Can you guys expand the number of regions with storage? I don’t know what I’m doing wrong but the only storage available (Texas) never overlaps with available compute.
GPT-3 was another buckets worth of evidence in favor of the scaling hypothesis. Performance kept improving (and cost to train kept increasing) as more parameters were added. Even with 175 billion parameters, the performance had not yet plateaued. One take-away is that throwing a lot of compute at the problem helps tremendously :).
Tesla V100s have I/O pins for at most 6x 25 GB/s NVLink traces. So, systems with more than 6x GPUs cannot fully connect GPUs over NVLink. This causes I/O bottlenecks that significantly diminish returns of scaling beyond six GPUs.
This article provides an overview of their architecture that bypasses this limitation using additional high bandwidth links. Looking at the benchmarks, multi-GPU performance scales almost perfectly linearly from 1x GPU 16x GPUs.
I'm one of the engineers who worked on this project. Happy to answer any questions!
This is a one-line apt/aptitude installation for TensorFlow, PyTorch, CUDA, cuDNN, etc. When NVIDIA releases a new version of CUDA, you can simply apt-get upgrade to the latest version.
Do you support the 19.10 release? The inability to compile or package every version and variant of TF, and GCC9 conflicts with both the CUDA SDK and Tensorflow, is precisely why we created Tensorman.
Lambda Stack supports 16.04 and 18.04 at the moment.
It's value prop is to enable people to easily install TensorFlow / PyTorch and their dependencies in a container-less fashion. Though it doesn't provide the isolation of containers.
What I've learned from talking to customers is that many people don't care that much about handling multiple versions of the same framework. I wouldn't be surprised if you find that, like Lambda Stack, people are mainly using this product to easily get started with TensorFlow/Pytorch.
Now that TensorFlow 2.0 is out, we will see a much more stable API. People won't have to change their code if TensorFlow bumps up a dependency version. For many, this will reduce the impetus for moving to containers.
Install of PyTorch and CUDA is literally one line in anaconda. I don't remember what I did for tensorflow, but I have it up and running and I am quite the computer phillistine so it could not have been much harder.
Maybe the second part is more of a value prop or good point of focus?
- Do you actually need to raise? If so, is now the right time? If you don’t have a product with any users, it may be more valuable to prioritize this. Raising money is a huge distraction. Approach with caution.
- A clear, succinct, well-designed deck does make a difference.
- Talk to your users. How do they use your product? How often? Include this in your deck.
- Venture Deals and Mastering the VC Game are helpful books
- Read early slide decks of now successful companies. Many are available online (e.g. Airbnb).
- Warm intros help. If you know someone who knows a VC and can intro you, ask!
- Different VCs have different investment strategies. Your TAM might not move the needle on a 1B fund, but it could on a 20M one.
- The best story is a growth curve that’s up and to the right.
- For later stage: not to sound demeaning, but VCs often act like lemmings. An offer on the table makes rallying others easier — reach out to those who gave you a “VC pass” (ie nevder responded to your email, or didn’t follow up after a meeting) and see if they’re interested now.
- If possible, get feedback on your deck from someone who has successfully raised.
- Don’t tell VCs which other VCs you’re talking to. You’ll be tempted, but don’t.
- Take notes after each meeting. What were the objections? Stumbling points? Use this feedback to improve your deck.
- Giving a range for your valuation or amount you want to raise makes you appear indecisive and lacking in confidence. Give specific figures.
- Be capable of justifying why you want to raise X. How’d you come to this figure?
- Stories help. How’d you come to this idea? If you have direct exposure to the problem you’re trying to solve - especially if it’s a business problem - incorporate this into your pitch.
- Make sure you’re talking to people who can make a decision within the firm.
- Don’t copy and paste cold emails. Personalize them.
- This can be a discouraging process. But it’s a numbers game. You only need one yes to get the ball rolling.
MSRP may be $999, but you won’t find them for much below $1,200 for the time being.
GPU modules are manufactured in China. Their harmonized codes are covered in recently established tariffs. 10% tariffs are already hitting cards arriving at US ports. This tariff will increase to 25% on Jan 1.
Yes. 1080 Ti do not scale linearly, but in most of the benchmarks I’ve seen, two 1080 Ti are at least 1.75x faster than one when doing multi-GPU training.
1.75 / 1.36 (speed up of 2080 Ti over a single 1080 Ti) = 1.28. So expect 2x 1080 Ti to be about 30% faster.
You can see how multi-GPU training works with Titan V benchmarks in the link below. 1080 Ti have similar scaling profile.
Also, that’s good feedback on the CPUs bottlenecking. I’ll let our HPC hardware team know about this.
We are also looking into GPU direct storage to help resolve this:
https://developer.nvidia.com/blog/gpudirect-storage/