The new GPU-optimized family of instances (P3) are powered by Tesla V100s. More details in this blogpost.
Interesting to note that they have skipped the Pascal family of GPUs , incidentally Pascal P5000s and P6000s are offered by Paperspace.
Edit 1 - Stephen Merity has just published his first thoughts on P3 here. Considering his analysis is for AWD-LSTM and QRNN in PyTorch, I would say these are quite relevant from the course perspective.
Edit 2 - Smerity’s Tweet
TL;DR :- P3 instances are available in three instance sizes,
p3.2xlarge with 1 GPU
p3.8xlarge with 4 GPUs
p3.16xlarge with 8 GPUs
They are available in US East (N. Virginia), US West (Oregon), EU West (Ireland) and Asia Pacific (Tokyo) regions. Customers can purchase P3 instances as On-Demand Instances, Reserved Instances, Spot Instances, and Dedicated Hosts.
P.S : Hope they don’t ask the user to raise a service limit request. Haven’t found any pricing details. 
On demand prices are available here. Just make sure you select the available regions. It costs $3.06 per hour. Spot prices are available here. Right now going at $0.3391 per hour.
The p3.16x large instances with 8 GPUs are already under demand. 244$ at its peak.
Any hunch who might be using these? Large industrial research labs?
Sadly, no PyTorch support yet.
hmm, for P3 isntances you need CUDA9 and CUdnn 7, but looks like people making it work:
I haven’t tried it yet…
These drivers and libraries have already been added to the newest versions of the Windows AMIs and will be included in an updated Amazon Linux AMI that is scheduled for release on November 7t’h. (regd. CUDA9 and CUdnn 7)
I guess some teams in Microsoft 
Nvidia has released docker containers with Pytorch support for P3 instances.
These were the other part of the ‘treats’ I talked about - I knew in advance these were coming
You can also just conda install pytorch.
Are people finding they need to make a service limit request to get access? Or are you finding you already have access.
Looks like we have direct access. I could just spin up an instance without raising any request.
Does this mean we can use the P3 instances without the pytorch support for CUDA9 / CuDNN7 support. I’m assuming that we should be able to run an AMI optimised for CUDA8 (with pytorch) and still get some V100 benefits. But, I haven’t messed around with GPUs and drivers a lot. Can you clarify this a bit ?
You should be fine to pip or conda install Pytorch on the new instances. I’ll be trying it out today.
Sweet. Let us know how this goes, and if there are speedups without the CUDA9/CuDNN7 drivers. (Not exactly sure how/what to run to test those yet.)
Actually, on CUDA8 drivers it just blindly does not see V100 GPUs on P3 instances.
With CUDA9 I am getting PyTorch compilation errors. But I suspect I messed installation somewhere.
Still trying to make it work though…
@jeremy - I was able to launch p3.2xlarge instance earlier but I am not able to do that from yesterday. I have raised service request to get access. I am assuming this might be true for other students also. If possible, you can contact your amazon contact to get access for the amazon accounts shared in the sheet.
I have requested it, quite a few times. Never quite got a clear answer however… 
We can ask all students to check if they have access if not, raise a request for it. I am not sure how long it will take for AWS to approve it. I am yet to hear from amazon about the request I raised.
I have access to the p3 instance, but I’ve had an AWS account since the beginning of the year, if that helps
I opened a case with AWS three days ago but I have no access to a p3 instance yet. Maybe there is a shortage, at least in the AWS Ireland zone (I am based in Europe). I assume p2 instances are an excellent choice for the course anyway.
I have opened case to access p3 and p2 instances. I don’t have access to any of gpu instance (created the aws account for the course only)
