HPE Hardware Accelerated Graphics for Desktop Virtualization

HPE Hardware Accelerated
Graphics for Desktop
Virtualization
Contents
Purpose of this document.................................................................................................................................................................................................................................................2
Abbreviations and naming conventions ..............................................................................................................................................................................................................2
Concepts and technology .................................................................................................................................................................................................................................................2
Bare Metal OS ........................................................................................................................................................................................................................................................................2
GPU pass-through .............................................................................................................................................................................................................................................................3
Software-virtualized GPU............................................................................................................................................................................................................................................4
Graphics-accelerated desktop sessions and application virtualization.............................................................................................................................5
Hardware-virtualized GPU .........................................................................................................................................................................................................................................6
NVIDIA GRID vGPU ..........................................................................................................................................................................................................................................................7
AMD MxGPU ..........................................................................................................................................................................................................................................................................8
Planning considerations for implementing hardware-accelerated desktop virtualization technologies .............................................. 10
Determining the right GPU and platform for your use case ................................................................................................................................................... 10
Virtualization solution feature comparison and considerations .......................................................................................................................................... 11
The HPE ProLiant WS460c Gen9 Graphics Server Blade .............................................................................................................................................................. 13
The HPE Multi-GPU Carrier Card .................................................................................................................................................................................................................... 16
HPE ProLiant WS460c Graphics Server Blade supported GPUs, specification, and configuration options ............................... 16
WS460c Graphics Server Blade supported operating systems matrix................................................................................................................................ 18
Technical white paper
Technical white paper
Page 2
Purpose of this document
• Overview of concepts and hardware-accelerated virtual desktop technologies.
• Overview of features and options for the HPE ProLiant WS460c Graphics Server Blade.
Abbreviations and naming conventions
Table 1. Abbreviation and naming conventions
Convention
Definition
Bare Metal OS
Operating system installed directly on the system, not virtualized
GPU
Graphical processing unit (graphics card). It is important to note that some graphics cards have more than one GPU processor
GPU compute
Synonymous to GPGPU. See GPGPU
GPGPU
General-purpose graphics processing unit—GPU technology that performs application computation traditionally handled by the CPU
GRID vGPU
NVIDIA® technology for sharing true virtual GPU (GRID vGPU) hardware acceleration between multiple users
HDX
Citrix® set of advanced desktop remoting technologies to deliver a high definition experience
HDX 3D Pro
Feature of XenDesktop® HDX protocol for delivering high-end 3D professional graphics
Hypervisor
Virtualization host platform (VMware vSphere® Hypervisor, Microsoft® Hyper-V, Citrix XenServer)
MxGPU
AMD Multiuser GPU (MxGPU) technology
PCoIP
Teradici remote desktop protocol used in VMware View®
RDP
Microsoft Remote Desktop Protocol
RemoteFX
Microsoft’s set of advanced desktop remoting technologies
RFX
Microsoft RemoteFX
RGS
HP Remote Graphics software
SR-IOV
Single Root I/O Virtualization
VDI
Virtual desktop infrastructure
vGPU
NVIDIA GRID vGPU: hardware-virtualized GPU solution used on VMware vSphere® and Citrix XenServer
vSGA
VMware®-specific terminology: software-virtualized GPU (API capture model)
vDGA
VMware-specific terminology for GPU pass-through
VM
Virtual machine
vRAM
GPU video RAM
Concepts and technology
This section contains a conceptual overview of the technologies behind Hardware Accelerated Graphics for Desktop Virtualization. This is a
high-level discussion on the differences between the technologies as well as how the major desktop virtualization providers implement these
technologies into their products.
Bare Metal OS
This method is the classic workstation and PC blade remoting architecture (see figure 1). The client OS is installed directly on the blade or server
hardware and no virtualization is used. End users connect to the workstation using remote protocols such as HP RGS, Microsoft RDP, and Citrix
HDX 3D Pro from client hardware. This method is still used today for users that demand the power and the performance of dedicated hardware.
Technical white paper
Page 3
Figure 1. Bare Metal GPU model (example)
GPU pass-through
It is also referred generically as “direct-attached GPU,” or vendor-specific “vDGA” (VMware) and “GPU pass-through” (Citrix). This method allows
PCI GPU devices to be directly mapped to a virtual machine for dedicated one-to-one use by the VM (see figure 2). The virtual machine has full
and direct access to GPU, including the native graphics driver, allowing for full workstation-class graphics and GPU computation functionality in a
virtual machine. Typically intended for high-end 3D and GPU computational users, the GPU device is directly owned and managed by the
VM operating systems just as in a desktop workstation. The GPU driver is loaded within the virtual machine.
Figure 2. Pass-through GPU model
Technical white paper
Page 4
Enterprise hypervisors using this technology include the following:
• Microsoft Remote Desktop Session Host
• Citrix XenServer 6.0 and newer
• VMware vSphere 5.5 and newer
Advantages
• Up to six workstation-class VMs per host using the HPE WS460c Gen9 Graphics Server Blade with HPE Multi-GPU Carrier Card
• Support for all 3D technologies including DirectX, OpenGL, OpenCL, and NVIDIA CUDA via the native NVIDIA driver in the VM
• Best performing solution as the graphics driver resides in the VM. Virtual machines have full and direct access to a dedicated GPU
(not shared)
• Can mix accelerated and non-accelerated VMs on the same host to maximize resource utilization
Disadvantages
• Higher cost of ownership per connection as it has a dedicated GPU per virtual machine
• Lower VM density per host when compared to other virtualized GPU solutions
• Live migration of VM with pass-through devices is not supported
Software-virtualized GPU
It is also referred to generically as “shared GPU,” “API intercept model,” or vendor-specific of “vSGA” (VMware), and “RemoteFX vGPU” with
Microsoft Hyper-V RemoteFX. This method uses an API intercept model where the GPU is owned and managed by the hypervisor. All incoming
graphics API requests from the VMs are intercepted via the API capture driver in the VM and redirected to and executed by the hypervisor and
then sent back to VM (see figure 3). The VM does not have direct access to the GPU, and the GPU driver is loaded within the hypervisor. This
solution is primarily a 3D offload solution to save CPU cycles and increase host performance, but is not a high-end 3D rendering solution.
Figure 3. Software-virtualized GPU model
Technical white paper
Page 5
Enterprise hypervisors and servers using this technology include:
• Microsoft Hyper-V RemoteFX
• VMware vSGA
Advantages
• Scalability to over 50+ users per GPU depending on the workload.
• It can load balance between multiple installed cards as VMs start.
• Some solutions such as VMware vSGA can dynamically switch between GPU offload and full software rendering (CPU).
• Lower cost per user compared to other technologies, with high density of VMs per GPU.
• Allows each user to have power user performance with enhanced support for DirectX 3D and Windows® Aero.
• Live migration of VMs with vSGA supported.
• Up to 20 percent+ CPU utilization drop.
Disadvantages
• May exhibit unacceptable performance for mid- to high-end 3D knowledge or workstation user workloads.
• This solution is primarily a 3D offload solution to save CPU cycles and increase host performance, but is not a high-end 3D rendering solution.
• Potential application compatibility issues due to limitations of 3D the APIs supported:
– Very limited OpenGL support
– DirectX supported versions limited to DirectX 9 in some cases
Graphics-accelerated desktop sessions and application virtualization
A type of shared GPU virtualization: In this model, 3D applications are installed on the host system or VM and published as hosted, shared
desktops or as a hosted, published application supporting large number of sessions per host. If the application or session is running on a host
equipped with a supported 3D graphics card, each hosted application or session can utilize the graphics card for 3D rendering. Figure 4 shows
the conceptual structure of accelerated desktop and application publishing using Citrix XenDesktop or XenApp.
Figure 4. Graphics-accelerated session and application virtualization
Technical white paper
Page 6
Enterprise solutions that support 3D hosted applications and desktops:
• Citrix XenApp
• Citrix XenDesktop 7.x
Advantages
• GPU sharing with direct access to the video driver maximizing user density and cost.
• Ability to publish hosted applications with OpenGL and DirectX 3D support when a full desktop is not required.
• Lower cost of ownership when compared to other accelerated graphics for VDI technologies.
• Application virtualization is supported on both Bare Metal and GPU pass-through (for greater density). Both VMware vDGA and XenServer
pass-through are supported.
Disadvantages
• Some 3D applications may not work or be certified as published application or multi-user published desktops.
• It may exhibit unacceptable performance for high-end 3D user.
• The GPU can become a performance bottleneck as many users draw on the resources of one card; it is possible that one user can consume all
of the resources of the GPU.
Hardware-virtualized GPU
Hardware-virtualized GPU technology is conceptually a hybrid of software-virtualized GPU and pass-through GPU models. It offers the benefit of
GPU sharing (similar to software-virtualized GPU model) as well as gives the performance, functionality, and features of native high-end graphics
as it has direct access to GPU functionality (similar to pass-through model). Each virtual GPU has a set amount of video RAM (frame buffer) and
the number of virtual GPUs per physical GPU is determined by total video memory on the physical GPU. For example, if we have a physical GPU
that has 8 GB of vRAM, the vRAM can be divided into two virtual GPUs with 4 GB vRAM, or four virtual GPUs with 2 GB vRAM, etc. Depending
on solution, the virtual GPUs are managed by either a software manager that is installed in the hypervisor (NVIDIA GRID vGPU) or managed by
logic embedded on GPU hardware (AMD MxGPU). These solutions are discussed in more depth in next section.
Hardware virtualized GPU advantages
• Increased number of true hardware graphics-accelerated VMs per host, supporting up to 16 users per physical GPU.
• Support for all 3D technologies including DirectX 9/10/11, and OpenGL 4.4, OpenCL, CUDA via the native GPU driver in the VM.
• The virtual machine has full and direct access to the GPU, including the native graphics driver, for full workstation performance.
Hardware virtualized GPU advantages/considerations
• Hot VM migration not supported at this time (e.g., VMware vSphere® vMotion®, XenMotion).
• With NVIDIA vGPU, vRAM amount is guaranteed. The GPU resource, however, is shared among all VMs configured to use it, allowing any
VM to have up to 100 percent of the GPU resource if no other VMs are using the resource, but VMs running heavy workload can take
performance from other GPUs.
• With MxGPU, there is some overhead associated with MxGPU/SR-IOV and vRAM availability is slightly reduced. For example, if you create two
VFs on an 8 GB physical MxGPU-enabled GPU, the VFs created will have slightly less than 4 GB of vRAM.
Technical white paper
Page 7
NVIDIA GRID vGPU
NVIDIA GRID vGPU is a software-defined solution where the physical GPU is broken into multiple virtual GPUs managed by the NVIDIA GRID
vGPU Manager that resides in the hypervisor that resides in the hypervisor, see figure 5a. All vGPUs receive a dedicated amount of physical
video memory and share the resource of GPU. The amount of video memory and displays supported is defined by GRID vGPU types or profiles.
vGPU types are defined by: video memory size, maximum supported display heads, and maximum resolution supported. For NVIDIA GRID
vGPU v3.1, the list in table 2 shows a few of the available predefined vGPU types/profiles for the NVIDIA Tesla M6. Once installed, vGPU types
can easily be added by editing the VM setting and choosing the desired vGPU type.
Figure 5a. NVIDIA GRID vGPU hardware-virtualized GPU model
The maximum number of vGPUs supported depends on the vGPU type selected. For example, if you have an NVIDIA Tesla M60 that has two
physical GPUs on a single card with 8 GB of vRAM on each, then one GPU could have 2 to 4 GB vGPUs and the other could have 8 to 1 GB
vGPUs. Only one vGPU type can run on a single physical GPU at a time. For example, if you only have a single NVIDIA Tesla M60 (one card with
two physical GPUs), once you start a vGPU profile on a physical GPU (8Q for example) no other vGPU profile type can start on that physical
GPU. You can however start a new vGPU profile type on another physical GPU. With NVIDIA vGPU, vRAM amount is guaranteed. The GPU
resource is shared among all VMs configured to use it, allowing any VM to have up to 100 percent of the GPU resource if no other VMs are using
the resource. In general, this works very well as most workloads are sporadic. However, if one or more heavy workloads are running, it may reduce
the performance of other VMs running on the same GPU.
Technical white paper
Page 8
GRID vGPU key benefits
• Up to 16 vGPUs/users/VMs per GPU with direct access to high performance GPU.
• Support for provisioning technologies such as VMware-linked clone pools.
• Ease of configuration, vGPU types appear as VM configuration option.
• vGPU type changes without host system reboot.
Table 2. Example of NVIDIA GRID vGPU v3.1 M6 types
Type
Use case
Frame buffer
Display heads
Max. resolution
Use/GPU
M6-8Q
Designer
8192
4
4096x2160
1
M6-4Q
Designer
4096
4
4096x2160
2
M6-2Q
Designer
2048
4
4096x2160
4
M6-1B
Power user
1024
4
2560x1600
8
M6-8A
Virtual application
8192
1
1280x1024
1
Enterprise hypervisors using hardware-virtualized GPU technology support by HPE ProLiant WS460c include:
• NVIDIA vGPU
– Citrix XenServer 6.2 and newer
– Citrix XenServer 7.0 and newer
– VMware vSphere 6.0 and newer
AMD MxGPU
AMD MxGPU hardware-virtualized GPU solution is based on the industry-standard Single Root I/O Virtualization (SR-IOV) specification. It has
been used in network cards for some time. AMD MxGPU is the first truly all-hardware–virtualized GPU. The GPU can be used as a dedicated
resource using GPU pass-through or it can be split into up to 16 virtual functions (VF), see figure 5a. Examples of the VF types that can be
created are shown in table 3. Once the VF’s are defined, the physical GPU presents itself as multiple small GPUs. For example, take a single
physical MxGPU-enabled GPU and create four VFs. After the host is rebooted, the single MxGPU is presented to the host as four separate GPUs.
Once configured, VF’s are connected to the VMs in the same way as pass-through GPUs. The solution offers hardware partitioning of the GPU,
creating virtual functions with hardware enforced memory isolation and deterministic performance. If a GPU is configured into four VFs, each
VF/VM/user will get 25 percent on the GPU resource, no less, no more, even if other VMs are not using the resource. This allows for a quality of
service that is both measurable and predicable. Because MxGPU technology exposes all GPU functionality to VF, there are no API limitations
allowing support for DirectX, OpenGL, and OpenCL.
Technical white paper
Page 9
Figure 5b. AMD MxGPU hardware-virtualized GPU model
Table 3. Example of AMD FirePro S7100X MxGPU VF definitions
# of VFs per GPU
Frame buffer
2
3840
4
1920
8
960
16
480
Technical white paper
Page 10
MxGPU key benefits
• Predictable performance allows for a quality of service that is both measurable and predicable.
• True hardware-based GPU virtualization.
• Full API support including DirectX, OpenGL, and OpenCL.
• VF memory isolation provides security between VFs.
Enterprise hypervisors using hardware-virtualized GPU technology include:
• AMD MxGPU
– VMware vSphere 6.0 and newer
Planning considerations for implementing hardware-accelerated desktop
virtualization technologies
Determining the right GPU and platform for your use case
The following chart (figure 6a and 6b) shows a comparison of GPU acceleration on desktop virtualization technologies, GPU types, and what
industry segment and use case they are best fitted to.
Figure 6a. Virtualized graphics technology use cases and segment positioning
Technical white paper
Page 11
Figure 6b. 3D graphics application alignment
Virtualization solution feature comparison and considerations
Table 4. Solution feature comparison
VMware
vSGA
VMware vDGA
with NVIDIA GPU
VMware vDGA
with AMD GPU
Microsoft
Server 2012 R2
for RemoteFX
Citrix XenServer
GPU pass-through
with NVIDIA
Citrix
XenServer GPU
pass-through
with AMD
VMware or
Citrix
w/ NVIDIA
vGPU
VMware
AMD MxGPU
OpenCL
N/A
2.x1
2.x1
N/A
2.x1
2.x1
2.x1
2.x1
OpenGL
2.x
2.x, 3.x, 4.x
DirectX
9
Displays supported
83
2
2.x, 3.x, 4.x
N/A
9, 10, 11, 12
9, 10, 11
9, 10, 11
83
63
83
2
2
2.x, 3.x, 4.x1
2
2.x, 3.x, 4.x
2.x, 3.x, 4.x1
9, 10, 11, 12
9, 10, 11
9, 10, 11, 12
9, 10, 112
83
63
83
63
2
Only supported using 8 GB profile (M60-8Q, M6-8Q).
Contingent on actual GPU used.
3
GPU and protocol support may vary.
1
2
XenServer considerations
• XenServer with NVIDIA vGPU
– The NVIDIA vGPU Manager must be installed on the hypervisor to use vGPU features.
– WS460c Gen9 Blades require XenServer 6.5 or newer.
– XenMotion, Storage XenMotion, and VM suspend are not supported at this time.
2
2.x, 3.x, 4.x
2
Technical white paper
Page 12
– VMs can be migrated and started on any server with a compatible GPU configuration.
– CUDA and OpenCL are supported with NVIDIA GRID 3.1 and above.
– Support for linked clones.
• XenServer with AMD MxGPU
– XenServer and XenDesktop do not support AMD cards at time of this writing.
• GPU pass-through
– GPU pass-through technology is one-to-one, meaning each GPU is directly attached to a GPU and is not shared between VMs.
– GPU pass-through technology allows for maximum guaranteed GPU performance as each VM has a dedicated GPU attached and it is not
shared in any way.
– GPU to VM density is determined by the number of GPUs. This model uses one GPU per VM, no sharing.
– XenMotion, Storage XenMotion, or VM suspend are not supported at this time.
– VMs can be migrated and started on any server with a compatible GPU configuration.
VMware considerations
• VMware with NVIDIA vGPU
– The NVIDIA vGPU Manager must be installed on the hypervisor to use vGPU features.
– VMware vSphere® ESXi™ or newer is required for latest vGPU features.
– Live VM migration (vMotion) is not supported at this time. Standard migration is supported as long as both hosts have equivalent vGPU
features configured.
– CUDA and OpenCL are supported with NVIDIA GRID 3.1 and above.
– Support for linked clones.
– Variable performance, determined by number of VMs sharing the GPU and the type of workload the user are doing. VMs are sharing the
resource of GPU. One VM running heavy workload may decrease performance on other VMs.
• VMware with AMD MxGPU
– The AMD MxGPU driver must be installed on the hypervisor to use MxGPU features.
– VMware vSphere ESXi 6.x or newer is required for vGPU features.
– Live VM migration (vMotion) is not supported at this time. Standard migration is supported as long as both hosts have equivalent vGPU
features configured.
– CUDA and OpenCL are supported.
– Deterministic predicable performance: The GPU is partitioned and each VF/VM/user gets a specific and consistent performance. If GPU has
four virtual functions defined, each will get 25 percent of GPU resource, no more, no less. This can be good for deployment needing a
guaranteed quality of service. However, one user cannot take advantage of extra GPU resource if it available.
• VMware vDGA
– VMware vDGA pass-through GPU technology allows for maximum graphical performance (workstation grade) as each VM has a dedicated
GPU attached.
– GPU to VM density is determined by the number of GPUs; this model uses one GPU per VM, no sharing.
– CUDA and OpenCL are supported.
Technical white paper
Page 13
• VMware vSGA
– Live migration is supported on virtual machines that have 3D graphics enabled.
– There are significant limitations in API support (DirectX/OpenGL/OpenCL) and overall performance may not be adequate for most
workloads.
Hyper-V RemoteFX considerations
• The GPU has a dedicated amount of video RAM. Virtual machines consume a specific amount of video RAM based on the maximum number
of monitors and resolution set for each virtual machine. This will dictate the maximum number of virtual machine per physical GPU.
• RemoteFX uses software-virtualized GPU API intercept technology, allowing multiple virtual machines to use the resources of that GPU. This
model uses the GPU as an offload engine for the software 3D drivers in the VM. It is not directly rendering 3D content of VM, which means
there is significant performance and API (DirectX, OpenGL) support limitations.
• Multiple physical graphics cards enhance performance and scalability; if multiples are installed, Hyper-V will load balance between cards as
virtual machines start up.
The HPE ProLiant WS460c Gen9 Graphics Server Blade
The HPE ProLiant WS460c Graphics Server Blade portfolio (formerly WS460c “Workstation Blade”) has been the cutting edge of remote
workstation (i.e., centralized in a data center) computing and graphics virtualization since its inception. Rather than placing the workstation’s
computing power at the user’s desk, the computing power (in the form of a server blade) is moved to the data center, where systems can be
more easily, securely, and economically managed, shared, virtualized, and accessed from anywhere.
The WS460c Graphics Server Blade is ideal for Bare Metal or virtualized multi-tenancy high-end graphics users. Its features enable users to
complete large 3D model visualizations with uncompromised workstation-class performance. New graphics virtualization technologies such as
vGPU and MxGPU allow for a much broader range of workload types to take advantage of graphics virtualization and increased density in
virtualized environments. The WS460c Graphics Server Blade portfolio of options allow for maximum flexibility from a single ultra-high-end 3D or
GPU compute user, to a few mid- to high-end users, or many lower end users that demand less computational horse power but require full
fidelity graphics-accelerated desktops.
Figure 7a. HPE ProLiant WS460c Gen9 Graphics Server Blade
Technical white paper
Figure 7b. HPE ProLiant WS460c Gen9 Graphics density comparison for competition and other platforms
Figure 8a. WS460c single-wide configuration options
Page 14
Technical white paper
Figure 8b. WS460c Gen9 w/ expansion bay configuration options
Figure 8c. WS460c Gen9 GPU support
Page 15
Technical white paper
Page 16
The HPE Multi-GPU Carrier Card
The HPE Multi-GPU Carrier Card (see figure 9) for the WS460c Gen8 and Gen9 Graphics Server Blades is the industry’s first MXM (small form
factor GPU card) carrier card technology with four MXM slots. Providing high-density, high-end 3D graphics for GPU-accelerated desktop
virtualization with support for up to six GPUs in an HPE ProLiant WS460c Gen9 Blade form factor to support all accelerated graphics for desktop
virtualization technologies.
Figure 9. HPE Multi-GPU Carrier Card
HPE ProLiant WS460c Graphics Server Blade supported GPUs, specification, and configuration options
HPE Multi-GPU Carrier Card options and configurations
Table 5. HPE Multi-GPU Carrier Card options (up to two Multi-GPU cards)
HPE Multi-GPU with 3 NVIDIA
Quadro K3100M
HPE Multi-GPU with
2 NVIDIA Tesla M6
HPE Multi-GPU with 2 AMD
FirePro S7100X
Max. Multi-GPU cards supported in WS460c
2 (6 GPU)
2 (4 GPU)
2 (4 GPU)
GPU
3 NVIDIA Kepler GPU (K3100M) per
Multi-GPU Card
2 NVIDIA Maxwell GPUs (K5000-class)
per Multi-GPU Card
2 high-end AMD Tonga (FirePro
S7100X) per Multi-GPU Card
CUDA/Stream cores per Multi-GPU Card
768 (192 per GPU)
3,072 (1536 per GPU)
4,096 (2048 per GPU)
Memory size
12 GB GDDR3 (4 GB per GPU)
16 GB GDDR5 (8 GB per GPU)
16 GB GDDR5 (8 GB per GPU)
Max. power
130 W
100 W per card
100 W per card
Form factor (Multi-GPU)
PCIe 3.0 Single Slot
PCIe 3.0 Single Slot
PCIe 3.0 Single Slot
PCIe
x16
x16
x16
PCIe generation
Gen3 (Gen2 compatible)
Gen3 (Gen2 compatible)
Gen3 (Gen2 compatible)
Cooling solution
Passive
Passive
Passive
Max. # of user
1 per GPU
3 per Multi-GPU Card
16 per GPU (vGPU)
32 per Multi-GPU Card
16 per GPU (MxGPU)
32 per Multi-GPU Card
Technical white paper
Page 17
HPE ProLiant WS460c Graphics Server Blade supported NVIDIA GRID/Tesla cards
Table 6. GRID configuration options
NVIDIA GRID K1
NVIDIA GRID K2
NVIDIA Tesla M6
GPU
4 Kepler GPUs (K6000-class)
2 Kepler GPUs (K5000-class) (K5000-class)
Maxwell M5000-class
CUDA cores
768 (192 per GPU)
3072 (1,536 per GPU)
1536
Memory size
16 GB GDDR3 (4 GB per GPU)
8 GB GDDR5 (4 GB per GPU)
8 GB GDDR5
Max. power
130 W
225 W
100 W
Form factor
Dual slot ATX, 10.5"
Dual slot ATX, 10.5"
x16 MXM
Aux. power
6-pin connector
8-pin connector
N/A
PCIe
x16
x16
x16
PCIe generation
Gen3 (Gen2 compatible)
Gen3 (Gen2 compatible)
Gen3 (Gen2 compatible)
Cooling solution
Passive
Passive
Passive
# of users
4–64
2–64
1–16
OpenGL
4.3
4.3
4.4
Microsoft DirectX
9/10/11
9/10/11
9/10/11/12
HPE ProLiant WS460c Graphics Server Blade supported AMD MxGPU cards
Table 7. AMD MxGPU configuration options
AMD FirePro S7100X
Stream processors
2048
Memory size
8 GB GDDR5
Max. power
100 W
PCIe
x16
PCIe generation
Gen3
Cooling solution
Passive
# of users
1–16
OpenCL
2.0
OpenGL
4.2
Microsoft DirectX
11.1
Technical white paper
Page 18
HPE ProLiant WS460c Graphics Server Blade supported Quadro full-size cards
Table 8. Quadro configuration options
NVIDIA M6000
NVIDIA M5000
NVIDIA K6000
NVIDIA K5000
NVIDIA K4000
GPU
Maxwell
Maxwell
Kepler
Kepler
Kepler
CUDA cores
3072
2048
2880
1536
768
Memory size
12 GB GDDR5
8 GB GDDR5
12 GB GDDR5
4 GB GDDR5
3 GB GDDR5
Max. power
250 W
150 W
225 W
122 W
80 W
Form factor
Double-width
Double-width
Double-width
Double-width
Double-width
PCIe
x16
x16
x16
x16
x16
PCIe generation
Gen3
Gen3
Gen3
Gen2
Gen2
Cooling solution
Active
Active
Active
Active
Active
OpenGL
4.4
4.4
4.3
4.3
4.3
Microsoft DirectX
9/10/11/12
9/10/11/12
9/10/11
9/10/11
9/10/11
HPE ProLiant WS460c Graphics Server Blade supported MXM cards
Table 9. MXM configuration options
NVIDIA Tesla M6
AMD FirePro S7100X
AMD FirePro S4000X
NVIDIA Quadro K3100M
CUDA/Stream cores
1536
2048
640
768
Memory size
8 GB GDDR5
8 GB GDDR5
2 GB
4 GB
Max. power
100 W
100 W
45 W
75 W
PCIe
x16 MXM
x16 MXM
x16 MXM
x16 MXM
PCIe generation
3
3
3
2
Cooling solution
Passive
Passive
Passive
Passive
OpenGL
4.4
4.4
4.3
4.3
OpenCL
2.0
2.0
2.0
N/A
Microsoft DirectX
9/10/11/12
9/10/11
9/10/11
9/10/11
WS460c Graphics Server Blade supported operating systems matrix
The following information complements the Server Blade documentation with supported operating systems and the hypervisor specific to the
WS460c Graphics Server Blade being used for Hardware Accelerated Graphics for Desktop Virtualization.
Technical white paper
Page 19
Table 10. WS460c-supported operating systems
Bare Metal
Windows 7
32 bit
Bare Metal
Windows 10
64 bit
Bare Metal
Windows 7
64 bit
Bare Metal
VMware
Windows 8.1 vSphere
6.x
64 bit
Citrix
XenServer
6.x
Citrix
XenServer
7.x
Windows
Server
2008 R2
Windows
Server
2012 R2
Red Hat®
Enterprise
Linux®
Workstation
6.x/7.x
WS460c Gen9 with
Single MXM GPU
No
Coming soon
Yes
Yes
Yes
Yes
Yes
No
Yes
Yes
WS460c Gen9 with
Graphics Expansion4
No
Coming soon
Yes
Yes
Yes
No
Yes
Yes
Yes
Yes
WS460c Gen9 with
Multi-GPU Carrier
No
No
No
No
Yes
Yes
Yes
No
Yes
No
WS460c Gen8 MXM
GPU
No
No
Yes
No
Yes
Yes
Yes
Yes
No
Yes
WS460c Gen8 with
Graphics Expansion4
No
No
Yes
No
Yes
Yes
Yes
Yes
No
Yes
WS460c Gen8 with
Multi-GPU Carrier
No
No
No
No
Yes
No
Yes
No
No
No
Table 11. WS460c Gen9 maximum supported graphics cards per platform and configuration
Hardware
Bare Metal
Microsoft
2008 R2
Microsoft
2012 R2
VMware
vDGA
VMware
vSGA
VMware
w/ NVIDIA
vGPU
VMware
w/ AMD
MxGPU
XenServer GPU
pass-through
AMD FirePro S7100X
Max. 1
No
Max. 45, 6
Max. 45, 6
No
No
Max. 45, 6
No
No
NVIDIA Tesla M6
Max. 1
No
Max. 45, 6
Max. 45, 6
Max. 45, 6
Max. 45, 6
No
Max. 45, 6
Max. 45, 6
NV Quadro M5000
Max. 15
No
Max. 15
Max. 15
No
No
No
Max. 25
No
NV Quadro M6000
Max. 1
No
Max. 1
Max. 1
No
No
No
5
Max. 1
No
NV Quadro K3100M
Max. 1
No
Max. 65, 6
Max. 65, 6
No
No
No
Max. 66
No
NV GRID K1
No
No
Max. 15
Max. 15
Max. 15
Max. 15
No
Max. 15
No
NV GRID K2
No
5
5
5
5
vGPU/
MxGPU
No
Max. 1
Max. 1
Max. 1
Max. 1
No
Max. 1
No
NV Quadro K4000
5
Max. 2
No
Max. 2
Max. 2
No
No
No
5
Max. 2
No
NV Quadro K5000
Max. 15
No
Max. 15
Max. 15
No5
No
No
Max. 15
No
NV Quadro K6000
Max. 1
No
Max. 1
Max. 1
No
No
No
Max. 1
No
NV Tesla K40
No
No
No
No
No
No
No
No
No
NV Tesla K20
No
No
No
No
No
No
No
No
No
NV Quadro 4000
No
No
No
No
No
No
No
No
No
NV Quadro 5000
No
No
No
No
No
No
No
No
No
NV Quadro 6000
No
No
No
No
No
No
No
No
No
NV Quadro 3000M
No
No
No
No
No
No
No
No
No
NV Quadro 1000M
No
No
No
No
No
No
No
No
No
NV Quadro 500M
No
No
No
No
No
No
No
No
No
Teradici APEX MXM
No
No
No
No
No
No
No
No
No
NV Tesla M2070Q
No
No
No
No
No
No
No
No
No
AMD FirePro S4000X MXM
No
No
No
No
No
No
No
No
No
5
5
5
5
5
5
5
5
5
HPE WS460c with Graphics Expansion allows for all full-size cards and HPE Multi-GPU Carrier Cards to be installed.
Requires Graphics Expansion Blade.
6
Requires HPE Multi-GPU Carrier Card.
4
5
5
5
Technical white paper
Table 12. WS460c Gen8 maximum supported graphics cards per platform and configuration
Hardware
Bare Metal
Microsoft
2008 R2
Microsoft
2012 R2
VMware
vDGA
VMware
vSGA
VMware
NVIDIA
vGPU
VMware
AMD
MxGPU
XenServer GPU
pass-through
XenServer
NVIDIA
vGPU
AMD FirePro S7100X
No
No
No
No
No
No
No
No
No
NVIDIA Tesla M6
No
No
No
No
No
No
No
No
No
NV Quadro K3100M
Max. 1
Max. 1
Max. 67, 8
Max. 67, 8
No
No
No
Max. 67, 8
No
NV GRID K1
No
Max. 17
Max. 17
Max. 17
Max. 17
Max. 17
No
Max. 17
No
NV GRID K2
No
Max. 1
Max. 1
Max. 1
Max. 1
Max. 1
No
7
Max. 1
Max. 17
NV Quadro K4000
Max. 27
Max. 27
Max. 27
Max. 27
No
No
No
Max. 17
Max. 17
NV Quadro K5000
Max. 17
Max. 17
Max. 17
Max. 17
No
No
No
Max. 17
No
NV Quadro K6000
Max. 1
Max. 1
Max. 1
Max. 1
No
No
No
Max. 1
No
NV Tesla K20
Max. 1
No
No
7
Max. 1
No
No
No
7
Max. 1
No
NV Quadro 5000
Max. 17
Max. 17
Max. 17
Max. 17
No
No
No
Max. 17
No
NV Quadro 6000
Max. 1
Max. 1
Max. 1
Max. 1
No
No
No
Max. 1
No
NV Quadro 3000M
Max. 1
Max. 1
Max. 6
Max. 6
No
No
No
Max. 6
No
NV Quadro 1000M
Max. 2
Max. 2
Max. 8
Max. 8
No
No
No
Max. 8
No
NV Quadro 500M
Max. 2
No
No
No
No
No
No
No
No
Teradici APEX MXM
No
No
No
Max. 1
Max. 1
No
No
No
No
NV Tesla M2070Q
No
No
No
No
No
No
No
No
No
AMD FirePro S4000X
MXM
Max. 2
No
No
No
No
No
No
No
No
7
8
7
7
7
7
7
7
7
7
7
7, 8
7, 8
7
7
7
7, 8
7, 8
7
7
7
7
7, 8
7, 8
Requires Graphics Expansion Blade.
Requires HPE Multi-GPU Carrier Card.
Learn more at
hpe.com/us/en/integrated-systems/bladesystem.html
Sign up for updates
© Copyright 2012–2016 Hewlett Packard Enterprise Development LP. The information contained herein is subject to change without
notice. The only warranties for Hewlett Packard Enterprise products and services are set forth in the express warranty statements
accompanying such products and services. Nothing herein should be construed as constituting an additional warranty. Hewlett Packard
Enterprise shall not be liable for technical or editorial errors or omissions contained herein.
AMD is a trademark of Advanced Micro Devices, Inc. Microsoft, Windows, and Windows Server are either registered trademarks or
trademarks of Microsoft Corporation in the United States and/or other countries. Citrix and XenDesktop are registered trademarks of
Citrix Systems, Inc. and/or one more of its subsidiaries and may be registered in the United States Patent and Trademark Office and in
other countries. VMware, VMware View, VMware vSphere, VMware vSphere vMotion, VMware vSphere Hypervisor, and VMware vSphere
ESXi are registered trademarks or trademarks of VMware, Inc. in the United States and/or other jurisdictions. NVIDIA is a trademark
and/or registered trademark of NVIDIA Corporation in the U.S. and other countries. Red Hat is a registered trademark of Red Hat, Inc. in
the United States and other countries. Linux is the registered trademark of Linus Torvalds in the U.S. and other countries. Intel Xeon is a
trademark of Intel Corporation in the U.S. and other countries. Adobe is a trademark of Adobe Systems Incorporated.
4AA4-1701ENW, August 2016, Rev. 7