Nvidia shows signs in [2018]

Status
Not open for further replies.
Four US national labs to deploy Nvidia DGX-2 systems
November 12, 2018

Featuring 16 V100 GPUs split across two server boards, along with two Intel Xeon Platinum CPUs, 1.5 terabytes of system memory and 30TB of NVMe SSDs, the DGX-2 is capable of two petaflops of deep learning computing power. The 10U, 10kW system is interconnected by Nvidia's interconnect fabric, NVSwitch.

The systems will go to Brookhaven National Laboratory, in Upton, New York; Oak Ridge National Laboratory, in Oak Ridge, Tennessee; Pacific Northwest Laboratory, in Richland, Washington; and Sandia National Laboratories, in Albuquerque, New Mexico.
https://www.datacenterdynamics.com/news/four-us-national-labs-deploy-nvidia-dgx-2-systems/
 
Holy shit what a bloodbath. Lots of theories vindicated today. Turing was definitely priced to stay out of Pascal’s way. How insane was the crypto craze that these companies over estimated demand by such a significant margin.
 
How insane was the crypto craze that these companies over estimated demand by such a significant margin.
For several months on end they had well-financed miners beating down the doors to their board partners, trying to buy everything they could produce. Hundreds of millions of dollars in orders, with no end in sight and badly damaging the consumer market in the process. So the answer is "very, very insane."
 
In NVIDIA's Q3 results, the company confirmed "excess channel inventory of mid-range Pascal products", they likely got rid of the excess high end Pascal in Q3. And just like AMD they expect the excess GPUs to clear out in one or two quarters.

https://techreport.com/news/34264/nvidia-grew-q3-revenue-and-profits-as-inventory-clouds-gather

But AMD at least has some time to clear the inventory till Navi. Nvidia might have the rest of the Turing lineup ready and can't sell it because of the Pascal excess inventory. Starting 6 months late with a 12nm product might really hurt them, when amd launches 7nm shortly after. It all depends whether 2060 and lower are 12 or 7nm and on which timeframe they were planned.
 
Interesting Jen-Hsun Huang tidbits from Nvidia's Earning transcript.
The ramp of T4 is completely related to customers porting their model on top of our platform. And the inference model is really complicated. This is one of the things that I've talked about in the past that on the one hand, people think that inference appears to be simple because there are so many ASICs built being talked about. The vast majority of the complexity of inference is actually in the optimizing compiler on top.

The TensorRT, fifth-generation optimizing compiler that we announced just recently took 3, 4 years to build. And then on top of that, in order to get it to scale as quickly as what people saw in Google -- Google's cloud requires us to build something called a TRT server, an inference server that allows multiple models to run on top of Kubernetes in the cloud. That piece of software is also super complicated to write. And so the pieces of technology that we're putting together have come together.
...
And at this time and as we look out into the horizon, the T4 Cloud GPU is just unquestionably the most effective. It can run models, whether it's an image model or a recommendation model or a speech synthesis model. It is the highest throughput processor in the world at 70 watts, which fits into a hyperscale data center OCP server. It is also the lowest latency of any processor at inferencing in the world at less than 1 millisecond.
Clearing excess channel inventory will take a few quarters, similar to AMD's Earning statement.
To sum up, the crypto hangover has left the industry with excess inventory -- excess channel inventory. It will take 1 or 2 quarters to work through it.
https://seekingalpha.com/article/42...-results-earnings-call-transcript?part=single
 
In NVIDIA's Q3 results, the company confirmed "excess channel inventory of mid-range Pascal products", they likely got rid of the excess high end Pascal in Q3. And just like AMD they expect the excess GPUs to clear out in one or two quarters.
Is GP104 midrange? The latest nth iteration of 1060 is now using GP104 dies in an attempt to dump more Pascal stock into China.
 
I think both AMD and Nvidia:

  • Dramatically underestimated the amount of the inventory being purchased by the miners
  • Dramatically underestimated the impact of crypto frenzy petering out would have on the secondary market.

The first one is easy. While both have reported "high single digit" crypto-realted sales, I am highly confident that this was primarily mining SKUs or cards purchased from known miners directly form AIBs. There is not way they could have accurately gauged the considerable combined impact of mid-to-small scale miners who were buying cards by the pallet from the distributors or 10 at a time from electronics stores. I am convinced that, taken in aggregate, the overall mining demand accounted for as much a 30-40% of overall sales in certain months. The sudden drop in demand of this magnitude would be bad enough, but we are just arriving at second problem: once crypto mining became unprofitable, a lot of miners started liquidating their operations, flooding the secondary market with cheap cards. While it's not uncommon for their to be a spike in supply of used hardware, it is usually centered around a lunch of new product, when users sell their older cards to finance the upgrades. Not only are there literal tons of used Polaris/Pascal/Vegas for sale, but none of the proceeds from these will go to buy Turing or other cards, unlike the natural gaming-driven cycle. Nvidia's timing/pricing with Turing could have not been worse.

Cliff notes: Way more cards went to miners than AMD/Nvidia cared to admit, impact of crypto crash much worse than simple decrease in sales.
 
Meanwhile Nvidia stock is crashing like a rock, already by half in a couple of months.
A coincidence of multiple causes it looks.
Since the release of BFV raytracing patch on 14 Nov (or was that 15 Nov Quarter results/forecast) another cool 30 Billion dollar wipe, they better come up with a vastly better patch soon :)
 
Last edited:
Meanwhile Nvidia stock is crashing like a rock, already by half in a couple of months.
A coincidence of multiple causes it looks.
Since the release of BFV raytracing patch on 14 Nov (or was that 15 Nov Quarter results/forecast) another cool 30 Billion dollar wipe, they better come up with a vastly better patch soon :)
What does this have to do with ray-tracing?
 
Probably more related to inventory issues caused by cryptocurrency stuff, and the realisation that while NVIDIA is a very valuable company, it was probably overvalued.

Edit: whoops, too slow.
 
What does this have to do with ray-tracing?

Maybe stock watchers also read game reviews ?
After being overhyped with raytracing and Grays/s.
It's certainly not the only factor playing now and also affecting all big tech companies.
 
Last edited:
Tech stocks have been in a bubbles for the past several years. Bubbles burst.

Not sure if you've noticed but the global geopolitical situation hasn't been that rosey recently.

TBH I'd sooner point the finger at a looming trade war rather than poor scores in a game patch.
 
Tech stocks have been in a bubbles for the past several years. Bubbles burst.

Not sure if you've noticed but the global geopolitical situation hasn't been that rosey recently.

TBH I'd sooner point the finger at a looming trade war rather than poor scores in a game patch.

In Europe, we are a bit less affected by Trump madness.
 
In Europe, we are a bit less affected by Trump madness.

Capital knows few boundaries. Most of the big tech company stocks are listed on US-based exchanges.

If Trump is going to make it more difficult to build stuff in China and sell it in the US, the markets will care regardless of what happens in Europe and the rest of the world.

Even more so if some folks in the market take the view that many of the big tech companies have been a bit over-valued recently (eg. Apple, etc.).
 
Status
Not open for further replies.
Back
Top