this week!

With Persistent Memory Summit coming up this week, I am posting Five items for people to look for in evaluating persistent memory solutions and trends. I will be at the Summit so please feel free to argue with me in person!

  • A couple years back at the PM summit, Jim Pappas was adamant on stating that persistent memory is here today, right now, and is not a mythical future item (which it really had been for many years). I love that focus. There is enough real stuff to look at without getting sucked into tomorrow’s possible technology. Review where NVDIMMS, RDMA/NVMoF, Optane/3D XPoint, etc are today and how they are being implemented. We can discuss the mythical future technologies afterward over drinks!
  • At Flash memory summit, we provided our definition of persistent memory which references a PM summit paper from 2 years back. The key is that it needs to be persistent, it needs to be accessed like memory and it has to have a reasonable, memory like, latency. I love fast SSDs, but fast SSDs and virtual memory mapping are not persistent memory (IMHO). Look for true PM attributes and applications.
  • Intel has put a tremendous amount of work and marketing into Optane DIMMS (Apache Pass, Optane DC persistent memory). Cascade lake has many skus to take advantage of these and Intel has product availability and inventory to support it. Sales of Optane and the use of true PM (app direct mode) is a test case for future market revenue projections. Look for what Intel and their customers are saying about sales and ramp of Optane DC PM. We projected in August 2019 that Intel DC PM sales (DIMMS only) would grow from $700M in 2020 to >$2B in 2024. It FEELS like Intel is slow to meet that revenue projection… but lets see what Intel says when asked (Hint: Earnings announcement is same time)!
  • Intel builds on a custom protocol and uses its memory controller to manage persistent memory. The rest of the industry has protocols planned for a more open architecture. NVDIMM-P, CCIX, GenZ, OpenCAPI, CXL are proposed to allow multiple technologies to achieve memory like access and speeds. How are those doing and who is implementing solutions today? Who will win?
  • Applications! Everyone is looking for the killer app…. We know of some specific applications of persistent memory (large datasets, quick recovery from reboot, journaling info) but what are the applications that will make is so that 50% of servers MUST have persistent memory? Look for the volumes required and the timing for those.

As you would expect, we have summaries and opinions for most of these topics and more topics like cost, pricing, technology timing and scaling, revenue, etc. Please call or text to discuss.

Mark Webb

Friday, January 10, 2020

Memory Pricing: Myth, Math, and Reality

With the bottoming of the memory market and reports of NAND price and DRAM price increases, People want to know how all these numbers pan out. No one knows for sure...  But…there are some market forces and reasons why numbers vary

Updated thoughts on pricing models and how to interpret

  • Memory is a commodity. Period. 
  • DRAM exchange/ Inspectrum is usually directionally correct, but wrong in magnitude and in absolute numbers when comparing to actual company ASP
    • Spot pricing is open market and not related to what major customers pay
    • Customers react differently in magnitude to increases and decreases in exchange price
    • Even reported contract price is not REAL contract price and does not match company ASPs
    • Apple does not pay more in price because the exchange price increases unless they think they cannot get shipments. They prefer to get companies into a bidding war. When prices drop, large companies do expect suppliers to cut price…. With or without a pricing contract. And they usually do because it means there is excess supply
  • Exchange prices and analyst reports are often for a particular part. That part may increase in price but because it did, people bought a different, cheaper part.
    • Hypothetical example: Hynix 256Gbit TLC, 64L NAND price increases 5%. But the recently introduced 512Gbit TLC, 96L Gbit is 15% lower price per Gbit. And the 512Gbit 96L QLC part is 20% low price per Gbit. End result: analyst reported NAND price went up. Hynix company ASP dropped 10%. Both are true.
    • Confirmation example: In 2017/18 Micron did not report any annual NAND price increase. In fact, ASP dropped 7% in F2017, 8% in F2018. Similar data for other companies… Yet most people will still tell you NAND prices went up 30-50% during that time. DRAM did Increase 50% during that time for Micron.....  However all exchange prices show 100% price increase
  • Actual pricing is a response to market growth, Inventory, and customer relationships. We can discuss how to track these and some basic conversion factors from exchange price to ASP.
  • The number one input to memory company profits is pricing.... so anything they can do (legally hopefully) to get pricing to increase will help memory companies. Unfortunately or fortunately, they are just not very good a collusion! but there are legal tactics that they do all the time.
  • Cost for each company matters. One model is that pricing always drops til at least one company loses money. Today, not all memory companies have the same cost. In 2021, there will be differences in cost with different leaders than 2019.
Where are we today?:

  • We had a shortage of NAND and DRAM through about Q4 2017 and Q3 2018 respectively. Since then both were in major oversupply. Inventory at customers and suppliers skyrocketed. Reports are that inventories are now nearing target. Suppliers report shortages on some line items.
  •  Well publicized supply limitations (power outage, Fires, Japanese chemicals, Reduced wafer starts) have people optimistic on pricing. DXI skyrocketed. All of these require discussion to see if they matter and how much.
  • Dramexchange has reported spot NAND price increases since April 2019, We are just now seeing companies report modest price increase (they are ALL losing money in NAND so they need to increase price).  Dramexchange has started to report spot DRAM price increases in the last 30 days.
  • It is a cyclical commodity market. (did I mention this before????)

So lots of positive reports on pricing these days. The key is how much will companies see and will it last. We track all of the metrics and can discuss how these impact different companies

Mark Webb
MKW Ventures Consulting

Thursday, December 19, 2019

Thoughts on Micron FQ1 Earnings Announcement

Micron announced earnings last night and gave updates on DRAM, 3D XPoint and NAND. Overall numbers were close to estimates. Guidance for FQ2 (Dec-Feb) were lower revenue and earnings due to seasonality. Some interesting items in the presentations.

Some Thoughts:

1) Micron increased NAND bits shipped significantly with slightly increased price. This is a positive. Often we see a big increase in bits at lower price ("we sold out") or flat trend in bit shipments at increased price ("we held out"). Since Micron and the industry overall is losing money on NAND, the prices need to be flat or increasing to get the market healthy. This is a good sign to get to break even and pay for development.

2) Micron continues to press forward on new DRAM technologies. For all the hype about 1Z, the fact is that most of the worlds bits are shipped on 1X or higher. Micron states they they will cross over to have most bits on 1Z/1Y in the next year. A recent Techinsights presentation also stated that Micron is about caught up with Samsung on cell size on new technologies. This has been a push from Sanjay since taking over and is a big success.

3) Earnings are expected to drop again in FQ2 for Micron due to seasonality and Sanjay stated that he is "optimistic" that this will be the financial bottom for Micron. The question is how fast will the upcycle improve and how good will it get this time. In case you were wondering, memory is a cyclical business for mathematical reasons that can be shown with an excel spreadsheet (it has a inherent root cause). For significant increase in earnings, we need to see higher memory prices in 2020. Which requires a shortage of memory. Which requires elimination of some inventory. Which requires supply growth lower than demand growth.

4) Micron floated some interesting number on the future of memory industry. DRAM Bit growth will be 15% in 2020, vs 20% in 2019. NAND Bit growth will be 30% in 2020 vs 40% in 2019. <20% ("mid to high teens") and 30% growth for DRAM and NAND respectively will be the future CAGR. These numbers are lower than historical but are becoming consistent from each supplier. Micron also states that DRAM bits per server will grow 20%  per year. All of these numbers include all of the hyped buzzword technologies. (AI, ML, Neuromorphic, 5G, Autonomous, etc). So again it is nice to have consistent bit growth numbers to plug into spreadsheets (with scenarios as well)

5) Micron gave some updates on NVM technology. Micron announced the X100 3D XPoint SSD earlier in the year and we gave our analysis of how this is built, components, and what the ramp might look like. Other than that, Micron was vague and discussed 3D Xpoint as a promising technology to be developed over the next few years. Micron also announced actions to deal with 3D XPoint which they said is declining in production volume and causing underload charges. At FMS, we presented details on 3D XPoint revenue and manufacturing for the next few years and we have new details since August and what Micron specific impact might be.

On the NAND side, Micron reiterated that 128L will be replacement gate and will have limited ramp on select products by end of Fiscal 2020. Until then, Micron will use existing technologies to support NAND sales. This is a great way to ramp a new technology but IF there is an upturn, it leads to problems on how (or if) to spend capex to meet increased demand... spend on old or new (unproven) technologies.

We give clients inputs on how to monitor all of this and whether actions and reports by companies are supporting or changing the models. Plus we have detailed cost models for all technologies comparing each company.

Mark Webb

Wednesday, December 4, 2019

5 Thoughts on China Memory Progress, December Edition

Recent headlines highlight ChangXin Memory (CXMT) plans to ramp manufacturing and also referenced YMTC plans in NAND. Lets “fact check” the status and announcements.

Chinese and all memory companies make optimistic claims stating 10’s of Billions have been spent. They show cartoons of factories that don’t exist yet or are empty. These claims are easy to challenge as companies track shipments to China and China reports equipment delivered by area. So it is very easy to verify the actual amount of capex and model where it is going. Lots of money is going into Chinese owned memory companies but it is no where close to $30B yet and there are not significant shipments yet.

  • Report: CXMT is running 19nm DRAM in production at 20K wafer per month and will ramp to 60K by end of 2020.  Production was started in September
    •   Facts: We know based on shipping reports that CXMT is spending $200-250M per month on tools. We know that this is roughly on pace to tool out 1/3 of a large fab. CXMT will have tools to run 20K wafer per month soon and is on pace to ramp those tools to a 40K per month pace at least. The hype over the last 2 years is becoming a reality… it’s a real fab with real tools now!
  • Report: YMTC is running 10-20K wafer per month of NAND and will quadruple output in 2020.
    • Facts: YMTC/Tsinghua is among the leaders in hype in the past. They were supposed to have technologies shipping for the past two years and have shipped little so far.
    • Facts: YMTC is also taking tool shipments at a high pace. 2018 showed tool deliveries of ~$1B dropping dramatically at the end before resuming in June 2019 at a $200M/month pace. We will revisit the plans to quadruple output by end of 2020 but it is possible with current spending level if it continues through 2020
  • Report CXMT has production 19nm technology being used for LPDDR4.
    • Facts: CXMT has made a number of public statement so we don’t need to guess. The technology is based on Qimonda, they hired former Qimonda people as well as people with DRAM experience from Samsung, Micron, and Hynix presumably. So they have a technology and some expertise
    • Rumor (or Alternative Facts!): A recent report online stated that the yields on this technology are very low and this was confirmed with other sources. This is to be expected but they might have some work to do
  •  Report: YMTC Technology is 64L Xstacking and will ramp to 128L in 2020.
    • Facts: We know they are installing equipment and sources state that have low volume and they are sampling some 64L. The technology might not be competitive but if they get shipments out and supply local customers, then profits don’t matter at this time. 
  • SUMMARY PREDICTION (Facts, Alternative Facts, Rumors):
    • CXMT and YMTC have their initial fab line set up and are ramping with large tool deliveries. This is new! They plan to ramp to 60K+ wafers per month by end of 2020. Both are running technologies that are not mature and are probably 1 year plus from maturity. We also have no public samples of either technology (despite my attempts to buy them!)
    • CXMT and YMTC Combined will be less than 2% of the bits shipped end of 2020
    • CXMT and YMTC Combined will be 5% of the bits shipped end of 2025
Help needed: Samples from either technology! THANKS!

Mark Webb