With reference to Kioxia's presentation at IEDM, as commented on by Ian Cuttress on 30-Dec-2019, I have some comments.
My general point is that I think Kioxia is trying to do some spin doctoring here, to promote the idea that NAND flash is the way to go for Data Centre storage in general.
And to try to downplay the idea that SCM will start to catch up, and then pass it.
Why?
Maybe because it's not something they think they will be good at.
Or maybe because they already have a big capacity for NAND, and want to squeeze as much as they can from it.
The stakes are very big here, as you can see from their first slide.
In 2020, Enterprise is by far the largest sector, and they think it's going to almost triple in size in 5 years.
Whereas, PC's was the smallest by far in 2020, and although they think it will grow faster, by maybe 4-5 times, but is still 2nd to last in size, and thus is still dwarfed by Enterprise.
Entertaintainment and Mobile don't grow much.
So if you consider what Enterprise is replacing, HDD, it doesn't have to be anything very fancy in terms of speed relative to other SSD NAND, just better than the current HDD it's replacing.
You are not competing against SSD here, but HDD.
So for either process, SCM or NAND flash, the main advantage of more layers wrt HDD, is COST, or I think more accurately, COST/"endurance units".
And there is a difference depending on the R/W ratio of the data centre you are trying to replace.
My next point, is that the Kioxia fancy normalised and then normalised again graph is VERY misleading.
Smoke and mirrors if you think about the math behind it.
See Graph 1.
No wonder they didn't want to give out anything in detail.
Because it doesn't stand up to anyone who thinks very much about it.
If you take the graph, and instead of using Cost per layer, you look at Total Cost, you get a very different picture.
See my Graph 2.
The TOTAL SCM Cost not only increases quite fast as a function of adding more layers, but it accelerates as well.
That is, it's a function with a reasonable squared term as well.
See the power linear regression I did for it.
Wherease, the NAND Total Cost is quite flat, and does not rise very much.
In linear terms, it rises one tenth as fast as the 3D SCM cost does.
And does not accelerate very fast (as in, the squared term is very small).
OK, so what?
Well, they've done a trick as well.
Is it correct to say that the cost of one layer of NAND is equivalent to one layer of SCM?
No.
Not now.
Let's say one layer of SCM is twice the cost of NAND, just to illustrate my point.
So you add 1.0 to the function for Total Cost SCM, and then to get it back to that graph they want us to believe, you divide by number layers.
This gives you Graph 3.
Aha!!!
That tells a COMPLETELY different story from Graph 1.
OK, so SCM cost starts off high, we know that.
But then, using the CURRENT trends, as you ADD layers, it drops, and bottoms out at TEN layers.
And not FOUR.
And after than, it does not rise very fast.
Compare it to the same curve they gave us NAND flash: Graph 1.
OK, there is an offset at the start, but it doesn't start to pull apart very fast.
And the point to take home from THAT, boys and gurlz, is that you just need an overall unitised drop in cost for SCM, or rather a single overall cumulative increase in efficiency of SCM to close the gap.
Or perhaps a lot of small things that add up to it.
And you do NOT need a CONTINUOUS increase in efficiency.
That would be harder to get.
And since SCM is still new, that's to be expected.
Rough edges, round them off, add a bunch of small improvements, and you close that gap.
Nothing really drastic or game changing needed.
And if you read the R&D papers, that's what they say too.
The SCM processes are OK, just young.
Therefore, don't put your money on Kioxia NAND to dominate the Enterprise market in 2025.
I think SCM will be a good bet.
My general point is that I think Kioxia is trying to do some spin doctoring here, to promote the idea that NAND flash is the way to go for Data Centre storage in general.
And to try to downplay the idea that SCM will start to catch up, and then pass it.
Why?
Maybe because it's not something they think they will be good at.
Or maybe because they already have a big capacity for NAND, and want to squeeze as much as they can from it.
The stakes are very big here, as you can see from their first slide.
In 2020, Enterprise is by far the largest sector, and they think it's going to almost triple in size in 5 years.
Whereas, PC's was the smallest by far in 2020, and although they think it will grow faster, by maybe 4-5 times, but is still 2nd to last in size, and thus is still dwarfed by Enterprise.
Entertaintainment and Mobile don't grow much.
So if you consider what Enterprise is replacing, HDD, it doesn't have to be anything very fancy in terms of speed relative to other SSD NAND, just better than the current HDD it's replacing.
You are not competing against SSD here, but HDD.
So for either process, SCM or NAND flash, the main advantage of more layers wrt HDD, is COST, or I think more accurately, COST/"endurance units".
And there is a difference depending on the R/W ratio of the data centre you are trying to replace.
My next point, is that the Kioxia fancy normalised and then normalised again graph is VERY misleading.
Smoke and mirrors if you think about the math behind it.
See Graph 1.
No wonder they didn't want to give out anything in detail.
Because it doesn't stand up to anyone who thinks very much about it.
If you take the graph, and instead of using Cost per layer, you look at Total Cost, you get a very different picture.
See my Graph 2.
The TOTAL SCM Cost not only increases quite fast as a function of adding more layers, but it accelerates as well.
That is, it's a function with a reasonable squared term as well.
See the power linear regression I did for it.
Wherease, the NAND Total Cost is quite flat, and does not rise very much.
In linear terms, it rises one tenth as fast as the 3D SCM cost does.
And does not accelerate very fast (as in, the squared term is very small).
OK, so what?
Well, they've done a trick as well.
Is it correct to say that the cost of one layer of NAND is equivalent to one layer of SCM?
No.
Not now.
Let's say one layer of SCM is twice the cost of NAND, just to illustrate my point.
So you add 1.0 to the function for Total Cost SCM, and then to get it back to that graph they want us to believe, you divide by number layers.
This gives you Graph 3.
Aha!!!
That tells a COMPLETELY different story from Graph 1.
OK, so SCM cost starts off high, we know that.
But then, using the CURRENT trends, as you ADD layers, it drops, and bottoms out at TEN layers.
And not FOUR.
And after than, it does not rise very fast.
Compare it to the same curve they gave us NAND flash: Graph 1.
OK, there is an offset at the start, but it doesn't start to pull apart very fast.
And the point to take home from THAT, boys and gurlz, is that you just need an overall unitised drop in cost for SCM, or rather a single overall cumulative increase in efficiency of SCM to close the gap.
Or perhaps a lot of small things that add up to it.
And you do NOT need a CONTINUOUS increase in efficiency.
That would be harder to get.
And since SCM is still new, that's to be expected.
Rough edges, round them off, add a bunch of small improvements, and you close that gap.
Nothing really drastic or game changing needed.
And if you read the R&D papers, that's what they say too.
The SCM processes are OK, just young.
Therefore, don't put your money on Kioxia NAND to dominate the Enterprise market in 2025.
I think SCM will be a good bet.
Facebook
Twitter