K1/K2 vs. K340/K520


so I’ve done quite some research on the internet and haven’t really found answers for all my questions. So I hope that I’ll be able to find some support here. I will will just go ahead and shoot question after question at you - I hope you don’t mind. ;)

  1. I've heard that K1/K2 video cards are for professional working environments, that K340/K520 are newer video cards and that the K340/K520 are rather designed for cloud gaming. My question is: What technical difference really makes the K340/K520 better for cloud gaming? Could K1/K2 video cards be used for cloud gaming? Wouldn't those be better as they provide vGPUs?
  2. In the specifications of the K340/K520 is mentioned that 4-24/2-16 simultaneous players are possible. I'm wondering: How can there be that many simultaneous players without the ability to virtualize the GPU? Can several virtualizations use the same GPU?
  3. Furthermore, how can a shortage of RAM be avoided when playing the latest video games? Some of the NVIDIA GRID video cards only support 4GB RAM. Does that mean that only 1, max. 2 GTA V players can be supported with one video card?
  4. Is the technology which delivers a H.264 encoded video from the video card only available in the NVIDIA GRID video cards? What about normal video cards like a NVIDIA Titan X? Do those also provide such functionality?
  5. Last but not least: Is there any news about new GRID video cards being in development?

I hope that’s not too much. Thanks for all your help upfront.

  1. The K340/520 are the same silicon as K1/K2 but with different memory and clock speeds. K1/2 can be used for cloud gaming. Both are the same age and generation so 340/520 is not newer.

  2. This is down to the way the solution is deployed using code that the middleware vendor develops using the GRID SDK. They do use virtualisation technologies.

  3. Frame buffer is usually allocated on a fixed basis to the runnning session (this is how vGPU works)

  4. It’s a feature of the GRID SDK, so is available to Quadro cards of K2000 and greater. It’s not available on the GeForce product line.

  5. Announcements will be made in the coming months.


thanks you very much for your support so far, I really appreciate it!

However, I have some followup questions:

  1. So K1/2 and K340/520 are all great for cloud gaming but come with different technical specifications. However, I am not sure what specifications are the most important for gaming. Memory type (GDDR3 vs GDDR5)? Memory clock? Number of CUDA cores per GPU? Core clock?
  2. I can't really find a noticable technical difference between a K2 and a K520. The only difference I found in the technical specifications was that the K2 has a core clock of 745MHz and the K520 a core clock of 800MHz and that the K2 has a vGPU and the K520 doesn't. Everything else about the memory, bios, GPU and board are exactly the same. A reference to both PDF files including the technical specs.: http://www.nvidia.de/content/grid/pdf/GRID_K2_BD-06580-001_v02.pdf, http://wfcache.advantech.com/www/certified-peripherals/documents/ags-grid-k520-prl_Datasheet.pdf
  3. I understand that the frame buffer is allocated on a fixed basis but is there a way to add additional RAM to the video card? Either by adding special RAM on the mainboard or to somehow share memory in any other way? I read about NVIDIA TurboCache.
  4. How does process of virtualising GPUs & VRAM work in detail? Is it possible to set up a dynamic allocation of resources, so that only resources that are currently needed are assigned to a user?
  5. In case there are several video cards installed in a server, is it possible to virtualise the entire resources of all video cards together or is it only possible to virtualise resources per video card?
  6. Do these video cards support DirectX 12?
  1. K340/520 are designed specifically for gaming. K340 has 1/4 the Graphics Memory of the K1, but higher clock + higher power draw.

  2. Clock speed is the prime difference.

  3. No.

  4. https://gridforums.nvidia.com/default/topic/362/where-does-gpu-virtualization-occur-/

  5. Virtualisation is per GPU. Unless you choose an API Intercept solution such as VMware’s vSGA or Microsoft’s Remote FX Virtual Graphics Adapter

  6. Yes, once the driver is released.


thank you once again for your support! Though, I still have some followup questions:

  1. Is it possible to support more than 60fps with a K520 or is the amount of fps limited by the encoder?
  2. Does the vGPU technology on a K2 save additional milliseconds in comparison to a K520?
  1. It depends on how you implement it. The GPU is capable, but is the remoting protocol?

  2. What do you mean by "save additional milliseconds?"

Thanks for your quick response!

  1. What are you refering to when talking about the remoting protocol? Do you mean a protocol such as RTSP? I am just wondering whether these NVENC hardware encoders on those Kepler GPUS support more than 60fps. Because until now I always read about 720p@30fps and 1080p@60fps. I am interessted in 1080p@120fps streams.
  2. By saving additional milliseconds I refer to the timeframe in which the user presses a button until the fully encoded H.264 video is sent back to the user. Can the K2's vGPU technology impact that timeframe positively due to its hardware support?
  3. Can several NVIDIA GRID video cards function together in one server independently or is a SLI mode required? I ask because I've seen in NVIDIA GRID servers (http://www.pto.hu/wp-content/uploads/2013/03/nvidia-grid-vca.jpg) numerous NVIDIA GRID video cards function together and I was wondering if and how they are connected.

Remoting protocol means how are you transmitting the session from the host to the client. e.g. VMware PCoIP, Citrix ICA/HDX, VNC, RGS etc. Those are encoding guidelines for NVENC and the number of streams are a guide for how many streams at that resolution & frame rate. 1080p @ 120fps is going to be a challenge from a protocol perspective, most are limited to 60, and the bandwidth required for 1080p @ 120 will be substantial.

It makes no difference. Network latency will have a far greater impact.

It depends on the implementation. VCA is designed to leverage all the GPU’s. Virtualisation doesn’t allow for it.

What exactly are you trying to achieve?

  1. The remoting protocol will be none of the mentioned. It'll be an in-house developed piece of software. Therefore, I only require the information whether or not the NVENC encoder (NvFBC and NvIFR of the GRID SDK) are capable of encoding such a number of frame rates or not. Limitations of the remoting protocol exist, sure, but are not relevant at the stage of deciding which video card to buy.
  2. 'VCA is designed to leverage all the GPU's. Virtualisation doesn't allow for it.' - I'm not quite sure if you understood the question the way I intended to ask it. Please excuse me, let me rephrase:

    I am trying to set up multiple virtualised environments on the same server. For example: 4x Windows 10 with each 1 GPU with 4GB VRAM - all on one server. Now, with one K520 that’s impossible as it only has 8GB VRAM. But with 2 K520s it would be. That’s why I was asking if just sticking the 2nd K520 in another PCIe would be enough or if any special setup software- or hardware-wise like a SLI mode would be required.

  3. How important is the amount of cores in a CPU for the ability to support as many streams as possible? In graphs regarding the K340/K520 I've seen 12 core CPUs in use. Would a 6 core processor with for example higher clock rate perform worse/better?