NVDA: GTC Europe is a must watch

Investment Analysis Clubs Saul’s Investing Discussions
1

https://www.ustream.tv/gpu-technology-conference

It happened this past week in Munich, Germany. NVDA held another GTC Conference. Jensen have a 2 hour keynote as he usually does. The stock market was volatile with tech, including NVDA’s stock, dropping big. Yet NVDA, the company, marched on. Introducing new products yet again, expanding its TAM again, executing on its stategy to take over the world of computing. Well, that’s not really NVDA’s stated strategy, but that seems what is happening. I love watching all of Jensen’s keynotes. Jensen has an amazing way of making it easy for someone to understand.

Moore’s Law has ended. For several decades, CPU computing improved by 100x every 10 years. Then it stopped. Now there is Jensen’s Law. Jensen joked about this because it doesn’t yet have a name but it deserves a name. NVDA’s GPUs together with the entire stack of SW, AI algorithms, etc, etc. is advancing accelerated computing by 1000x every 10 years. This about that for a second. This is about 10x every 3 years. When a new technology or product comes along that is 10x better than the substitute complete disruption happens in about 7 years. This means that NVDA’s pace of disruption across computing is outpacing a normal 10x disruption time. NVDA’s pace of innovation is its real competitive advantage and it is why other new technologies can’t come in and disrupt NVDA. NVDA is simply moving too fast for any other computing technology or approach to insert itself. I don’t think it can happen because NVDA is already disrupting itself at a pace that by far exceeds what is minimally to disrupt (10x better with complete disruption in 7 years). By the time 7 years has gone by NVDA has already disrupted by 10x two times sequentially. So Jensen’s Law is 10x better than Moore’s Law. As long as NVDA’s technology improvement continues at its current pace and NVDA continues to systematically target and disrupt new gigantic markets, its revenue will keep growing at an amazing pace. People have talked about not inventing in NVDA because of the law of large numbers. NVDA is already a $150B market cap company so therefore its growth must slow. Just look at the markets NVDA is targeting: they are huge. Every few months NVDA finds another $10-20B market to target and the products it introduces are sooooo much better that companies must adopt them or they will let behind by those who do adopt them. This week NVDA introduced RAPIDS which is disrupting Machine Learning and Data Analytics. Watch Jensen’s keynote and see for yourself what RAPIDS is all about. Watch the demo to see what it can do and how much better it is than the alternative. The More You Buy, the More You Save. It is happening: NVDA’s products must be bought; the alternative is to get left behind. See which companies have already adopted it. See which SW is compatible with it. Now, Saul will probably watch the keynote after he reads this post. Then he may go buy some more shares of NVDA. Then he may sell after a few days. Saul, I’m teasing. But the objection that NVDA is too big to keep its growth pace? I don’t think it applies due to my arguments above. The objection that NVDA has to sell 60% more each year than the year before? Instead look at how big the markets are that it just entered in the past 6 months alone. Look at the improvements to its existing markets. Look at how it is maintaining its margins.

Now, near the end of the presentation Jensen talked about Xavier, NVDA’s GPU for autonomous devices, not just autonomous cars but anything that moves. It is a brain for anything that moves. It is not just hardware but the entire stack. NVDA is not a hardware company. Hardware is just how it monetizes. It is also starting to motive in other ways. Service contracts. Services such as CLARA and Constellation will be monetized as services. NVDA is enabling computing that is so disruptive to the status quo that it must be adopted. That is computing. But the really, really big opportunity for NVDA is just starting to be realized. Brains for things that moves. What is intelligence worth? What is it worth to make a device truly smart and independent? NVDA is beginning to launch new smart things into the world. Some might consider it new life.

I’ve posted a lot about NVDA in the past. Here is a link to a post that links to some of the interesting posts of mine over the past 1 1/2 years:

https://discussion.fool.com/nvda-summary-of-gauchochris-posts-33…

Chris

45 Likes
2

Now there is Jensen’s Law. Jensen joked about this because it doesn’t yet have a name but it deserves a name. NVDA’s GPUs together with the entire stack of SW, AI algorithms, etc, etc. is advancing accelerated computing by 1000x every 10 years. (Emphasis different than original.)

This is known as the Law of Accelerating returns and explained very nicely here by Ray Kurzweil. Ray is certainly the world’s greatest living inventors and currently is director of engineering at Google. Worth a read:

http://www.kurzweilai.net/the-law-of-accelerating-returns

Jeb
Long NVDA and GOOG
Explorer Supernaut
You can see all my holdings here: http://my.fool.com/profile/TMFJebbo/info.aspx

1 Like
3

Chris:Now there is Jensen’s Law. Jensen joked about this because it doesn’t yet have a name but it deserves a name. NVDA’s GPUs together with the entire stack of SW, AI algorithms, etc, etc. is advancing accelerated computing by 1000x every 10 years.

Jeb: This is known as the Law of Accelerating returns and explained very nicely here by Ray Kurzweil. Ray is certainly the world’s greatest living inventors and currently is director of engineering at Google.

Moore’s Law: CPU processing power doubles every 2 years. Five doublings every 10 years is then 16x in 10 years.

Moore’s Law has reached its limit and is now dead.

Now there is a new law in town.

Jensen’s Law: GPU (full NVDA stack inclusive) computing power increases 1000x every 10 years. That equates to about 10x every 40 months or 3 sequential doublings every 10 years.

Kurzweil is an interesting and brilliant man. I’ve heard him speak at a lecture that I attended and I’ve read his book, The Singularity is Near.

Jensen’s Law is its own law and, yes Jeb, I agree that you when you plot progress of Moore’s Law and Jensen’s Law together then you can see the Law of Accelerating Returns. Again, NVDA’s speed of innovation is its true competitive advantage and it is also an incredible moat…by the time a potential competitor has something new the 10x that they were trying to disrupt has now become 100x or 300x so NVDA is no longer disruptable. Someone posted a link to a presentation about the automobile disrupting the horse and carriage. The point is that when a new technology/product has a 10x benefit to the current technology/product then disruption usually occurs and is complete within 7 years. NVDA advancing computing speed every 40 months by 10x then means that NVDA will disrupt itself about every 3 years. However, currently NVDA is still disrupting CPUs as it targets markets and develops products/solutions for those target markets in a sequential and methodical manner. As an example, look at NVDA’s new autonomous machine product that is now available for developers. It can already drive cars autonomously. Now, think about this technology improving by 10x in 40 months and 100x in 80 months and 1000x in 120 months. What will happen? This box that fits in a car or a forklift or farm equipment or a large robot today will in 5 years be smaller, lighter, faster, and probably less expensive. In 10 years, the power consumption will be much, much lower and it will be cheaper, lighter, and tiny. The applications for “autonomous anything that moves” will expand by A LOT as the technology gets faster, smaller, cheaper, and with much lower energy consumption requirements. NVDA’s products are beginning to move from centralized applications to decentralized applications. And the decentralized markets are very much larger than the centralized ones. So when Jensen said on Cramer a few months ago that he sees a path for NVDA to grow 10x. I think the stock price will also grow 10x. NVDA is not currently subject to the Law of Large Numbers. When will NVDA get to a $1.5T market cap? I don’t know when, but it seems pretty clear to me that it will happen. As long as Jensen’s Law is intact and NVDA continues to execute as well as it has been on knocking down market after market (huge ones like the $100T automotive market, like to $100B medical imaging market, and the many other $10B+ markets that it has already targeted), then there is no reason why revenues can’t keep growing at an incredible pace.

Just my opinion that to me seems obvious.

Chris

22 Likes
4

Great post Chris.

And GTC Europe was very cool, as always.

It is amazing how pervasive Nvidia’s technology is. It is showing up everywhere. Have you seen the Agrobot?

https://youtu.be/bXQg_M7_6_E

5 min video about a strawberry picking robot powered by Nvidia Jetson.

They have created quite the ecosystem. From training to edge. Think about it. AI Deep Learning Training has gone from nothing to ubiquitous in like 3 years. And that is basically all on Nvidia GPUs inside Nvidia’s ecosystem with CUDA and Nvidia’s optimized sodtware Stack. The only accelerators anywhere that work with everything. Now that all of that foundation has been laid and the ecosystem is in place, it’s time to capture more markets where all that intelligence can be put to use. Agrobots everywhere. The potential is huge.

Deep Learning Training and high performance computing have been the drivers for the huge Data Center Growth to date. That business is still small to where it will be. Machine Learning with RAPIDs will add another layer. Been looking at posting about that but haven’t gotten to it. The quick is that Deep Learning relies on layers of algorithmic convolutions to solve a problem (“what is this object?”). That layered approach is where the term “deep” is derived. GPUs perform parallel computing and were ideal for this type of computing. OTH, Machine learning problems start with inputs from a software engineer and work with assumptions to solve(“if someone is this tall they probably weigh this much”). Until now that has been best accomplished in sequential calculations. RAPIDs has changed that so that the calculations can now be run in parallel. This of course greatly speeds up the performance. ML is responsible for the AI like Netflix recommenders for what shows you might like. Another layer of AI that Nvidia can now capture into its ecosystem. Probably as big as Deep Learning.

The next waves are inference and autonomous machines. I mean doesn’t that agrobot just show you how we’re scratching the surface here. These are tremendous opportunities, never available before.

Take a look at this. You build a data center of DGX’s for training your autonomous fleet you intend to launch in a year or two. But you’re also a multi billion dollar vehicle conglomerate that has all kind of typical business problems that could greatly benefit from powerful data analytics. Oh yeah, your sweet AI system can do that too!

https://blogs.nvidia.com/blog/2018/10/10/data-center-ai-infr…

I have no idea how market cap for NVDA will go. But I can tell you that they currently have about $13B/yr in revenue. They will have substantially more than that in 3-5 years. We have a lot of really great companies discussed here. I’m in about 13 right now. But I think hands down Nvidia’s potential revenue (3-5yrs) to current revenue is the widest gap of them all. And they have a 36% net profit margin. After spending $.5B/qtr on R&D. Yeah some kid with $10M in venture capital is going to come out of his closet lab and steal Jensen’s leather jacket.

Darth

14 Likes
5

Moore’s Law: CPU processing power doubles every 2 years. Five doublings every 10 years is then 16x in 10 years.

Nope.
Five doublings is 2^5 which is 32x.

Moore’s Law has reached its limit and is now dead.

Nope.
This is an incorrect understanding of Moore’s Law. First, of course, it isn’t an actual law. But if you google what Gordon Moore actual said it has to do which the “number of transistors” that you can put on a chip.

http://hasler.ece.gatech.edu/Published_papers/Technology_ove…

And it was originally stated as doubling every year, then maybe a decade later to every 18 to 24 months, then every 2 years. Now that doubling rate is slowing even more.

This bounty of transistors for chip designers resulted in a similar rate of increase in performance of general purpose CPUs – for a while. Part of the gains were due to architectural design improvements…and another part was due to clock speed increases that were made because as transistors shrunk they also got faster. That mostly ended a decade ago – leaving designers only gains from architectural improvements.
Most gains in CPU designs, except for adding more cores, has also slowed greatly.
When workloads are capable of massive parallelism, then GPUs are the best (currently available) technology. But advancing them still needs more transistors.
The actual end of Moore’s Law is also the end of advancement for GPUs on currently known silicon technology.

Jensen is smart enough to know this.

Mike

6 Likes
6

This bounty of transistors for chip designers resulted in a similar rate of increase in performance of general purpose CPUs – for a while. Part of the gains were due to architectural design improvements…and another part was due to clock speed increases that were made because as transistors shrunk they also got faster. That mostly ended a decade ago – leaving designers only gains from architectural improvements.
Most gains in CPU designs, except for adding more cores, has also slowed greatly.
When workloads are capable of massive parallelism, then GPUs are the best (currently available) technology. But advancing them still needs more transistors.
The actual end of Moore’s Law is also the end of advancement for GPUs on currently known silicon technology

Mike, I see that you are questioning whether NVDA’s historical 1000x improvement rate per 10 years can continue. Do you know how much of their past improvement has been due to transistor density increases on silicon? You seem to know a lot about it. How much more can NVDA’s transistor density improve to reach the state of the air and how much of their 1000x improvement in the past has been due to the same reason why Moore’s Law is ending?

Chris

P.S. my error was that there are only 4 2 year doublings in 10 years, not 5 doublings.

2 Likes
7

P.S. my error was that there are only 4 2 year doublings in 10 years, not 5 doublings.

Nope. My math says the 10 divided by 2 is 5. Therefore 5 doublings in 10 years.

Mike, I see that you are questioning whether NVDA’s historical 1000x improvement rate per 10 years can continue.

Let me answer that a few times each based on a different assumption.

  1. The big strength of GPUs over CPUs is the massive parallelism. The primary way that you implement parallelism in hardware is by adding more transistors to have more parallel data paths. You can only increase as fast as you get transistors by shrinking them. So, no, they can’t get 2^10 faster in a decade (that is 10 doublings). FAB technology would need to accelerate, rather than slow down as is really happening. Of course you could make chips bigger – but that only works one time and Nvidia already makes chips as big or bigger than anyone. Physics of defects is your enemy here as yeilds go down.

  2. Software and algorithms. Improvements in these are continually happening but don’t really count for many benchmarks…since you count how many floating point multiply-adds you can do per second, In AI there can be improvements such as using small data sizes, such as 16 or 8 bits. This makes the application run faster…which is good, but it didn’t really make the computer go faster. Even if you count this, it only provides a one-time speedup, since next you you can’t go to 4 bits, then 2 bits then 1 bit and get the exact same accuracy.

  3. Memory bandwidth. This is going to limit everything to some extent. You can’t just get 1000x faster in true performance without getting additional memory bandwidth. Sure, you can add bigger and better caches…you need lots of new transistors for this (see #1). You can’t just declare you want memory chips that are faster. First, Nvidia does’t make these. Second, laws of physics and heat still apply. New tricky designs will allow some specific applications (like AI) to have non-general purpose parallel computing tasks reduce the needed memory bandwidth – but this is probably a one-time 2x or 4x gain…not a compounding year after year improvement. Some of this is already in their latest chips. (For example, convolutions used in AI are many multiply-adds. One half of the parameters, the trained AI weights, are used over and over again so you can make a design that takes advantage of this – but you can’t get rid of the other half of the parameters – that is the data you are trying to analyze)

Mike

3 Likes
8

It is also worth noting that not all problems are suitable for parallelization. Some are obviously well suited for it and these really shine on a GPU, but there are others which are inherently serial and those will not be sped up by a GPU.

9

You can only increase as fast as you get transistors by shrinking them.

…or by making the chip larger. Part of Nvidia’s historical improvements included making larger dies. Note that Nvidia has already hit the reticle limit with a 815mm² die on the V100. I’m agreeing with Mike that the historical 1000x improvement included a bunch of easy gains that are very unlikely to be repeated.

1 Like
10

Note that Nvidia has already hit the reticle limit with a 815mm² die on the V100.

Why not go to 815mm^3?

Denny Schlesinger

11

Why not go to 815mm^3?

How do you get the heat out?

Have you seen the size of the cooling on the 2D parts?
You’ll need a heat exchanger 800x as big for what you propose.

I think that wafers are thinner than 0.5 mm, typically.
Even assuming 1mm, you are suggesting ~800x as much heat. The middle of the chip would probably melt. I guess you could run the clock speed at 1/10th as much and/or add some active cooling…a serious engineering and yield issue.

Note that HBM memory is already stacked 8 layers high. And this is placed on an interposer to the side of the GPUs to aid in cooling. And memory chips do not generate nearly as much heat as logic chips.

So going 3D in chips is defintely happening…but probably at a rate of 1/10th of Moore’s Law for traditional 2D layout…just as a guess. (Moore’s Law would suggest doubling the number of layers across the board every 1.5 to 2 years.

Mike

4 Likes
12

How do you get the heat out?

I forgot all about that! How about a liquid nitrogen bath? Superconducting semiconductors?

I better stick to what I know! :frowning:

Denny Schlesinger