메모리반도체-Waiting For The World To Change:

exponential[èkspounénʃəl]~기하급수적인,급격한

fail-safe~안전장치의, 절대 확실한

compel [kəmpél]~강요하다, ..하게하다.

what if~ 이라면 어떨까? ~하면 어떻게 될까?

 

내년에 마이크론의 EPS하락을 예상하는 애널리스트의 분석에는

3분기부터 마이크론의 이익을 증가시킬 낸드 출하량 증가는 고려가 안된 것 같다.

 

2018.06.05 William Tidwell

Waiting For The World To Change:

The Memory Industry Confronts The Long Traverse To Memory-Centric Computing

마이크로프로세서의 속도증가율을 디램이 못따라간다는 사실을 우리는 안다.

Low memory industry PEs are currently chained to perceptions of revenue growth unpredictability.

The introduction and implementation of the memory-centric computing paradigm will lead to a fundamental upward rating of industry PEs.

Gen-Z is the best way forward to realize the promise of memory-centric computing.

“We all know that the rate of improvement in microprocessor speed exceeds the rate of improvement in DRAM memory speed, each is improving exponentially, but the exponent for microprocessors is substantially larger than that for DRAMs. The difference between diverging exponentials also grows exponentially; so, although the disparity between processor and memory speed is already an issue, downstream someplace it will be a much bigger one. How big and how soon? […]

우리 모두 CPU의 속도 향상이 디램의 속도가 향상되는 것보다 더 빠르다는 것을 안다.

둘다 기하급수적으로 속도가 향상되지만 CPU 속도가 디램 속도보다 더 빨라진다.

이 불균형이 예전부터 이슈였지만 어느 곳에서는 더 큰 문제가 될것이다.

We want to look at how the average access time changes with technology,

so we'll make some conservative assumptions; as you'll see, the specific values won't

change the basic conclusion of this note, namely that we are going to hit a wall

in the improvement of system performance unless something basic changes.” -

"Hitting the Memory Wall – Implications of the Obvious” –

William Wulf and Sally A. McKee, University of Virginia - 1994

_________________________________

“Now we see everything that's going wrong
With the world and those who lead it
We just feel like we don't have the means
To rise above and beat it

So we keep waiting (waiting)
Waiting on the world to change
We keep on waiting (waiting)
Waiting on the world to change” -

Waiting on the World to Change – John Mayer

Memory industry investors who have done even a little research into industry demand drivers have been hearing about this thing called “memory-centric” computing for some time now.

I remember hearing the phrase in a 2014 during a Micron (MU) Analyst Conference

when Mark Durcan used it to support his assertion that the trend in computing overtime would be to favor

the value of memory over other system components.

I recall being quite intrigued by the concept and, in doing further research on the term,

found that it wasn’t all that new. Researchers such as William Wulf and Sally McKee (quoted above) had coined

the term “the memory wall” as far back as 1994,

and since that time the notion of “memory-centric” computing had been seen as a solution to this problem.

메모리 산업 주식에 투자하는 사람은 '메모리중심 컴퓨팅'이라는 것에 대해서 들어봤을 것이다.

나는 이 말을 2014년 마이크론 사장이 다른 시스템 부품과 비교해서

메모리반도체가 갖는 우월한 가치를 주장하기 위해서 했던것으로 기억한다.

1994년 윌리엄 울프와 샐리 맥키같은 분석가들이  '메모리 벽'이라는 말을 만들어 낸 이후,

'메모리 벽'을 해결하기위한 방법으로 '메모리중심 컴퓨팅'이 주목받았다.

A quarter century away now, we’re finally getting very close to solving the problem of the memory wall.

We’re actually going to implement memory-centric computing and in so doing we’re going to solve

another intractable problem as well - the problem of memory industry revenue.

 

Stick with me here as we wade through some wonky material.

This upcoming technical revolution in computing is going to put money in memory industry investor’s pockets.

once folks understand what’s coming down I think it’s going to lead to

a decisive rerating of memory industry PE multiples.

To fully understand how and why that’s going to happen we need to first understand the problem –

memory industry revenue – and then we’ll dive into what’s going to happen technically

that will fix that problem. Here goes.

 

So the problem is revenue – more specifically, the reliability of revenue growth.

The memory industry historically has been very bad at this. Here’s a graph of industry revenues over time.

 

(source: Statistica - here) (2018 estimate from the author)

So revenues are trending up but the pattern is unmistakable – sudden climbs and long declines through time. Unsurprisingly, many supply-side analysts are predicting a fall from 2018 highs through 2019 and 2020. (I wonder where they got that idea?)

And when we contrast the revenue graphic with the bit supply chart things get even more interesting. In the chart below from the recent Micron Analyst’s Conference the blue line shows bit growth and the bars represent capex investment required to achieve that bit growth.

 

 

Industry DRAM Bit Growth Vs. CapEx (Source – Micron 2018 Analyst and Investor Meeting)

So is there a correlation between industry bit supply and revenue? No. The two, per se, have nothing to do with each other. Maybe this chart will bring it all into focus:

메모리 산업의 빗공급량과 매출은 상관 관계가 있는가? 아니다,이 둘은 서로 관계가 없다

 

It should be obvious by now that we’re looking at charts that, taken all together, describe a commodity business. Substitute “corn” or “soybeans” for “DRAM” and we could be seeing the same set of charts.

우리는 명확하게 상품 챠트를 보고있는 것이다. 디램을 옥수수나 콩으로 대체했을 경우에도 우리는 같은 모양의

챠트를 보게 될것이다.

So it’s not the amount of product that drives revenue. Nor is it the rate of change in industry supply, per se. No. The real key is simply the relationship between supply and demand, and the price swings that ensue when supply and demand get out of balance.

매출을 증가시키는 것은 생산량이 아니고 공급 증가율도 아니다.

중요한 요소는 공급과 수요의 문제이다. 수요 공급이 불균형일때 가격이 변동한다.

Now all this is hardly breaking news. Please excuse me for this little exploration of the obvious but I thought it might be instructive to remind ourselves of this history. And the reason I did that is because I still hear plaintive cries on our board that Micron’s PE is too low. Why, we ask, can’t it be higher than the absurd 5 to 6 handle that the stock has had for some time now? To which the answer is: See the charts above. The fact is, sell-side analysts who seem to reflexively talk about cyclic moves in industry pricing have history on their side. Those of us who are arguing that “this time it’s different” - well, what do we have on our side?

One thing we have for sure is industry supply discipline. We have made this argument in past articles and I won’t repeat that argument here. Some analysts still don’t see this and they’re flat wrong. This is a relatively new phenomenon, arguably only four-years-old – dating from the Micron acquisition of Elpida/Rexchip – and its effects have been clouded by the historic crash in PC industry demand in 2015/16. Nevertheless, it is real and the fact that we have only three suppliers of 93% of the world’s DRAM supply marks a fundamental change in industry pricing dynamics.

Analysts who miss this point have got their noses so deep in the charts above they literally can’t see the forest anymore. But that still doesn’t fully change the charts above – supply discipline on its own can’t fix the whole problem. We’re still stuck with the challenge of “value.” Commodities, after all, are fundamentally objects of commerce for which there is no intrinsic, differentiable, value. Commodities are fungible and one source of the product can be easily substituted for another. DRAM, while much more differentiated than a few years back, is still largely either one of two things – server DRAM or mobile DRAM. Beyond that fact, back in the heart of the “processor-centric” computing era of just a few years ago, memory had limited intrinsic value relative to the whole system. The processor/SOC combination was the key to PC performance with memory being largely a sideline – a budget problem that the OEMs solved by taking advantage of the relative lack of supply discipline within the memory industry.

The memory-centric computing revolution is going to stand that factor on its head. Memory becomes the value in the system – the design center of the computing system solution. That couldn’t be true in the old paradigm because memory was the creature of the cpu/SOC. The memory/CPU/SOC combination was one architectural monolith that couldn’t be broken apart. Memory innovation remains captive to the processor/memory/memory bus monolith, and has to be designed and implemented in tandem with the other two. Think for a moment of the delayed introduction of 3D XPoint. Optane “Apache Pass” DIMMS are delayed until the release of the “Cascade Lake” series of Xeon processors in 2019.

With the realization of the memory-centric computing paradigm, this monolith will be broken apart. That will allow memory innovation (and with it, memory differentiation) to proceed independently of the processor/SOC. And this in turn will establish and develop a much higher value profile for memory. This will be the final nail in the commodity meme stranglehold on memory company valuations. The memory-centric computing revolution is going to solve the “value” problem. Memory innovation will accelerate, and as it accelerates more and different types of memory will be brought to market. They will have different characteristics – stronger and relatively weaker aspects - and each can be stitched into a hybrid memory fabric that best serves the application requirement.

Finally, memory-centric computing is going to super-charge demand for memory. The argument I’m going to make is that there is a technological transformation on the horizon that will fundamentally change the character of demand. Quite simply, the current computing paradigm makes provisioning memory very difficult. We’re going to discuss the reasons why later on but for now let’s imagine a world in which that problem is solved. My contention is that once we have solved the memory-centric computing problem, memory consumers are going to buy a lot more memory. And in so doing, of course, companies like Micron and Intel (INTC) are going to sell a lot more DRAM, 3D XPoint, STT-MRAM, and whatever other new memories that accelerated memory innovation brings to market. The overall increase in demand levels will lead to a re-pricing of the industry PE multiple. This, of course, will be good news piled on more good news for Micron and its fellow industry participants.

So yeah, solving the memory-centric computing model problem is important. Let’s talk about how it’s going to happen. The following slide from a 2015 presentation by HP shows why we need to solve the problem: