I'm aware, but I also would guess that quick time to market is of importance to make sure you get your part of the cake.
How much competition is in this space already? Does Nvidia have a similar dedicated product?
Anarchist4000
Veteran
That's it. Thanks.
Quite the opposite actually if you read and understand it. Just need to infer what technologies Vega or HPC Zen would bring to those capabilities. Then a little time for the software to mature. Hardware companies and ISVs would likely be developing for upcoming capabilities and not existing hardware. So while the article is talking about a Tonga, Vega with access to system memory could change those capabilities substantially.
So Nvidia does have a x86 processor from Intel or AMD working with NVLink to provide direct access to system memory, added full hardware virtualization, and scrapped a lot of their software tools that are no longer needed?
Well if you have 127 friends all doing essentially the same thing you would need 128 models, each on their own VM. If it's just you, go buy yourself a workstation instead of a VM. The whole point of a VM is sharing the hardware among a collection of users. Creating a giant machine that only one user utilizes at a time sort of defeats the purpose.
Yet they are still limited to 16 users in a office type environment, because the horse power necessary can't do workstation type applications in VM's.
This is not necessary for the time being, and software can't utilize this right now either. Software will never get scrapped, unless hardware is fully complaint with each other, and that we know won't happen in the short term or possibly even mid term.
You can't do that right now, you don't have the horsepower to drive that, its not ONLY about the amount of memory you have. Per unit you can have 1 or 2 people doing this at once with the type of things I'm doing. You still need a lot GPU horsepower to drive things. Even with a cluster of 4 GPU's your limiting factor will be the GPU horsepower.
What you are looking at it with GPU vm's right now, anything more than 6 uses, basic office type stuff. 2-6 users all you can do is view a CAD file. 1-2 users basic low end workstation performance.
Just load up a dense model in 3dmax and start fiddling with it.
On my system Titan P with 64 gigs of ram the specs I just stated for my model, if I uses more than 6 sets of 8k textures, vram starts page flipping. So I know that is my limit there along with the amount of vertices needed to be stored for all the change states the 64gb of system ram is close to the max, if I start going beyond those figures, it slows down a bit too much and it slows down my work flow.
This means, a VM graphics system if I was to use one, I would need these kinds of specs dedicated to myself to run it. Right now AMD can't provide this to me. nV can though. Will Vega solve this, I don't think they will be able to out do nV in the graphics segment anytime soon, since CUDA is driving many of the rendering applications I use.
Btw the specs I mentioned are above and beyond what is typically used. We use these things to get the textures for lower details meshes which will run a lot better on a VM system, but up till that point we need those type of models to create the other assets. (I pointed this out because we are talking about crazy amounts of ram being addressed to just the GPU, which I don't know of anyone that would need that type of amount right now or even in the near future, when that needs comes up yeah it then becomes a viable option, up till then its a nice feature but won't be taken as a driving force for sales).
Last edited:
Ecosystem is most important, so SR-IOV validation, VM software enablement and then application certification (for the "professional" VDI space) are, in large, more key than the generation of hardware underneath.
Just so we are clear, this is 16 users per GPU and you can enable 4x FirePro S7150x2's in a 4U server, which would enable up to a max of 128 users per server. In this scenario you're actually more likely to be CPU bound than GPU bound.
yes is a typical office type environment I agree, for workstation CAD, 3D file manipulation, it will become GPU bound.
Large design deployments will be using PLM systems that access the CAD design files (hence need the software and a capable platform ecosystem) even without heavy load on the GPU; rule of thumb is about 10:1 PLM users to designers.
Anarchist4000
Veteran
CAD is probably not the best example of what I'm describing. Point being, it could allow a multitude of relatively infrequent users to be sharing a larger GPU without partitioning resources. Cases where memory capacity in aggregate would likely become an issue. It could be a bunch of users updating their spreadsheets from tablets.
As for your GPU bound cad file, is that because of compute power or memory limitations? A lot of these current capacity issues you are mentioning seem like they may be alleviated by better access to system memory. If using a fast interconnect, the available VMs are likely limited by CPUs in the system. Even for IBM with NVLink I'm not sure they are linking more than one GPU at a time to the CPU. It would ultimately be wasteful as it only helps if you have system memory bandwidth to feed it. I'd imagine a system with large bus to GPU makes a good CAD workstation, but a VM might be a bit much unless you had a really big system and didn't oversubscribe active users.
As for your GPU bound cad file, is that because of compute power or memory limitations? A lot of these current capacity issues you are mentioning seem like they may be alleviated by better access to system memory. If using a fast interconnect, the available VMs are likely limited by CPUs in the system. Even for IBM with NVLink I'm not sure they are linking more than one GPU at a time to the CPU. It would ultimately be wasteful as it only helps if you have system memory bandwidth to feed it. I'd imagine a system with large bus to GPU makes a good CAD workstation, but a VM might be a bit much unless you had a really big system and didn't oversubscribe active users.
Ah interesting I thought that would have to rely upon traditional Xeons and more likely Skylake Xeon when released, be curious to see how they manage to do this with KNL as it ideally requires optimised code compared to traditional x86.
Separate to this are the recent high profile acquisitions highlighting that they do feel a bit of pressure IMO and the importance of latest trends, but they do have a strong market share they need to defend from being squeezed in the HPC-enterprise-traditional servers and I think they would like to restrict IBM/Nvidia/Cray/Fujitisu-ARM.
For those curious, here is a recent presentation of KNL with additional slides by Senior Principle Engineer at Intel (some of this has been in earlier info but not all): http://cgo.org/cgo2016/wp-content/uploads/2016/04/sodani-slides.pdf
Cheers
Last edited:
If I start manipulating vertices, and polygons, those calculations are done by the GPU, with dense models, this gets very tasking on the GPU, not its memory interface. Textures are what gets the memory and bandwidth needs, but if I don't have enough memory or bandwidth I can always go back to using procedural textures and shaders, which is what I did back when we were limited to less then a 512mb for graphics memory.
Easy to see this when you start using multiple view ports when manipulating vertices. The more view ports that you have, and I typical use the standard 4 views, top, left or right, top or bottom, and perspective, when I'm doing this, but if the system starts slowing down, I will switch to a 3 or 2 view port layout.
Lets just say we are using cad for automotive design. Animated those polygons also are GPU based, and they tend to use fairly dense meshes as well, not as dense as what I'm talking about.
Last edited:
The flipside on CAD it tends to a more bursty workload, concentrated when a sudden model rotation takes place or you want to view the model with the materials. Unlike games, its not pushing to deliver a constant 60FPS from the draw call level.
Anarchist4000
Veteran
For CAD I think more along the lines of assembly lines, cars, aircraft, and smaller models for CNC machines. Tasks that a company could have a lot of designers working on, but not load down the entire system consistently. In your case if multiple people had large files constantly loading the device the VM is somewhat pointless. Even for memory consumption a single user likely avoids the limit, but multiple would likely hit it.
True
If viewing the same file, your memory should stay the same, only processor power would be needed......
if two people are working on the same file, yeah, but I don't see why two people would be doing that, I can see the hierarchy system within a cad file being used with multiple sub files and those sub files are being used by different people but not the parent file.
Pretty much results are you still need the same total amount of horsepower, bandwidth, memory as each individual system in a traditional setup, but with this new tech, you are consolidating everything into one area, and along with it the data too. From an organizational standpoint and operations its great, it doesn't mean you need less of anything though.
Last edited:
Anarchist4000
Veteran
These users likely wouldn't be working on the same file. Likely just parts or separate projects. With the bursty nature as Dave mentioned you wouldn't need the same quantity of hardware. That's generally the point of VMs is that you can consolidate everything and share stronger hardware among users. Take the idle time from one user and give it to another. If a user is consistently utilizing the resources they'd be better off with a workstation than VM. If you had 100 engineers that needed to design bolts, would you want to provision a workstation for all of them or use a VM to reduce hardware costs?
There's a big list of applications there and CAD is only touching on one of them. I could still see mobile devices leaning on VMs for processing power in any number of applications.
Ideally that is what will happen, but reality that doesn't happen over night, also with other competitors in the market who are more solvent and have more features and have better hardware, how do you expect expect AMD to make headway?
As we were talking about it also depends on what each designer is doing. You think the guy that does airfoil designs will want to be on the same server as everyone else?
The friend of mine that works at cargril, his dad was an army engineer, back in the 80's they were using things similar to virtual machines with CAD (of course back then CAD was being run off of CPU's). The reason why they didn't take off because the needs of the design group and others were just too different to create a system that was flexible enough and vertical enough to sustain what they need. This problem has never gone away.
These are not new technologies. They are new in the sense new features are there that make them more flexible, but the root of the problem still exists.
Firstly you need the organization to change their perspective on their business model, that will change their needs from a infrastructure point of view. And this will take many years for large companies to do. Even 5 years or 10 years can be short time for something like this to change. Newer smaller companies can adapt quicker though but still time to do something like is extensive. Bigger companies if they move too fast and change their business model can break their back.
As we were talking about it also depends on what each designer is doing. You think the guy that does airfoil designs will want to be on the same server as everyone else?
The friend of mine that works at cargril, his dad was an army engineer, back in the 80's they were using things similar to virtual machines with CAD (of course back then CAD was being run off of CPU's). The reason why they didn't take off because the needs of the design group and others were just too different to create a system that was flexible enough and vertical enough to sustain what they need. This problem has never gone away.
These are not new technologies. They are new in the sense new features are there that make them more flexible, but the root of the problem still exists.
Firstly you need the organization to change their perspective on their business model, that will change their needs from a infrastructure point of view. And this will take many years for large companies to do. Even 5 years or 10 years can be short time for something like this to change. Newer smaller companies can adapt quicker though but still time to do something like is extensive. Bigger companies if they move too fast and change their business model can break their back.
Last edited:
Thank you for providing more insight into this area I'm obviously not too familiar with!
I only know of something like this. What I don't know about are the capabilities besides claimed support for 16 virtuel users and a 100W TDP:
http://images.nvidia.com/content/pdf/grid/data-sheet/nvidia-m6-datasheet.pdf
or this:
http://images.nvidia.com/content/pdf/grid/data-sheet/nvidia-m10-datasheet.pdf
Claimed 16 users for each of the 4 GPUs.
But I think both are rather recent too.
Last edited:
D
Deleted member 2197
Guest
Tesla M6 - 16 vGpu instances
Tesla M60 - 32 vGpu instances
Tesla M10 - 64 vGpu instances
http://images.nvidia.com/content/pdf/grid/guides/nvidia-grid-datasheet.pdf
Tesla M60 - 32 vGpu instances
Tesla M10 - 64 vGpu instances
http://images.nvidia.com/content/pdf/grid/guides/nvidia-grid-datasheet.pdf
Last edited by a moderator:
Since this thread was last discussed we've had both Alibaba adopt FirePro S7150 x2 and Google adopt FirePro S9300 x2 for cloud service offerings.
