I was also surprised (and sent it around to our team internally last night). We're skipping Maxwell entirely as you can see from my previous comment threads.

For display it's still a fine part. The P100 is also a beast, so its overkill for most people just doing Remote Desktop. So perhaps the M60 (like with Azure) fills this market segment for them, and they don't mind the hardware diversity.

[Edit: Too sleepy. A post down below reminds us that these are G-series and G is for Graphics. So yeah, I assume they just didn't want to wait for enough P4 parts in volume or will quickly make another such announcement about the Pascals].

Disclosure: I work on Google Cloud.

There was an announcement last year that AMD FirePro™ S9300 x2 were coming to Google cloud. Any reason why those have not hit general release yet? I have some OpenCL workloads that I could put to good use on those.

Still on it, and now coming really soon (hopefully). There were a lot of interesting challenges (power, software, datacenter process) along the way, since this is very new hardware to Google.

Out of curiosity, why OpenCL on AMD rather than OpenCL on say a K80 or AWS's M60s or upcoming P100s? (That is, are you blocked on having an AMD part, or do you have a great reason to prefer AMD for compute)

Disclosure: I work on Google Cloud (and want to sell you GPUs!)

Because the code is already validated / tuned and deployed on AMD parts. For a while NVidia dragged their feet on OpenCL so validating the app on their hardware was not possible.

Great feedback, thanks.

FWIW, we haven't invested in opencl for anything due to the fragmentation. It's great to see google pushing an alternative.

Overall though, I would basically consider opencl "AMD's thing".

Not original poster, but for us is the lack of OpenCL2.0 on nVIDIA HW (mainly shared data structures) Only OCL 1.2 on Linux from nVIDIA

I think "generation" here is referring to EC2 generations and not GPU generations – AWS tends to use that term to refer to new instance types being released.

((Next Generation) (GPU EC2 Instances)) rather than ((Next Generation GPU) (EC2 Instances)) :)

"I am in the market for a cloud GPU offering, and I have to say the big cloud providers are very uncompetitive here, only offering these old, slow GPUs. "

It's one thing if one of them is like that, but if all of them are like that, maybe it's not because of the cloud providers?

Oh, there are smaller vendors renting out Pascal architecture GPUs. Yet Amazon and Azure are stuck at prev-gen (M60), while Google is all the way back at prev-prev-gen (K80).

Would you mind specifying these smaller vendors and a few details about your experiences?

I don't want to promote anyone in particular, but there are a few vendors here, for example:

https://www.ovh.ie/dedicated_servers/gpu/ https://www.exoscale.ch/gpu/ http://www.cirrascale.com/cloud/plans.aspx http://www.nvidia.com/object/gpu-cloud-computing.html

Last link has a list of vendors. The one I used gave you a bare bones machine running some Linux distro, with no platform bullshit around it. Exactly what you need for deep learning work.

Or GTX 1080 ti. Aren't the Tesla class like the P100 mostly super overpriced for deep learning because their only main advantage is Double (64 bit) float support, and no one really needs that? Plus half float (16 bit) support, which is not super widely used (but certainly more than double). Something like 95%+ of Deep Learning must be done with single floats (32 bits) right now afaik, making this a fairly dubious expense

> Aren't the Tesla class like the P100 mostly super overpriced for deep learning

Yes. Nvidia has real competition from AMD in consumer graphics but no competition in enterprise. Their consumer cards are sold at a competitive price while their enterprise cards are marked way up.

If cloud providers were really smart, they'd release cheaper/just-as-fast instances with consumer cards, but Nvidia probably doesn't want to do that deal; they'd prefer to push teslas as the standard for GPGPU, even if they have to mark them down for cloud providers.

As DannyBee said in a sibling comment:

> It's one thing if one of them is like that, but if all of them are like that, maybe it's not because of the cloud providers?

What is your point?

There may be non-technical reasons beyond "Those guys are just stupid" that we don't offer GTX 1080s.

Why don't you offer GTX 1080s?

I'm suggesting Nvidia is offering great deals on teslas to cloud providers but not offering those same discounts on 1080s; they're doing that because it lends support to their pitch to enterprises that they need Teslas instead of their competitively priced consumer line; if everyone was paying list price, cloud providers would be offering 1080s.

Even with all that, cloud GPU is not cost effective and end users are better off buying their own "consumer grade" stuff.

Looks to be about twice the cost of the M60 on AWS, but you should get quite a bit more performance out of that system.

The g2s were Kepler, the g3s are Maxwell. 'Next generation' is technically correct.

Current gen is Pascal though, so 'Prev generation' is more correct.

In architectural terms perhaps. But the M60 is most definitely a current generation Tesla product.

These are G-instances and graphics is one of the primary use cases for this instance (and for the previous G2 instance). The M60 is the top part with GRID support built in. A GRID license that allows you to use the GRID driver is part of the offering.

Disclosure: I own HPC for AWS (among other things) and used to own instances

Not really. Typically there were reasons you would stick with older generations, such as M60s having very poor fp64 performance. If you want fp32, it gives a bump over the K80s currently in there. However, the P100 would have been the most logical path since it doesn't cripple fp16/32/64.

His point is:

They didn't say newest generation, only next and next may in this case refer to

- "next after kepler" (parent's assumptions) or

- "next after current" (other commenters' assumption)