AMD Multiuser GPU virtualization demonstrated at VMworld 2015

[HEXUS]

Said to be the "world's first hardware-based GPU virtualization solution".

Read more.

[sykobee]

This is for backend computation right?

We're no further towards a single PC driving an entire office (or part of an office) directly - e.g., 8/16 monitors for 8 users in a single system, 8 keyboards/mice into the system, a 16/32 core CPU, a redundant, reliable storage system and 64GB-256GB RAM.

Have one of these systems per "office desk unit (8 users)" in an office, and drop hardware support costs massively...

[l.set]

This is for backend computation right?

We're no further towards a single PC driving an entire office (or part of an office) directly - e.g., 8/16 monitors for 8 users in a single system, 8 keyboards/mice into the system, a 16/32 core CPU, a redundant, reliable storage system and 64GB-256GB RAM.

Have one of these systems per "office desk unit (8 users)" in an office, and drop hardware support costs massively...

This is (along with NVIDIA GRID) designed primarily for VDI/SBC, so a single server can support multiple GPU-accelerated users. Having a single PC acting as a single point of failure for that number of users would be pretty unwise and impractical imo. Giving everyone a cheap zero/thin client and getting them to connect over the LAN to a set of VDI/SBC servers means that everyone has a resilient session (and keeps everything centralised for security etc.). If it's for a small office then you could always drop a pair of tower servers in it and do the same thing. For most businesses though, hardware costs aren't usually the killer, it's software licensing and keeping compliance.

[crossy]

This is for backend computation right?

Wrong, OpenCL is part of the benefit, but the article does say "they have full access to native AMD display drivers for OpenGL, DirectX" and if their target market is CAD/CAM then it's graphics horsepower that's required (rather than a graphics-based number cruncher).

If you think about it this is a pretty neat play. After all it's going to be easier to justify that screaming professional graphics card costing 1,000's (or even tens of thousands) if you can say to the bean counters that it's a shared resource. But, as I.set correctly points out, the problem with that is that the software vendors seem to have traditionally penalised shared environments. And if you think domestic DRM is bad, you ain't seen nuttin' compared to the hoops needed to license some commercial software packages...

I'd be fascinated to understand (more research required) how this'll work in practice. So is the video signal going over DisplayPort or GBit. What really interests me though would be if this tech works it way down to the home market. The idea that I can build a real barn-burner of a PC and host virtualized Windows and Linux on it AND game in that virtual environment is something I like very, very much.

[zaph0d]

This is for backend computation right?

We're no further towards a single PC driving an entire office (or part of an office) directly - e.g., 8/16 monitors for 8 users in a single system, 8 keyboards/mice into the system, a 16/32 core CPU, a redundant, reliable storage system and 64GB-256GB RAM.

Have one of these systems per "office desk unit (8 users)" in an office, and drop hardware support costs massively...

What you're talking about is the way offices had computers in the 70's and early 80's. Everyone had a Terminal which was basically a monitor and keyboard connected to a mainframe computer, usually running a UNIX derivitive.

During the 80's it became more cost effective to ditch the large (and very expensive) mainframe and give all office workers their own computers.

But nowadays it looks like you're correct in that the performance of hardware over software is such that you could get mainframe-like performance from a single workstation class computer. The problem is still cost though, individual computers from a business perspective are about £100ish/per(in bulk) and use about 35w/hr each.
A workstation for say 15 users would cost easily £3000+ and consume 250+w/hr, so it all boils down to - will you save enough money on energy bills long term to offset the initial cost of purchase and is the support nighmare of having 15 non-productive workers per hardware/software failure instead of 1 worth the risk?

TBH I think offices will be better sticking with the current 1 pc per person approach, it's a little more expensive in terms of energy but cheaper in initial purchase and safer in support systems.

[l.set]

This is for backend computation right?

Wrong, OpenCL is part of the benefit, but the article does say "they have full access to native AMD display drivers for OpenGL, DirectX" and if their target market is CAD/CAM then it's graphics horsepower that's required (rather than a graphics-based number cruncher).

If you think about it this is a pretty neat play. After all it's going to be easier to justify that screaming professional graphics card costing 1,000's (or even tens of thousands) if you can say to the bean counters that it's a shared resource. But, as I.set correctly points out, the problem with that is that the software vendors seem to have traditionally penalised shared environments. And if you think domestic DRM is bad, you ain't seen nuttin' compared to the hoops needed to license some commercial software packages...

I'd be fascinated to understand (more research required) how this'll work in practice. So is the video signal going over DisplayPort or GBit. What really interests me though would be if this tech works it way down to the home market. The idea that I can build a real barn-burner of a PC and host virtualized Windows and Linux on it AND game in that virtual environment is something I like very, very much.

You're right on the commercial software compliance part. Try dealing with MS when it comes to licensing Office and Windows in a virtual environment will make you want to cry (MS do everything in their power now to lock you into SA or Office365). Then there's Adobe, AutoDesk etc. Normally I have to use something like AppSense or RES WSM.

As for in practice, the setup you want can be done. The multi-GPU cards designed for this stuff, like GRID, don't have any external display connectors, instead everything is done through the remoting protocol to the VM, like PCoIP or ICA. Only certain cards support virtualising the GPU, but a lot more GPUs are supported to be passed through the hypervisor to the virtual machine with a one-to-one PCI mapping (vSphere and XenServer both support it). Hell, one of my main demos to customers is showing playing Tomb Raider on a Win 8 VM over our corporate LAN, with the server sat in our test lab at the other end of the building. Really the limitation is latency, well, that and the relevant VMware/Citrix licensing.

[crossy]

What you're talking about is the way offices had computers in the 70's and early 80's. Everyone had a Terminal which was basically a monitor and keyboard connected to a mainframe computer, usually running a UNIX derivitive.

Erm, and what about the mini's that Digital were making? Not everyone could afford a big, humpin' IBM mainframe. And don't forget that mainframes were very good at high IO workload, but absolutely hopeless at something that needed a lot of cpu. On the other hand DEC's VAX 11/780's (and successors) were pretty darn good at that, especially hooked up to one of the various 3rd party co-pro's like FPS. IBM's mainframes don't run Unix (except as a VM?) and DEC's had a long and illustrious part in the history of Unix, although you really were supposed to be running VMS.

During the 80's it became more cost effective to ditch the large (and very expensive) mainframe and give all office workers their own computers.

But nowadays it looks like you're correct in that the performance of hardware over software is such that you could get mainframe-like performance from a single workstation class computer. The problem is still cost though, individual computers from a business perspective are about £100ish/per(in bulk) and use about 35w/hr each.
A workstation for say 15 users would cost easily £3000+ and consume 250+w/hr, so it all boils down to - will you save enough money on energy bills long term to offset the initial cost of purchase and is the support nighmare of having 15 non-productive workers per hardware/software failure instead of 1 worth the risk?

TBH I think offices will be better sticking with the current 1 pc per person approach, it's a little more expensive in terms of energy but cheaper in initial purchase and safer in support systems.

Again, mainframes are brilliant at high IO loads - that's why IBM still sell zSeries boxes. Other thing that mainframes are damn good at is reliability - uptimes for the mainframes here make the Wintel guys cry. Trouble for the mainframers is that a lot of the Unix server manufacturers have upped their game and made RAS features a big deal. E.g. IBM's high end System p boxes can be configured with a lot of redundancy in them - e.g. hot swappable PSU's. And those can run Linux as well as their tailored versions of Unix, (although actually IBM's zSeries have a pretty darn good setup for running Linux VM's - ideal for that high-transaction database server?!)

Similarly there's a lot of power-saving features in these "big iron" boxes where processors or even execution units can be powered down if not in use. And that's where virtualization developments such as these GPU ones come in - they let you make a lot better use of the hardware, and avoid (expensive?) unused resource. The desktop PC's are great for the day-to-day Office, design, etc stuff, but there's more to computing than that. And where I'm curious is what effect increasing virtualisation and better tablet devices is going to have.

[lumireleon]

I remember using the old timer 2004 3DS MAX that had a support for multi PC rendering of projects, this idea from AMD will be killer but will the current APPS support it? like when am flying XPLANE 10 on a slow pc yet the rendering workloads is done by the server