Slashdot Mirror

← Back to Stories (view on slashdot.org)

Google Cloud Platform Cuts the Price of GPUs By Up To 36 Percent (techcrunch.com)

Posted by BeauHD on from the new-and-improved dept.

In a blog post, Google's Product Manager, Chris Kleban, announced that the company is cutting the price of using Nvidia's Tesla GPUs through its Compute Engine by up to 36 percent. The older K80 GPUs will now cost $0.45 per hour while the more powerful P100 machines will cost $1.46 per minute (all with per-second billing). TechCrunch reports: The company is also dropping the prices for preemptible local SSDs by almost 40 percent. "Preemptible local SSDs" refers to local SSDs attached to Google's preemptible VMs. You can't attach GPUs to preemptible instances, though, so this is a nice little bonus announcement -- but it isn't going to directly benefit GPU users. As for the new GPU pricing, it's clear that Google is aiming this feature at developers who want to run their own machine learning workloads on its cloud, though there also are a number of other applications -- including physical simulations and molecular modeling -- that greatly benefit from the hundreds of cores that are now available on these GPUs. The P100, which is officially still in beta on the Google Cloud Platform, features 3594 cores, for example. Developers can attach up to four P100 and eight K80 dies to each instance. Like regular VMs, GPU users will also receive sustained-use discounts, though most users probably don't keep their GPUs running for a full month.

28 comments

  1. Look out Nvidia by Anonymous Coward · · Score: 0

    Teh G gonna eat your meat!

    1. Re:Look out Nvidia by Anonymous Coward · · Score: 0

      This is all well and good, as I use it quite heavily. My only
      problem is the huge network lag when driving my triple 4k
      monitor setup. It can take a couple of minutes to scroll a few
      lines in vim (which is why I needed 4k monitors). I guess I
      pine for the olde days when the GPU was local to the CPU.
      I guess this is progress...

      CAP === 'corrects'

  2. Bitcoin mining? by invictusvoyd · · Score: 2

    Would it be profitable?

    1. Re:Bitcoin mining? by Anonymous Coward · · Score: 0

      No.

    2. Re:Bitcoin mining? by Anonymous Coward · · Score: 1

      NO, not even close. GPU mining on bitcoin hasn't been profitable for years.

    3. Re:Bitcoin mining? by Anonymous Coward · · Score: 0

      only if you pay with a stolen CC

    4. Re:Bitcoin mining? by Anonymous Coward · · Score: 0

      Not even then as there are far better ways to turn the credit card to cash rather than give it to google for a fraction of a cent on the dollar.

    5. Re:Bitcoin mining? by Anonymous Coward · · Score: 0

      I already bought six!

    6. Re:Bitcoin mining? by Anonymous Coward · · Score: 0

      Market rate to rent GPUs for crypto mining is in the region of $0.10 per Geforce1080 equivalent GPU per hour. Broadly, the Geforce 1080 offers roughly 3x the mining capability of something like a K80.

      My supplier is currently offering a block of mining power equivalent to 5k Geforce 1080 cars at $0.09 per GPU per hour.

      I took advantage of this offer over the weekend, and made a reasonable profit (about 30% return on investment), but it was not a click and forget, and it was only viable because I was mining a new cryptocurrency with an insanely volatile market. You need to keep a close eye on the mining too, as mining servers are frequently fraudulent, so you need to have monitoring scripts calculating statistics, and have backup servers ready to go, if the p-values on your returns start dropping.

    7. Re:Bitcoin mining? by Doc+Right · · Score: 0

      You have 1300 altcoins to choose from.

  3. How long will it take by Anonymous Coward · · Score: 4, Funny

    How long will it take before someone thinks "hey, you know what would be great? If you could own a VM. Like a cloud computer, but for you, personally. A personal computer, so to speak."

    1. Re: How long will it take by Anonymous Coward · · Score: 0

      Like a vps? Um...

  4. Quality editor work by Anonymous Coward · · Score: 3, Funny

    Boss: Could you edit this a bit so it doesn't look like a direct copypaste?
    Editor: Sure thing, I'll change "hours" to "minutes" on this bit.

    1. Re: Quality editor work by Anonymous Coward · · Score: 1

      Source said minutes. Source corrected. copy pasta detected

  5. GPU coin mining by Anonymous Coward · · Score: 0

    For LTC or ETC maybe, not for BTC

    1. Re:GPU coin mining by Anonymous Coward · · Score: 0

      Not at that price it isn't. It isn't even close to being viable/profitable.

  6. K80 P100 by wjcofkc · · Score: 2

    There was such thing as a K80, but nostalgia ensues none-the-less. I'm going to have to dig up some old hardware install OS\2 Warp now, all for the joy of setting jumpers.

    --
    Brought to you by Carl's Junior.
  7. Too expensive by 110010001000 · · Score: 1, Funny

    My Voodoo2 card outperforms all of those options, plus I can play Quake on 800x600 with ultra high settings on it.

    1. Re:Too expensive by DontBeAMoran · · Score: 1

      Your Voodoo2 is no match for the Bitchin' Fast! 3D 2000.

      --
      #DeleteFacebook
    2. Re:Too expensive by hawk · · Score: 1

      Which itself is no match for the Radeon 9500 ASC http://www.bbspot.com/News/2003/02/ati_ascii.html, enhanced for high-power ASCII gaming . . .

      hawk

  8. Sustained Usage Discount by dave562 · · Score: 2

    Like regular VMs, GPU users will also receive sustained-use discounts, though most users probably don't keep their GPUs running for a full month.

    Challenge accepted!

    Hold my beer...

    1. Re:Sustained Usage Discount by DontBeAMoran · · Score: 1

      Simply mine Litecoin and/or Monero. That's not much of a challenge.

      --
      #DeleteFacebook
  9. Cloud pricing complexity by swb · · Score: 1

    Is cloud pricing complex because it has to be -- so many options and permutations of configuration that it addresses a market need for flexibility?

    Or is it done because of economic viability? In other words, if there was just some standard monthly rate for a VM regardless of utilization the price would end up being so high that it would be non-competitive? I accept that even in this situation there would be some variability -- maybe CPU count, memory quantity, disk consumption, but more or less flat rate for those quantities.

    Or is it a case of deliberately opaque pricing -- make it seem cheap ("only $0.75 per minute, it's like free!") and very difficult to calculate what it would actually cost to run workloads on it? The enterprise sales gimmick of never making it possible to make

    I'm sure its a combination of all 3, but I'm inclined to think that there's a strong combination of lack of economic viability and deliberate opacity.

    1. Re:Cloud pricing complexity by Anonymous Coward · · Score: 0

      My experience is that cloud pricing is only really complex if you consider lots of options complex. If you don't want to account for demand etc you can buy a dedicated resource and pay accordingly, however the people paying a notable amount for cloud resources are more likely to want to optomise their investment than many consumers buying services they don't understand in detail (bandwidth use for broadband for example). I've not used a huge amount of cloud based services, but even for me the options have been useful. I've moved a number of AWS instances between standard instances to t2 instances (more capable than standard, but earn/burn credits so you can't run them full on constantly) to increase capacity on some services by ~3x without increasing cost, some of them then moved back to standard (actually compute heavy) instances when our set up was tidied up to allow instances to be added and removed as load required. I've seen other uses where non-time critical processing is done when load is low so the rate is discounted (it isn't discounted if you have it reserved already).

      All the above are example of where multiple options allows higher cost effectiveness and so a service just offering a single cost all you can eat price would either have to hope that it got a disproportionate number of underutilized services or more likely charge so much that only people absolutely maximising utilisation would be willing to pay it.

    2. Re:Cloud pricing complexity by coofercat · · Score: 1

      In my experience, Google makes the pricing more opaque than it needs to. I have experience of a couple of other providers, who charge for the resource you've been allocated, regardless of how you use it (we just use VMs, no fancy-schmancy stuff). They may have some sort of fair-use clauses that mean they'll throttle your VM if you burn too much CPU for too long - I'm not sure. Google charge for idle, and then for how much you use it, and then give you a committed use discount. It's really hard to know anything with that much complexity.

      Also, Google do a really bad job of reporting what you're costing. It's hard to see what any given VM adds to the total, and hard to predict what a change will do the end-of-month bill. Their reporting can't show you "last 30 days", it only shows you "since last billing date", so you can't even get a rough idea of trends in your usage/price. I asked them about this and they said I should dump all the billing data into one of their database/big data tools and then slice it and dice it as I want. Seems a bit much to ask me to pay for more services to find out how much the existing ones are costing me.

      There are also a load of weird restrictions on things you wouldn't really expect. For example, if you want network packet forwarding in your VM, you have to create your VM from scratch to do it - you can't enable it retrospectively. Another one is if you put a local SSD on your VM, you can't reboot your VM. Yep, once you connect one of those, if you type 'reboot', your VM will go down, but not come back up (with no option for backup/recovery). You have to destroy it and start again. I'm not quite sure what the intended use-case is there. Oh, and changing the number of CPUs, or amount of RAM, or pretty much anything about the system requires a reboot - almost nothing happens dynamically. Better detatch that SSD before you reboot!

      Then, Google insist on messing with /etc/hosts, the system hostname, /etc/resolv.conf, and a bunch of other config files on your system. If you set them the way you want them, Google will put them back at some future time of their choosing, so you're not so much renting a VM as you're running your apps in someone else's computer.

      One thing Google has going for it is that it's cheap. The SSDs are cheaper than other people's standard disks (for example), and putting a couple of extra CPUs into a VM is pretty affordable. They also gave us something like $350 signing credit, which means we didn't have to pay two providers while we migrated from our old provider to them. The other thing they have is their name - you tell anyone "oh yeah, we're hosted at Google", then they immediately feel a bit safer than if you say "yeah, we're at rockbottompricehosting.com" - just as well, because they'd probably fail to evaluate as well as rockbottomhosting.com for a lot of their technical stuff. I also like their firewalling options, although have found sharing networks between two firewall zones a painful experience.

      So yeah, it's more complex than it needs to be. It gets even more complex when you have more complex requirements than "just give me some VMs".

  10. Oooh, this is so exciting by grungeman · · Score: 0

    I think I wet my pants!

    --

    Signature deleted by lameness filter.
  11. The P100 machines cost $1.46 per *hour* not minute by Anonymous Coward · · Score: 0

    The older K80 GPUs will now cost $0.45 per hour while the more powerful P100 machines will cost $1.46 per minute

    You have the units for the P100 wrong.

  12. Bounds checking rates? by augustz · · Score: 1

    How about a bounds check on some of these articles?

    $1.46 per minute is $767,000 per year.

    That rate makes no sense, no matter what the mark-up is to be in the cloud. Does no one read these before they go live?