What I Learned at NVIDIA GTC 2019 (San Jose)

First of all, it is always humbling to be dumbest person in the room.  PHDs everywhere (some people even with multiple PHDs)!  I think everyone should feel this once in your life because it really helps get you motivated to get back to learning new stuff on a daily basis (and hopefully not be the dumbest person in the room next year).  If you are not following @NVIDIAVirt or @NVIDIA on Twitter, you really should.  NVIDIA is always out there educating the community on what people are using GPUs to do.

Secondly, every single time I see AI (for Artificial Intelligence) in a presentation… I automatically think my first name.  Al (that is an L).  Maybe it’s just me.  I guess I should start putting my name in capital letters from now on to make it easier.

Anyway, so while the focus of the conference is everything is bigger, better, faster and cheaper than last year, they also were showing use cases: from autonomous driving to video production/ray tracing to machine learning/neural networks to healthcare.  They also discussed NVIDIA’s acquisition of Mellanox.

For the purposes of this blog, we will be focusing on more of the “enterprise” aspect of conference rather the other applications.  Some of the announcements were prior to NVIDIA GTC but were still discussed at the conference.

NVIDIA RTX Server and Quadro Virtual Data Center Workstation (Quadro vDWS)

Part of the keynote was to announce the NVIDIA RTX™ Server which “delivers a highly configurable reference design for on-demand rendering and virtual workstation solutions in the data center”.  The major manufactures that have created these reference designs are HP Enterprise, Lenovo, Dell EMC and Super Micro. By adding the NVIDIA Quadro vDWS, an organization can enable remoting of the secure pixels to address your use case (AKA remote display to a physical server than to a virtual machine):

• Geographically dispersed workforce
• Frequent or In-frequent third party contributors
• Large datasets with many collaborators
• High security datasets
• Certification for mission-critical software

Supported protocols:

Teradici Cloud Access Software with PCoIP protocol support for Quadro vDWS GPU acceleration for RTX Server will be in technical preview starting April 15.

Opinion: I understand PCoIP being the primary protocol, but  I would like to see what other protocols that clients have already like Citrix HDX and VMware Blast Extreme.

NVIDIA Tesla T4 vs Tesla M10 for vGPU

The M10 has been the density card of choice since it was released with some server vendors supporting 2 or 3 cards per server.  While the promise of the 512MB profile sounded great, the reality was less than stellar due to limitations with NVENC issues (affecting both Citrix and VMware : here and here) in the 512MB profile. So that resulted in a single M10 card supporting 32 1GB users and most servers only supporting 2 M10s.

With the new T4s NVIDIA is hoping to deliver a high performance/high density solutions.  Here are some of the T4 advantages over the M10 that you might want to look at:

1. (Possibly) Better Density – The 16 GB of GPU GDDR6 memory means 16 x 1GB profiles per card but since you get upwards in 6 T4s in a common server (with some servers supporting upwards of 10 T4s).  This will of course depend on your server vendor and model you choose so make sure to check the NVIDIA Certified Servers HCL for your current vendor/model or before you make a purchase.
2. “VDI by Day, Compute by Night” –   The GTC session by NVIDIA Solution Architects Konstantin Cvetanov,  Shailesh Deshmukh and Eric Kana “S9670 Virtual Desktops by Day, Computational Workloads by Night – An Example Infrastructure” can be viewed here.
   While conceptually this has been discussed amongst the community as a possibility, blogger Tony Foster (AKA WonderNerd ) made it his mission to script this functionality.   The idea being you can use the GPUs for VDI acceleration and as users stop using VDI for the night, the GPUs can be switched to perform high performance compute (HPC) functions.  The script sees spare GPU capacity at night and boots VMs for HPC, then does the opposite when it notices that more VDI VMs are needed (i.e. HPC resources are spun down to allow VDI users to work). It’s only been tested on VMware with VMware Horizon VDI but the concept is interesting.

M10s will still be cheaper cards for the foreseeable future but the T4s are where NVIDIA is heading.

Update on Windows 10 & vGPU

Windows 10 requires a vGPU. This used to be optional, but Entisys360 has been recommending vGPU for any new Windows 10 VDI or hardware refresh for VDI for the past 2 years. (see a quick video I did recently)

Skipping vGPU isn’t an option any longer.  Collaboration Suites, Microsoft Office, HTML5 web apps, YouTube and 4K/multi-monitor requirements all need vGPU.  vGPU isn’t for CAD users only longer. Microsoft has driven this with their UI modernization of Windows 10 (AKA they want Windows 10 to look pretty).

Now that we are less than one year from Windows 7 Extended Support ending (January 2020  https://support.microsoft.com/en-us/help/13853/windows-lifecycle-fact-sheet), if you haven’t already deployed Windows 10, then you need to get there soon.

LakeSide Testing/Benchmarks (see graph below): Windows UI modernization has resulted in a 50% increase of CPU due to the graphical interface (30% percent increase in CPU when no GPU is present and moving from Windows7 to Windows 10 with 2017 releases.  With the additional updates to Windows 10 in 2018 there is an additional 20% increase in CPU when no GPU is present.)

So everyone knows you need a GPU to deliver the best experience in VDI, but many VDI deployments decide to cross their fingers and see if could get away without a vGPU/additional cost.  Some did get away without a GPU (for now), but many did not and now have to look at update the timing of their next hardware refresh.

As part of the initiative of “vGPU for All VDI Users”, NVIDIA announced the NVIDIA GRID Windows Migration Acceleration Program with some of the server vendors: Cisco, Dell, Fujitsu, HPE, Lenovo, NetApp and Super Micro.  This program allows you to buy servers with either NVIDIA T4 or NVIDIA M10 cards and the necessary software.

In addition, pricing for the NVIDIA cards and the subscription software licensing has been reduced

• 5% reduction in the cost of the NVIDIA cards
• Additional years added to the subscription at no additional cost (three-year subscription for the price of two years, and a five-year subscription for the price of three.

This can reduce the price of vGPU even further to <$3 per user month. (*Still waiting for the final slide deck to get the actual number)

NVIDIA DGX Pod and AI (Artificial Intelligence… not my first name)

The NVIDIA DGX-2 was announced at last year’s GTC.  If you don’t know what a DGX is, you can check it out at https://www.nvidia.com/en-us/data-center/dgx-2/. It’s designed for simplifying Artificial Intelligence as a compute block with 16 GPUs and local storage. It also includes a lot of the software for implementing Artificial Intelligence and Machine Learning (ML) within your organization.

Now to what was announced this year…

NVIDIA announced during the keynote the NVIDIA DGX POD which is a reference architecture to expand the DGX configuration with additional storage and the necessary components to add additional capacity.  NetApp, Pure Storage, Dell EMC, IBM SpectrumAI and DDN Storage are the first ones to deliver the NVIDIA DGX POD architecture. This means larger data sets can be used (because of the additional storage) at faster speeds (without having to redesign the network) with a better price than non-GPU configurations.

This should address some of the concerns some clients had based on their larger data sets needing more storage.

Some final notes on NVIDIA GTC 2019.
AI/ML/vGPU are finding more and more use cases within different types of organization that I thought were very interesting

• Healthcare
  • Deep Learning for imaging (e.g. radiology image scans to identify issues earlier or more accurately than a human)
  • Deep Learning for speech and Analytics/Machine Learning (e.g. predicting health and next/best course of action based on large data sets)
• Financial Services
  • Fraud detection (e.g. compare a claim against comparable past claims to quickly identify fraudulent claims)
  • Data analytics with Insurance claims (e.g. scanning photos of car damage to identify fraudulent claims by insured or repair shop)
• Retail
  • Loss prevention (e.g. scanning security cameras at a store to identify patterns of theft/security weakness, or to identify products disappearing into a coat pocket)
  • Store Analytics (e.g. identify patterns in data as to understand why a store is performing better or worse than another location using multiple forms of data from video outside store, video inside store, product placement, employee-client interaction, per store or per region pricing/discount behavior, etc.)
  • Autonomous Checkout (e.g. identify correctly and quickly any product as it leaves the shelf to be purchased by someone)
• Manufacturing
  • Predictive Maintenance (e.g. constantly analyzing data of systems to determine when parts need to be repaired/maintained rather than using average run time or average hours in use to increase efficiency and reduce unnecessary costs performing work too early)
• Smart Cities
  • AI City Intelligent Video Analytics (e.g. “Big Brother” performing security in public areas like parks, mass transit or airports by scanning faces)

And most importantly with vGPU : ENDPOINTS MATTER. Make sure to check your thin clients to see if they still perform or you can start looking at the IGEL OS enabled 64-bit ARM-based NVIDIA Jetson platform.

I’ve attended multiple NVIDIA GTC conference and it is amazing to see the innovation that NVIDIA is enabling with their technology beyond allowing my kids to play Fortnite. Can’t wait to see what another 12 months bring us in GPU innovation!