Nvidia unveils new GPUs for data centres

[Jamie McKane]

Nvidia unveils new GPUs for data centres

Nvidia Corp. announced new graphics chips and computers for artificial intelligence processing in data centres, claiming a leap forward in performance that may help cement its lead in the growing area.

The company’s Ampere chip design is 20 times faster than its predecessors and has the flexibility to be repurposed between the two key areas of artificial intelligence processing: training and inference, Chief Executive Officer Jensen Huang said.

 

[cguy]

... but can it play Sopwith?

 

[cguy]

20x faster is a bit of a stretch. 1.25-2.5x for most applications, 10x for some edge cases and 20x for some edge cases of edge cases.

 

[clynton]

Just because you're cribbing off of Bloomberg is no excuse for not proofreading

 

[magicfoot]

GPUs are cool for numerical processing such as ML training routines, but what tasks are they doing in a data centre? GPUs do not run generalized software such as MS SQL, Apache, or IIS, so what is their function in a commercial data center.

I do like Nvidia GPUs but this article sounds like hokum!

 

[Messugga]

GPUs are cool for numerical processing such as ML training routines, but what tasks are they doing in a data centre? GPUs do not run generalized software such as MS SQL, Apache, or IIS, so what is their function in a commercial data center.

I do like Nvidia GPUs but this article sounds like hokum!

This may be hard to grasp, but all data centers aren't designed purely for database and web hosting purposes...

 

[cguy]

GPUs are cool for numerical processing such as ML training routines, but what tasks are they doing in a data centre? GPUs do not run generalized software such as MS SQL, Apache, or IIS, so what is their function in a commercial data center.

I do like Nvidia GPUs but this article sounds like hokum!

Off the top of my head:
- ML training and inference happens at large scale in data centres
- Supercomputers are hosted in data centres
- Large scale scientific and engineering based computation/simulation is done in data centres (some overlap with above)
- Large scale 3D image/video rendering happens in data centres
- Quantitative computing for finance happens in data centres

 

[Johnatan56]

Off the top of my head:
- ML training and inference happens at large scale in data centres
- Supercomputers are hosted in data centres
- Large scale scientific and engineering based computation/simulation is done in data centres (some overlap with above)
- Large scale 3D image/video rendering happens in data centres
- Quantitative computing for finance happens in data centres

Just adding:
AWS: https://aws.amazon.com/nvidia/
MS: https://docs.microsoft.com/en-us/azure/virtual-machines/sizes-gpu
Google/Nvidia: https://www.nvidia.com/en-us/data-center/gpu-cloud-computing/google-cloud-platform/
From the one right above:


It's substantially faster to go and throw a learning task on a cloud instance, pay for a ton of GPU power for a few minutes/hours, than it is to buy your own stuff and let it run for days as you can't scale it.

Only reason I didn't go cloud route for uni stuff is that I also wanted another machine to game on, and my stuff is small enough that it only takes a day or two. For a larger business, having your data scientist wait days for a result is a lot of wasted money, especially as that result could be needed right now.

For those doing hobby stuff, this I've found can often be cheaper than AWS, so it's quite nice: https://vast.ai/
Wouldn't use for business as data encryption.

 

[UrBaN963]

You guys realise this is all just so the DB Managers can play Crysis on their servers right?

 

[magicfoot]

Just adding:
AWS: https://aws.amazon.com/nvidia/
MS: https://docs.microsoft.com/en-us/azure/virtual-machines/sizes-gpu
Google/Nvidia: https://www.nvidia.com/en-us/data-center/gpu-cloud-computing/google-cloud-platform/
From the one right above:
View attachment 840715
View attachment 840717
It's substantially faster to go and throw a learning task on a cloud instance, pay for a ton of GPU power for a few minutes/hours, than it is to buy your own stuff and let it run for days as you can't scale it.

Only reason I didn't go cloud route for uni stuff is that I also wanted another machine to game on, and my stuff is small enough that it only takes a day or two. For a larger business, having your data scientist wait days for a result is a lot of wasted money, especially as that result could be needed right now.

For those doing hobby stuff, this I've found can often be cheaper than AWS, so it's quite nice: https://vast.ai/
Wouldn't use for business as data encryption.

I am lamenting the fact that for the last five years I have been working with GPU facilities at EC2, Google, CHPC, Juelich, etc, but these are all projects for research and scientifically funded institutions only, not commercial. The only game in town for the use of GPUs at commercial organizations seems be some form of ML, and for this the commercials would probably use EC2 or other cloud variant as opposed to buying their own hardware. Can anyone name a large local commercial organization using ML with GPUs in-house, or any other large scale GPU profit-making application for that. I would like to believe that the Banks, P&Ps, Woolies, MTN, etc, are all diving into large scale GPU usage, but are they really? Can anyone cite the brand name using large commercial GPU deployments in SA? Iris petal categorization or Titanic disaster analysis need not be included

 

[Johnatan56]

I am lamenting the fact that for the last five years I have been working with GPU facilities at EC2, Google, CHPC, Juelich, etc, but these are all projects for research and scientifically funded institutions only, not commercial. The only game in town for the use of GPUs at commercial organizations seems be some form of ML, and for this the commercials would probably use EC2 or other cloud variant as opposed to buying their own hardware. Can anyone name a large local commercial organization using ML with GPUs in-house, or any other large scale GPU profit-making application for that. I would like to believe that the Banks, P&Ps, Woolies, MTN, etc, are all diving into large scale GPU usage, but are they really? Can anyone cite the brand name using large commercial GPU deployments in SA? Iris petal categorization or Titanic disaster analysis need not be included

Why would you do that in SA? ML is not latency critical so no reason to host in SA where a data center would cost a premium.
And buying hardware yourself is generally not a good investment either since you overcompensate so you can get it done faster and then it sits idle, or you under-compensate and then it takes longer/engineer idles, I can't think of any modern organization that would want to do so outside of maybe some very large video houses, and those large video houses would want to be able to scale in burst, since it would be for a movie or something, so you'd need organize the hardware, etc. which is outside of what your business is (so acquire, manage and sell off everything again).

Maybe a bank would want to do so since local/in-control, but they'd probs rather do it via AWS/MS and do the security checks with them, it would take a long time for them to migrate.

 

[Messugga]

I am lamenting the fact that for the last five years I have been working with GPU facilities at EC2, Google, CHPC, Juelich, etc, but these are all projects for research and scientifically funded institutions only, not commercial. The only game in town for the use of GPUs at commercial organizations seems be some form of ML, and for this the commercials would probably use EC2 or other cloud variant as opposed to buying their own hardware. Can anyone name a large local commercial organization using ML with GPUs in-house, or any other large scale GPU profit-making application for that. I would like to believe that the Banks, P&Ps, Woolies, MTN, etc, are all diving into large scale GPU usage, but are they really? Can anyone cite the brand name using large commercial GPU deployments in SA? Iris petal categorization or Titanic disaster analysis need not be included

ABSA for one. Std Bank as well.
My employer also plays. We have quants with R500+k workstations on their desks, nevernind what's in the data centers. Lots of pricing models and whatnot, can make use of GPU compute.
Engineering firms also heavily use this stuff for their modelling.

 

[Lupus]

GPUs are cool for numerical processing such as ML training routines, but what tasks are they doing in a data centre? GPUs do not run generalized software such as MS SQL, Apache, or IIS, so what is their function in a commercial data center.

I do like Nvidia GPUs but this article sounds like hokum!

NvENC for transcoding.

 

[newby_investor]

The other cool thing coming to this generation is the integrated Mellanox 100Gbit NICs. So your data can come in straight from Ethernet or Infiniband, without needing to go through host memory.

 

[cguy]

I am lamenting the fact that for the last five years I have been working with GPU facilities at EC2, Google, CHPC, Juelich, etc, but these are all projects for research and scientifically funded institutions only, not commercial. The only game in town for the use of GPUs at commercial organizations seems be some form of ML, and for this the commercials would probably use EC2 or other cloud variant as opposed to buying their own hardware. Can anyone name a large local commercial organization using ML with GPUs in-house, or any other large scale GPU profit-making application for that. I would like to believe that the Banks, P&Ps, Woolies, MTN, etc, are all diving into large scale GPU usage, but are they really? Can anyone cite the brand name using large commercial GPU deployments in SA? Iris petal categorization or Titanic disaster analysis need not be included

I worked for an engineering firm in SA that uses GPUs extensively. I work in finance in the US now and we use GPUs extensively too.