Thoughts on Nvidia: Ray tracing, deep learning startups and how semiconductor VC is like Biotech VC.

Image for post
Image for post

I thought I would share a few thoughts ahead of Nvidia’s analyst day around the RTX launch and my belief that the VC funded deep learning startups are not a significant threat to Nvidia’s franchise. Semiconductor VC is akin to biotech VC and the $1.25b invested just isn’t enough to be relevant.

This chart from Nvidia showing K/D (Kill/Death) ratios increasing regardless of skill (proxied by hours played per week) as the gamer plays the game with a more powerful GPU is critical to understanding the RTX launch.

Image for post
Image for post

Frames per second drives a higher K/D in shooters because you see your opponent before they see you as your GPU refreshes faster and is therefore more responsive to information coming from the servers. Gamers who play shooters pay up for faster FPS in the same way golfers pay up for better clubs (I don’t play but allegedly there have been big advances). My own personal experience as somewhat serious gamer is that my K/D is 35% higher on my PC (GTX 1080 ti with my new Blade 15 with an RTX 2080 arriving shortly) vs. playing on a level playing field on my Xbox One.

RTX gives less of a frames per second uplift than prior generations (35% vs. the GTX 1080 vs. 70% for the GTX 1080 vs. the 980) and DLSS has been disappointing so far (as pointed out by my favorite journalist focused on tech stocks) in terms of mitigating this. DLSS will inexorably improve as the algorithm is trained on more data, but today the top of the RTX product line is less of a “value” for shooter focused gamers in terms of FPS uplift per $ than prior generations. Recent price reductions on the RTX 2080 ($650) mean that the “value” has improved but this chart still directionally correct. Also interesting to see the “value” in the RTX 2060 and the GTX 1660.

Image for post
Image for post

However, frames per second are irrelevant above 60 FPS (movies run at 24 FPS) for games other than shooters. i.e. For games like Skyrim, Fallout, Assassins Creed, Red Dead Redemption, GTA, etc. And all of these games and their successors will eventually support ray tracing. Defining GPU performance solely as frames per second is *absurd* when it comes to games other than shooters. Ray Tracing is the biggest advance in the visual beauty of graphics since programmable shaders. For gamers who don’t primarily play shooters, RTX is incredible value.

Image for post
Image for post

And for shooter focused gamers, Nvidia is still the only game in town. The RTX 2080 outperforms its AMD comparable Vega VII despite being a node behind and devoting valuable silicon real estate to Ray Tracing. This is incredible engineering by Nvidia.

Image for post
Image for post

The Vega VII should outperform the RTX 2080 by at least 20–30% on a frames per second basis in non ray-traced games given it’s a node ahead and all of its silicon is focused on traditional GPU tech. This suggests Nvidia has at least a 25 to 35% architecture advantage over AMD. And the fact that AMD launched the Vega VII at price parity despite a performance disadvantage shows that they are committed to pricing discipline, which is a positive.

These dynamics — being good value for non-shooter gamers, still being the best GPU for shooters and potentially having relatively less crypto exposure — are why Nvidia’s share went up in the fourth quarter despite the “disappointing” RTX launch. And Nvidia gets its best chance ever to create a proprietary graphics standard around ray tracing and DLSS.

Image for post
Image for post

Semiconductor customers want buffer inventories to equal lead times. When lead times go up, inventories go up and vice versa. These inventory cycles are what drive semiconductor industry cyclicality today — not capacity cycles. Right now Nvidia is underearning as they are undershipping end demand as they work down channel inventories.

GPU lead times went up in 2018 because of crypto demand and buffer inventories went up to match these lead times. Nvidia thus overshipped end demand and overearned for several quarters (late 2017 through mid 2018). When crypto demand evaporated that cycle unwound: lead times came down, customers reduced buffer inventories which caused a further reduction in lead times and so on. Today, Nvidia is underearning as it is undershipping end demand which will continue until inventories are normalized. Long term investors should look through this period of temporary weakness to the true underlying earnings.

However, there are risks here. Nvidia’s ASPs are up significantly over the last 4–5 years. This pricing power has been an important part of the story. I think the real risk to the gaming business is that the ROI on crypto mining may have driven much of Nvidia’s recent pricing power. i.e. Gamers could afford to buy a better card because they could mine at night, in the winter, etc. when it was economical. (Winter being important b/c I have friends who turned off their heat and relied on their mining rig to heat the house). So it will be interesting to watch GPU pricing going forward. And, of course, crypto could always come back with Facebook’s upcoming Stablecoin a potential catalyst for this.

Moving on from gaming, a lot is written about the threat to Nvidia from deep learning startups and the $1.25b that has gone into funding them. $1.25b is nothing at the leading edge of semiconductors. This is very different from SaaS, consumer internet and more like biotech. The cost to design a leading edge semiconductor is $175m at 10 nanometer, $300m at 7 nanometer and $540m at 5 nanometer (roughly). Mask costs alone are 10s of millions. And you have to spend that steadily escalating amount every 2 years like clockwork.

Image for post
Image for post

$300-$500m is about the amount of money it takes develop a biotech drug. Tapeouts are much more predictable than they were (thank you EDA industry), but still a risk. The tapeout is like the phase 2 readout. If any of these deep learning startups miss on their tapeout, it is over. And if they have a great tapeout, they will still at best support 1–2 frameworks and a few algorithms. Nvidia supports all deep learning frameworks, all deep learning algorithms and their GPUs can be also used for traditional ML (via RAPIDS) and HPC. Higher utilization for Nvidia’s chips because of this flexibility = lower cost.

Supporting all of these frameworks is both difficult and expensive — witness how far behind AMD still is from a GPU software perspective. Nvidia generally made faster GPUs than AMD, but Nvidia’s software drivers were their real secret sauce in graphics — both more stable and more optimized (faster). i.e. Nvidia beat AMD in graphics partially due to better software engineering. Software is a core competence for Nvidia. The mental model I use for this is AI frameworks = game engines and AI algorithms = games. They can all be accelerated by recompiling code more efficiently.

This semiconductor treadmill is very hard to get on and even harder to stay on. “Tick, tock” sounds easy but it is really hard. And spending $500m plus every other year — just to develop the chip — let alone support it, market it, etc. is quite different than other tech industries.

And once successfully on this treadmill, the startups also have to compete with Google’s TPU and other forthcoming ASICs from cloud companies. Given Nvidia’s gross margins and the gross margins the startups have to target, these inhouse ASICs can be much worse technology and performance wise but still make economic sense. i.e. Nvidia really outperforms TPUs from a performance per watt perspective but it is still economic for Google to use TPUs in some cases given performance per watt per dollar (until the increased opex overwhelms the capex savings).

Lots of really smart AI people say encouraging things about the AI semiconductor startups. I would do the same! Nvidia having a semi-monopoly isn’t good for them. Some of the startups will succeed in specific use cases and take some share, but I don’t see major shifts in next 5 yrs. And of the startups, my money would be on Cerebras to succeed in training and a top secret co. in low power inference (I’m under NDA). For the ones that fail, I think will be tougher to sell to Intel given recent great engineering hires and more disciplined CEO (former CFO).

And all this is before Nvidia and Mellanox co-develop the next generation of interconnect and GPUs together. Mellanox was both accretive and strategic. Hearing more about the product roadmap is what I am most looking forward to at the analyst day.

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store