The New NVIDIA GRID
Today, August 17, NVIDIA publicly announced big changes in the GRID product line. The introduction of a brand-new GPU architecture, the new 5.0 GRID software and a licensing rebranding. Most excited about these announcements are the new GPU’s. These are based on the Pascal architecture and like its predecessor, the Maxwell GPU’s, they lead the way into accelerating datacenters all over the globe. With the new Pascal based GPU’s the Industry most powerful enterprise virtual workstation becomes available.
Originally NVIDIA focused the PC gaming market with delivering the best GPU’s. But along the way NVIDIA redefined computer graphics and revolutionised parallel computing. (take a look at this video for an explanation between parallel and sequential computing). Parallel computing powers todays Deep Learning (DL) and ignited Artificial Intelligence (AI). And let’s face it, AI is on everyone’s mind and it is already all around us. Take a moment and watch the opening keynote of the NVIDIA GPU Technology Conference (GTC). NVIDIA GPU’s are they super fast and efficient brains of computers, self-driving cars and robots.
NVIDIA revolutionising AI and DL with its GPU’s has a great benefit for GRID. All the knowledge and development comes straight towards the datacenter solutions with the Tesla datacenter GPU’s. The Tesla GPU’s are perfect for accelerating virtual desktops and delivering Windows and Linux desktops and apps to any user on any device, anywhere they want……… OK, as long as they have a connection.
Before we take a look how these new Pascal based GPU’s fit into the GRID product line lets take a look at the Pascal architecture itself. NVIDIA says, “Pascal architecture is purpose-build to be the engine of computers that learn, see and simulate our world.” But what are the technological advances?
First of all, Pascal GPU is the the world’s largest FinFET chip ever built. It’s engineered to deliver the fastest performance and best energy efficiency for workloads with near-infinite computing needs. The Pascal GPU has 150 billion transistors build on a 16 nanometer FinFET chip. That’s sounds awesome, but the real value to the ones using these GPU’s is it’s performance. Today, the Pascal is the most powerful compute architecture ever build inside a GPU. It is able to deliver over 5 TFLOPS of double precision performance for a High Performance Compute (HPC) workload. Pascal powered systems offer much faster and efficient performance over its predecessor, the current Maxwell GPU architecture. A brand-new scheduler increases the performance and the Quality of Service (QOS) by a much-improved resource allocation per user. The QoS scheduler is able to guarantee GPU cycles per VM. Naturally, NVIDIA has some nice numbers to let us understand how much faster the Pascal architecture is compared to the Maxwell but it all depends on the workload. Another rather big thin is the fact that Pascal is the first architecture to integrate the NVLink high-speed bidirectional interconnect. NVLink enables fast communication between the CPU and GPU, and between GPU’s. And luckily for us GRID users, all this Pascal development becomes available to us. Which graphics cards are we talking about when it comes down to GRID availability?* data not yet available
Pascal Based GRID
As you could have read in previous blogs (here and here) NVIDIA GRID is around for quite some time. Starting with the Keplar based K1 and K2, followed by the Maxwell based M6, M60 and M10. And today the new line-up, the P-series. NVIDIA is launching the P4, P40, P6 and P100. These four Pascal based graphics cards are all available with the new GRID 5.0 version. The P40 and P6 perfectly fit in the GRID product line as shown below.
The P40 and M60 are optimized for Performance, the P6 and M6 optimized for blades and the M10 for User Density. With the Pascal based GPU’s added to the GRID line-up we get more choice and possibilities to get the right fit for every case. The P100 and the P4 are also usable in a GRID scenario.
What does the introduction of these new Pascal bases GPU’s means for GRID?
First of all, we are able to use the latest and greatest GPU’s out there to accelerate virtual desktops and apps. Before the Pascal launch there where only three options. We had to choose from the M10, M60 and M6. All three cards had their ‘own’ optimisation, M10 for user density, M6 for blade and M60 for performance. With the P40 and P6 added we get more choices. We are more flexible and able to mix and match the right solution.
Especially when it comes to blade and performance optimised solutions.
With the P6 the amount of GPU memory (framebuffer) is doubled. This still in a maximum of 16 virtual desktops with 1GB framebuffer each compared to max 8 virtual desktops with 1GB framebuffer. Yes, I do know the maximum GRID vGPU instances with the M6 is 16, but this is with the 512MB framebuffer size. This 512MB vGPU profile is no longer available with the Pascal GPU’s. No worries, nobody was successfully using it, or perhaps better, no one was able to use it. Operating Systems like Windows 10, productivity applications like MS Office (Office365/Office216) and browser apps are increasing the need for GPU’s and 512MB framebuffer doesn’t make any sense these days.
At the performance optimised solutions, the P40 and the M60 also offer us more flexibility. The M60, with 2 GPU’s and 16GB framebuffer offers us 16 virtual desktops with a 1 GB vGPU profile. With the P40 this increases to 24, resulting in 150% sessions compared to the M60, and that’s on a single GPU. The P40 also introduces new vGPU profiles with the 6GB, 12GB and 24 GB vGPU profiles. Keep in mind that some OEM’s will offer up to three cards in a single chassis. Take a Dell PowerEdge R730 for example, stacking three P40 will result in 72 virtual desktops with a 1GB vGPU profile.
With the new Pascal based line-up a new GRID software release is also announced. This release, version 5.0, includes a new GRID Manager, guest drivers and a new license server. This new version of the license server, available as a Windows and Linux installer, introduces High Availability (HA). This means, productivity keeps going great in case of a license server failure.
More importantly, the 5.0 version brings us end-to-end monitoring. We are now able to get a detailed understanding of applications GPU usage. This fine-grained GPU understanding is personalised per user via guest level monitoring. This enables us to get a better understanding of how GPU’s are being utilised and provides us with the ability to do predictive management. This end-to-end monitoring fully integrates with your existing virtual desktop infrastructure. Integration with VMware vROps, Citrix Director or XenCenter, yes it’s there. Integration with eG Innovations, Lakeside Systrack, ControlUP and LiquidwareLabs is possible as well. Guest level monitoring and host level monitoring was introduced a while ago and today GRID is the only out there with these management and monitoring capabilities. Underneath a table of both the host- and guest level monitoring capabilities, with ‘per application’ included.
GRID Software Editions
Another announcement today was the rebranding of the virtual Workstation (vWS) edition. This edition is rebranded to Quadro Virtual Datacenter Workstation (vDWS). Quadro was already available to mobile workstations, desktop workstations and as external graphics and if I’m not mistaken a feature available with the previous vWS edition. Nevertheless, the Quadro certification can be very important in the aviation industry for example. With this Quadro certification, GRID offers customer to choose from 120+ certified servers based on 30+ vendors. For a complete list of servers and vendors, take a look at the GRID Certified Servers list.
With the new vDWS edition the complete list of GRID editions is as follow:
It is really astonishing in what pace NVIDIA is able to deliver new GPU architectures. When I take a look at the competition, NVIDIA keeps improving their products in a high pace. And this isn’t all about hardware. The software part of their solutions, and this is not only about drivers, keeps improving as well. The NVIDIA GRID software solution (GRID Manager, drivers and license server) keeps being improved. End-to-end monitoring for example doesn’t come overnight and really need development resources and an ecosystem. So, hands up for this. Another great achievement is XenMotion support for GPU accelerated VM’s. During Citrix Synergy (May 23 until 25th 2017) NVIDIA had a great demo at their booth. I took the liberty of creating short video demonstrating XenMotion with a vGPU enabled VM (take a look here). As you can see the screen only freezes for a couple of seconds. This was a Tech Preview, but in the near future Citrix is supporting this with XenServer. When VMware will follow, all depends on release cycles and priorities. Next up, if you would ask me are the ability to mix different vGPU profiles on the same physical GPU and the ability to overcommit the physical framebuffer
With today’s GRID line-up, we have more flexibility in finding the right fit for every challenge. But what about tomorrow? Clearly the Pascal architecture is the successor of Maxwell. Does this indicates the P40 to replace the M60 and the P6 replacing the M6 somewhere in the future? And what about the M10, the Maxwell GPU based card optimized for user density. There isn’t any news on some Pascal based card optimized for user density. And this will bring us to support. What will happen with support for the Maxwell based M10, M6 and M60?
Today’s official statement is, NVIDIA will support both the M-series for GRID as well as the P-Series. The new GRID 5.0 is backward compatible, so no worries. But at some point, NVIDIA will announce End Of Sale (EOS) for the M-series. When this will be, nobody will or can tell. The only thing I do know, when EOS is announced we have another 9 months to get our act together. For now, both the M-series and the P-series are supported and customers have to decide which one is a better fit based on their requirements, future plans and budget. Also, there isn’t a successor for the M10 yet. A Pascal based card optimized for user density. All Pascal based cards are single GPU.
To come to a conclusion on this final thought, my congratulations to everyone at NVIDIA involved in the new GRID product line. Great achievements to be honest. I am convinced a successful virtual desktop solution can only be done with the right products. This includes a GPU as well since it is our goal to deliver a physical desktop/laptop like, or perhaps better user experience. It makes sense to deliver centralized virtualized desktop and apps based on Windows and Linux. Keeping data secure in our datacenters, private clouds and keeping it close to the virtual desktops and client applications. Making them securely accessible wherever you are and on any device. With today’s increase on operating system and applications that really benefit from the presence of a GPU, these GPU’s are powering productivity and user experience.
Trackback from your site.