NVIDIA GRID Days- My thoughts and what I have Learned
Last week, March 1st and 2nd, I was invited at the inaugural NVIDIA GRID Days at NVIDIA HQ, Santa Clare, CA. Together with 12 other ‘lucky ones’ I was selected to participate in this very first get together of people who are actively working with NVIDIA GRID. For me as a solution architect at PQR working on GRID enabled virtualization projects this was a huge opportunity to meet the NVIDIA GRID team and to learn from there insights as well as providing feedback from the field. Take the GRID 2.0 licensing for example. A lot of our existing GRID 1.0 (K1 & K2) are upset with this new licensing model which was introduced with GRID 2.0.
This blog will cover some of my experiences at these two GRID days at NVIDIA. There is lots of information disclosed by the GRID team these two days. I’m not going to cover it all unfortunately. Enjoy reading en if you have any questions, feel free to ask.
For those you are new to GRID and do not have a clear understandings of the differences between GRID 1.0 and 2.0, I’ll be writing a blog on this shortly, just keep an eye out on this blog site.
The first day stared with a real treat. Rachel Berry invited Mark Templeton, the retired CEO of Citrix. He paid us a visit and shared his cool toys. He brought his Vintage Electric Bike (http://vintageelectricbikes.com) and his Tesla Model S P85D to have a spin around the parking lot. The design of the bike is extraordinary and riding it was an awesome experience. For Mark, sharing his toys is, as he sad, a fundamental truth.
“Sharing is a fundamental truth”
Super Uber Sensor
After getting a ride to NVIDIA headquarters in Mark’s Tesla he showed us a Super Uber Sensor which was a prototype. The sensor has it all (e.g. infrared, thermal, sonar, radar, audio, video etc.). It’s able to track ‘flight time’, so it gets to know the distance of objects and it can track all knows devices, say like your smartphone. Image the use cases with such a sensor. A store for example. It is able to track all movements of customers and you can analyze the pattern people walk inside the store, which products are selected at the shelves and eventually are bought. Or use it at home, or Airbnb, to make it a really smart building.
Mark provided a quick demo with the The Super Uber Sensor and I have to say it’s a great piece of technique and opens up lots of opportunities but isn’t there yet. Great demo though.
After all these fancy toys it was time for the real deal, the green stuff, the NVIDIA stuff!
NVIDIA is a company we all know for its presence in the world of gaming. Delivering high graphics to get the best gaming experience. Besides the huge presence they also play a huge role in professional visualization, datacenter and automotive markets. NVIDIA is involved in almost any automotive brand such as Audi, BMW, Mercedes, Volvo and Tesla. You can imagine the GPU power required to drive a car autonomously, understanding it’s surrounding and possible obstacles. Take a look at NVIDIA DRIVE Automotive Technology here for more information.
The GRID Days where all about the Multi GPU Technology in the Professional Visualization and Virtualization. In this space Quadro is the main brand people know. NVIDIA has combined the graphics capabilities of Quadro GPU’s and the high performance computing power of Tesla GPU’s. With Quadro and Tesla GPU’s its all about Performance, Reliability and Support.
NVIDA was founded in 1993 by Jen-Hsun Huang, Chris Malachowsky and Curtis Priem. Today Jen-Hsun is the CEO, quite a big achievement!
At NVIDIA HQ they have set up a demo room. A room packet with all the green awesomeness. The took us from ‘regular’ PC and server graphics to gaming towards self-driving cars and deep learning. The self-driving car was something we witnessed at morning with Mark Templeton’s Tesla. Although at first I expected the demo room was all about their products and less about the functionality and use cases I was pleasantly surprised. The showed us an example of deep learning where a random set of hundreds pictures where analyzed in a few seconds. The pictures where categorized by they algorithm using the GPU resources. Without the GPU this would take much and much more time. You can understand how self-driving cars benefit from these GPU’s since ‘they’ have to analyze each and every object instantly to keep us all safe. And perhaps Mark’s Super Uber Sensor can help with that as well.
GPU Technology Conference (GTC) and Deep Learning
Last year I had the privilege of presenting at GTC. You can find my recording at the NVIDIA GTC On-Demand site (http://on-demand-gtc.gputechconf.com/gtc-quicklink/5B6Vp1r).
Last year GTC was all about deep learning and I must say I have learned a lot. This was all new to me, coming from a user workspace virtualization background. I found it really interesting though to step into this ‘new world’. Deep learning is going to open up much of opportunities. There are security use cases with deep learning. Analyzing audio of video streams and be able to protect IP along the way.
So if you ever have the opportunity to attend GTC make sure you do. It’s a really worth it. GTC 2016 is just around the corner, starting April 4th until 7th. If you can’t make it you get a second change at the very first GTC Europe conference, held in Amsterdam. September 28th and 29th, save the data.
VMware Blast Extreme
Will Wade, NVIDIA GRID Product Manager, told us all about how VMware Blas Extreme benefits GRID technology. With GRID you are able use hardware acceleration for PCoIP. This offloads the CPU, increases FPS, reduce latency and the amount of bandwidth that’s consumed. So it increases user density and UX. Erik Bohnhorst presented us with some of results from the field. With GRID enabled Blast Extreme there was a 18% user density improvement (per host) and a 13% increase of frame rates. He measured a 19% decrease of bandwidth consumption. So, as I sad, it increases UX and user density which results in a lower TCO.
There’s a excellent blog about the performance increase by Erik Bohnhorst at the VMware Blogs here (http://blogs.vmware.com/euc/2016/02/vmware-horizon-blast-extreme-acceleration-with-nvidia-grid.html) And also make sure to read Rob Beekman’s blog (http://www.robbeekmans.net/uncategorized/vmware-horizon-view-7-blast-extreme/) if you want to know more about Blast Extreme.
NVIDIA GRID 101
Andy Currid took us all on a GRID 101 session. He explained the basics of vGPU, telling about all the different engines (e.g. Graphics Compute, Video Encode, Video Decode and the Copy Engine) and how they all work parallel to each other. And off course the framebuffer, which moves data between the memory and the CPU. He told us how each individual vGPU enabled VM gets its own ‘Virtual BAR’ which provides ‘exclusive’ memory access to the GPU. Access to these virtual bars is controlled by the CPU’s Memory Management Unit (MMU). The hypervisor both creates MMU page tables that allow the VM to access the Virtual BAR as well as blocks access to locations outside the Virtual BAR. The GRID in guest (VM) drivers post the workload towards the GPU via channels. Each channel is a queue of work from an application towards the GPU engines. Each VM has dedicated channels, exposed inside the VM Virtual BAR. There are thousands of channels available, don’t you worry of running out. The GPU hardware scheduler places the work from active channels onto GPU engines. Engines sequentially execute work, signals work completions to the NVIDIA driver inside the guest VM. The Scheduler instantly loads the engine with the next work item. All the GPU engines access the memory. Each vGPU is assigned a fixed range of frame buffers (FB) for its exclusive usage. GPU’s MMU manages all engine access to the FB and system memory. This prevents the engines of accessing outside the vGPU’s FB, or the guest VM’s system memory.
This is just a part of all the NVIDIA GRID 101 information that Andy shared with us. The Scheduling and Timesharing, Memory Management as well as Single Root I/O Virtualization (SR-IOV) was tremendous informative but to much detail for now. The NVIDIA GRID 101 sessions involved that much of technical information I can write a blog on it entirely.
My Final Thoughts
First off all a big thanks to everyone at NVIDIA who made the GRID Days possible. I know Rachel Berry, Amanda Saunders, Luke Wignall, Victoria Rege and Erica Lee have put a lot of effort into the GRID Days. The entire NVIDIA GRID Team was actively participating the entire event, providing us insights and collecting feedback. This is how a community program should be, give and take, learn from each other.
Also Jim McHugh #GRIDModelRocks, the VP en GM of NVIDIA shared his insights on the company and they way they are heading. He reassured us that he understands our, as well as our customers concerns when it comes to GRID 2.0 licensing model. Stay tuned for announcements at GTC 2016. For now, we have to work with the Kickstarter promotion on GRID 2.0.
Working with GRID for quite some time now and get a better understanding of the technique gets me really enthusiastic about it. Maybe its because it’s technical innovation, which always gets an IT pro to stay up at late, or the real visual measurable achievements you are able to accomplish with it. I don’t know for sure, but I do no we can all benefit from the GPU’s NVIDIA is providing us today. Whether you are a designing the next Boeing with Catia or just adding these really ‘cool’ animations into your daughters elementary school presentation in PowerPoint (running on your company VDI environment 😉 . The CAD designer requires a solid GPU to perform his job. Providing the designer a GPU powered VDI desktop now allows him to perform hos job regardless of his location. No longer the large datasets are copied back and forth and IP is protected. For the helping dad it’s all about the User eXperience (UX). He expects the same UX at his enterprise environment as he is being offered by obtaining a good piece of equipment at his local Best Buy.
I know the licensing of GRID 2.0 upsets a lot of people. Even alternatives are being considered. We have to wait on that until GTC 2016 to get a clear understanding. Also the fact of not being able to mix and match different GPU profiles on the same physical GPU has reached me by customers in the past. For test purposes I can understand this comes. I have been told this is due to hard- and software limitations, at least today. This will change in the future. Also the ability to overcommit the frame buffer, which isn’t very unusual in the world of virtualization, isn’t possible today. This is still under investigation, the same as vMotion/XenMotion for vGPU VM’s. There is still more awesomess to come…..
I know this blog doesn’t cover all the excitement of the GRID Days, but I hope it provides some insights. I probably do people short, like the Click to Photon setup (Round Trip from mouse click(client) to screen update), the Performance benchmark by Luke Wignall and Erik. It’s just all to much to write down into this blog. I look back at a couple of really good days with good content and opportunity to provide feedback to the GRID team. It’s really good to have met the GRID team face to face as well as all the community friends.
Some of the community friends have wrote excellent blogs as well, you’ll found them here:
Barry Coombs – NVIDIA GRID Days – day 1
Barry Coombs – NVIDIA GRID Days – day 2
Victoria Rege – Top Virtualisation Talent Comes to NVIDIA for GRID Days
Sean Massey – It’s time to Reconsider My Thoughts on GPUs in VDI…
Dane Young – NVIDIA GRID Days 2016 Recap
NVIDIA GRID Forums – #GRIDDays
Also make sure to read the by 3D Graphics for Virtual Desktops Smackdown TeamRGE
Thanks for reading and if you have any thoughts of your self, please share……
Although NVIDIA paid for the travel and accommodation while I visited the NVIDIA GRID days, this has not influenced the content of this post. The gift I received deserves a nice place at the PQR Experience Center (pics will prove).
Trackback from your site.