AMD Unveils World's First Hardware-Based Virtualized GPU Solution at VMworld 2015

[mxnerd]

AMD Multiuser GPU Enables Consistent, Predictable and Secure Performance for Virtualized Workstations

http://ir.amd.com/phoenix.zhtml?c=74093&p=RssLanding&cat=news&id=2083146

This is great news for those doing virtualization in corporate environment.

 

[Stuka87]

Really interested to see how well this works.

 

[soccerballtux]

intresting

 

[el etro]

OMFG, This is godsend!

God save AMD!

 

[SolMiester]

Wonder how this differs to NV vGPU tech which uses pass-through?...and not sure where they get the world first virtualised GPU platform?, NV has been doing this for awhile now/

 

[Elixer]

This would be kick ass if it works as advertised!
I can't wait for the day when you can seamlessly switch between OSes and still have full hardware acceleration.

 

[ViRGE]

Wonder how this differs to NV vGPU tech which uses pass-through?...and not sure where they get the world first virtualised GPU platform?, NV has been doing this for awhile now/

Agreed. I'm curious what makes this "hardware" and NV's solution "software." NV is time-slicing, but there's an MMU in place to enforce user isolation, and each user has access to the full capabilities of the platform (especially now that CUDA is available).

 

[Stuka87]

Wonder how this differs to NV vGPU tech which uses pass-through?...and not sure where they get the world first virtualised GPU platform?, NV has been doing this for awhile now/

NV does not have a virtualized GPU, they have passthrough as you stated. Passthrough allows one VM to use a GPU. Yes its times sliced to allow multi users, but they cannot share it simultaneously.

AMD's solution allows multiple VMs to use the GPU hardware in the server. This is a first.

This is huge for those of us with large VMWare deployments and have a need for hardware GPU's to handle OpenCL or the like.

 

[saratoga172]

Couple things.

There are currently three levels of GPU deployment with nVidia cards today.

Svga- software virtualized graphics adapter. You install the graphics driver on the host and the vm software virtualized the card. The guest os see it as a software solution. Only really good for standard office stuff.

Vdga- virtual dedicated graphics adapter. This is the pass through option. It's a 1 for 1 thing. Using grid K1 cards for instance we can setup 4 people since one card has 4 cards built in. Designed with cad use in mind but in my testing I'd be hard to find that acceptable.

Vgpu (newest of the bunch and requires vm 6+). This basically mixes the above two. It allows multiple people to share a card. Provides much better performance than svga.

It sounds like this new solution from AMD has actual virtualization tech built into the card like a hyper visor to allow it to be shared out. Haven't read much into it so I'm curious how it works from a tech standpoint but this should allow less interference and the vm to run more natively with the graphics.

 

[saratoga172]

I'm curious how exactly it's different from this because it's what it reminds me of. Haven't had a chance to implement it yet at work, but hopefully soon.
http://www.nvidia.com/object/virtual-gpus.html

 

[ViRGE]

Couple things.

There are currently three levels of GPU deployment with nVidia cards today.

Svga- software virtualized graphics adapter. You install the graphics driver on the host and the vm software virtualized the card. The guest os see it as a software solution. Only really good for standard office stuff.

Vdga- virtual dedicated graphics adapter. This is the pass through option. It's a 1 for 1 thing. Using grid K1 cards for instance we can setup 4 people since one card has 4 cards built in. Designed with cad use in mind but in my testing I'd be hard to find that acceptable.

Vgpu (newest of the bunch and requires vm 6+). This basically mixes the above two. It allows multiple people to share a card. Provides much better performance than svga.

It sounds like this new solution from AMD has actual virtualization tech built into the card like a hyper visor to allow it to be shared out. Haven't read much into it so I'm curious how it works from a tech standpoint but this should allow less interference and the vm to run more natively with the graphics.

I am not a systems administrator, but the more I read, the more I get the impression that whoever is in charge of promoting this for AMD is asleep at the wheel.

http://www.amd.com/Documents/Multiuser-GPU-Datasheet.pdf

I'll give AMD a pass for their datasheet being out of date with respect to GRID 2.0 and the increased user limits. Still, there are claims here in comparing it to GRID that are clearly wrong. "Dedicated share of local memory" GRID does, and "Stable, predictable performance" is a bit nebulous, but in a time-slicing system you certainly expect GRID to be able to deliver that (and I've never seen any claims to the contrary).

AMD's entire premise seems to be a year out of date, written on the basis that NVIDIA's vGPU capabilities don't exist. All of this only makes sense if you're comparing AMD's solution to a true software solution, e.g. RemoteFX and VMware vSGA.

Which isn't to kick AMD here. Based on this announcement they finally have a vGPU technology of their own, which is a good thing for them. It only brings them back up to parity with NVIDIA, but it's where they need to be.

 

[mxnerd]

http://gfxspeak.com/2015/08/31/hardware-virtualized-solution/

The multiuser GPU is designed to work in environments using VMware vSphere/ESXi 5.5 and up, with support for remote protocols such as Horizon View, Citrix Xen Desktop, Teradici Workstation Host Software and others. It uses VMware’s VIB driver with VMware certification, and AMD says no changes from pass-through are needed to run on vSphere/ESX 5.5 or vSphere/ESX 6.0. Other hypervisors with SR-IOV support can be enabled in the future.

We asked AMD if they plan to build a rack-mount unit as Nvidia has done with their Grid product. AMD said no, they did not want to compete with their customers and would support their customers in the development of such chassis. So there won’t be any specific AMD Virtualize boxes in the immediate future. However, we can envision AMD developing a generic box for white-box, tier three server suppliers, and VARs who want to offer a turnkey solution with virtualization.

AMD’s approach to specific GPU deterministic partitioning, rather than a dynamic load-balancing approach like Nvidia offers, is bound to have adherents. AMD said they did it based on feedback they got from users. AMD claims their direct hardware solution offers better throughput, and tighter communications with the drivers and hypervisor, and points out one user cannot affect another user’s GPU performance.

AMD will be showing this technology at the JPR Virtualize conference in San Francisco on 29 October: https://jonpeddie.com/events/details/jpr-virtualize-2015

 

[Glo.]

Agreed. I'm curious what makes this "hardware" and NV's solution "software." NV is time-slicing, but there's an MMU in place to enforce user isolation, and each user has access to the full capabilities of the platform (especially now that CUDA is available).

Asynchronous Compute comes to mind .

I think there is a lot of power and capabilities to come out from GCN GPUs in next months.

 

[Silverforce11]

Zlatan covered this previously in the Nano thread.

It was due to this:

Fiji has 4 ACEs and 2 HWS instead of the normal 8 ACEs.

But it seems the 2 HWS can perform the function of 4 ACEs AND this..

All newer GCN 1.2 cards have this configuration. There are 4 core ACEs. The two HWS units can do the same work as 4 ACEs, so this is why AMD refer to 8 ACEs in some presentations. The HWS units just smarter and can support more interesting workloads, but AMD don't talk about these right now. I think it has something to do with the HSA QoS feature. Essentially the GCN 1.2 design is not just a efficient multitask system, but also good for multi-user environments.

Most GPUs are not designed to run more than one program, because these systems are not optimized for latency. They can execute multiply GPGPU programs, but executing a game when a GPGPU program is running won't give you good results. This is why HSA has a graphics preemption feature. These GCN 1.2 GPUs can prioritize all graphics task to provide a low-latency output. QoS is just one level further. It can run two games or a game and a GPGPU app simultaneously for two different users, and the performance/experience will be really good with these HWS units.

http://forums.anandtech.com/showpost.php?p=37656793&postcount=204

Zlatan is spot on so far.

 

[zir_blazer]

AMD says it does GPU virtualization using SR-IOV, like some of the high end Intel NICs (Intel i350). Each GPU should provide several virtual PCI Device functions that you can do PCI Passthrough of, in a rather identical fashion to what's already available. However, it seems that you need heavy support to get SR-IOV working at the first place.

 

[Despoiler]

Slick work by AMD. I can't wait to see it in action.

 

[SolMiester]

I know a few Linux users who go NV due to AMD linux drivers...wonder what their take on this well be?

 

[zir_blazer]

I know a few Linux users who go NV due to AMD linux drivers...wonder what their take on this well be?

Something is really wrong there. Lately AMD is having stronger Linux support that nVidia. The Open Source Driver is better than nVidia one (nouveau) since AMD has been releasing a lot of documentation to make a working Driver from pretty much scratch, while Nouveau is brute reverse engineering. On close source Drivers, nVidia has been historically better on Linux, but I think that doesn't hold true anymore, either.
And, you're missing AMDGPU, Linux Catalyst Open Source replacement, but these are from AMD themselves. Looks promising.

 

[SolMiester]

Something is really wrong there. Lately AMD is having stronger Linux support that nVidia. The Open Source Driver is better than nVidia one (nouveau) since AMD has been releasing a lot of documentation to make a working Driver from pretty much scratch, while Nouveau is brute reverse engineering. On close source Drivers, nVidia has been historically better on Linux, but I think that doesn't hold true anymore, either.
And, you're missing AMDGPU, Linux Catalyst Open Source replacement, but these are from AMD themselves. Looks promising.

Good to hear!

 

[Jubassi]

yes, vmware is going to be so much better with this!

 

[SolMiester]

Except VMware is losing ground to hyper-v with free HA, and 2 free VM licenses...

 

[zir_blazer]

I'm always dissapointed that I never hear about Xen and Linux KVM... Since SR-IOV is supposed to provide multiple virtual devices from a single physical device and you do PCI Passthrough of those, you can make use of all the working Xen and KVM infrastructure.

Regardless, chances are that most of the users of these cards are going to purchase fully assembled OEM systems. And the price of the cards themselves will not be interesing for the home user that wants to do GPU Virtualization being from the premium professional segment - XenGT will take care of that for free on Haswell + systems.

 

[Headfoot]

Zlatan covered this previously in the Nano thread.

It was due to this:

Fiji has 4 ACEs and 2 HWS instead of the normal 8 ACEs.

But it seems the 2 HWS can perform the function of 4 ACEs AND this..



http://forums.anandtech.com/showpost.php?p=37656793&postcount=204

Zlatan is spot on so far.

If Fiji and 1.2 really do have HSA based virtualized multi tenant HW support, then Virge is totally right... their marketing is asleep at the wheel.

This is really cool stuff and not one major tech publication has gotten info on it from AMD. They should be making a halo around the products. They should be talking up Fiji the way Tesla talks up their products for better or worse

 

[Silverforce11]

If Fiji and 1.2 really do have HSA based virtualized multi tenant HW support, then Virge is totally right... their marketing is asleep at the wheel.

This is really cool stuff and not one major tech publication has gotten info on it from AMD. They should be making a halo around the products. They should be talking up Fiji the way Tesla talks up their products for better or worse

I think only Fiji, newer 1.2 GCN. But yes, their marketing is not taking advantage of this.

Imagine a cloud server that streams gaming services. You can have heaps of gamer streams from 1 Fiji, via parallel execution, no added *latency* (that's the key here, to maintain smooth gaming performance/response) for the games being rendered. It's an INSANE technology and they aren't even making a noise and selling it.

One of the biggest hurdles to cloud gaming is the added latency of the internet + graphics latency, gaming is not as smooth or responsive.

 

[Despoiler]

I think only Fiji, newer 1.2 GCN. But yes, their marketing is not taking advantage of this.

Imagine a cloud server that streams gaming services. You can have heaps of gamer streams from 1 Fiji, via parallel execution, no added *latency* (that's the key here, to maintain smooth gaming performance/response) for the games being rendered. It's an INSANE technology and they aren't even making a noise and selling it.

One of the biggest hurdles to cloud gaming is the added latency of the internet + graphics latency, gaming is not as smooth or responsive.

I'm pretty sure they will do just fine after this tech makes it's rounds in the various upcoming conferences and trade shows. People in this field will already know that SR-IOV will be faster and with low overhead compared to Nvidia's IOMMU pass-through implementation.