Google Compute Engine now has machine types with 96 vCPUs and 624GB of memory

[Guest]

Just yesterday (or 10/5/17 if you're a latecomer), Google announced that it is now allowing VM configurations with up to 96 vCPUs and 624GB of memory for its own Compute Engine machine types on their cloud platform.

Quote

Got compute- and memory-hungry applications? We’ve got you covered, with new machine types that have up to 96 vCPUs and 624 GB of memory—a 50% increase in compute resources per Google Compute Engine VM. These machine types run on Intel Xeon Scalable processors (codenamed Skylake), and offer the most vCPUs of any cloud provider on that chipset. Skylake in turn provides up to 20% faster compute performance, 82% faster HPC performance, and almost 2X the memory bandwidth compared with the previous generation Xeon.¹

4

This is a pretty sizable increase over what they offered in March, with VM configs of 64 cores and 465GB of ram. (3)

Quote

The new 624GB Skylake instances are certified for SAP HANA scale-up deployments. And if you want to run even larger HANA analytical workloads, scale-out configurations of up to 9.75TB of memory with 16 n1-highmem-96 nodes are also now certified for data warehouses running BW4/HANA.

1

And if the previous options are still not enough and you happen to have workloads compatible with what is mentioned above, you now have even greater options to use your compute budget.

Quote

You can use these new 96-core machines in beta today in four GCP regions: Central US, West US, West Europe, and East Asia. To get started, visit your GCP Console and create a new instance. Or check out our docs for instructions on creating new virtual machines using the gcloud command line tool. Need even more compute power or memory? We’re also working on a range of new, even larger VMs, with up to 4TB of memory. Tell us about your workloads and join our early testing group for new machine types.

Well, you have a beta to try it out and there are even greater offerings coming someday, somewhere...

 

Update 1:

Google Cloud Compute will allow you to sign up for a one year trial with $300 of credit (Annual not monthly) if you want to play around try it out. The downside is that during the trial, you can only make VMs with up to 8 cores and 52GB of ram. This configuration is estimated at $345.73/month if that gives you a better indicator of how much the full fat 96 core will cost.

 

What are your thoughts on this and will you end up using it?

Sources:

(Google) (Tech Crunch 1) (Tech Crunch 2)

This is my first tech news and reviews topic so if there are mistakes, please do say so. This post is also compliant with night theme.

[Cyracus]

The only things that matter, how much it costs, and how good is it at mining 

[Guest]

Just now, Cyracus said:

The only things that matter, how much it costs, and how good is it at mining 

Doubt it would even be remotely profitable at mining.

[Guest]

Update 1:

Google Cloud Compute will allow you to sign up for a one year trial with $300 of credit (Annual not monthly) if you want to play around try it out. The downside is that during the trial, you can only make VMs with up to 8 cores and 52GB of ram. This configuration is estimated at $345.73/month if that gives you a better indicator of how much the full fat 96 core will cost.

[ScratchCat]

But can it run minesweeper

This must be really useful for researchers who need to process large datasets infrequently, For any business it would be much cheaper simply to buy the hardware.

Naples should however decrease the prices of these instances dramatically while offering a greater per dual socket machine memory density.

[Guest]

6 hours ago, ScratchCat said:

Naples should however decrease the prices of these instances dramatically while offering a greater per dual socket machine memory density.

Heck, the core selector on their site still allows you to select what generation the cores are so I could see them adding an option for it as long as there are no issues involved with doing so. Things I could think of is an exclusive agreement between Intel and Google, or they made some very heavy optimizations for Intel's uarch. That or they want to milk the researchers of their money and not speed any additional money to set up a platform that would save their consumers money

[cj09beira]

1 hour ago, tjcater said:

Heck, the core selector on their site still allows you to select what generation the cores are so I could see them adding an option for it as long as there are no issues involved with doing so. Things I could think of is an exclusive agreement between Intel and Google, or they made some very heavy optimizations for Intel's uarch. That or they want to milk the researchers of their money and not speed any additional money to set up a platform that would save their consumers money

you can bet intel is doing everything they can to limit naples sales, shady contracts are probably already in place 

[ScratchCat]

4 hours ago, tjcater said:

Heck, the core selector on their site still allows you to select what generation the cores are so I could see them adding an option for it as long as there are no issues involved with doing so. Things I could think of is an exclusive agreement between Intel and Google, or they made some very heavy optimizations for Intel's uarch. That or they want to milk the researchers of their money and not speed any additional money to set up a platform that would save their consumers money

Probably the availability of AVX and other components means that  scientific systems can be processed much faster compared to a desktop use case. AVX512 doubled performance in some applications. Additionally they need to deploy servers globally and validate any configuration on a new architecture.

 

[LAwLz]

13 hours ago, ScratchCat said:

For any business it would be much cheaper simply to buy the hardware.

Not necessarily.

This and other services like Azure and AWS got lots of benefits. The biggest ones I can think of are Flexibility and the lack of maintenance necessary.

You might not always want the same level of performance. Just going "OK we need more cores and memory for the next month and then we are back to our regular plan" is worth a lot. If you buy hardware you don't get the same flexibility.

You also don't need to worry about taking care of the hardware. You just rent it and it "just works". The same can not be said for buying and doing upkeep on multiple server racks.

[ScratchCat]

15 minutes ago, LAwLz said:

Not necessarily.

This and other services like Azure and AWS got lots of benefits. The biggest ones I can think of are Flexibility and the lack of maintenance necessary.

You might not always want the same level of performance. Just going "OK we need more cores and memory for the next month and then we are back to our regular plan" is worth a lot. If you buy hardware you don't get the same flexibility.

You also don't need to worry about taking care of the hardware. You just rent it and it "just works". The same can not be said for buying and doing upkeep on multiple server racks.

Generally businesses require the same amount of compute power each month, however the cost and reliability points are valid.

[mrthuvi]

14 hours ago, tjcater said:

Heck, the core selector on their site still allows you to select what generation the cores are so I could see them adding an option for it as long as there are no issues involved with doing so. Things I could think of is an exclusive agreement between Intel and Google, or they made some very heavy optimizations for Intel's uarch. That or they want to milk the researchers of their money and not speed any additional money to set up a platform that would save their consumers money

 

I have read somewhere the opposite. Intel is the one making optimization for the big 7 (MS, Facebook, Google, Baidu, Amazon, Alibaba, Tencent)'s code/workload in the Chips.

[Guest]

7 hours ago, mrthuvi said:

 

I have read somewhere the opposite. Intel is the one making optimization for the big 7 (MS, Facebook, Google, Baidu, Amazon, Alibaba, Tencent)'s code/workload in the Chips.

That has been on going for a while but doesn't change that the big 7 would still be optimizing their code for these processors.