Recently I connected an NVIDIA RTX4500 Blackwell Server Edition Datacenter GPU on the Minisforum MS-A2 mini computer (aka homelab server) running VMware vSphere and NVIDIA vGPU.
As a proof of concept, I wanted to share how I did this, what parts I used, and share my success running NVIDIA vGPU.
This PoC can be used to run VDI workloads (example Omnissa Horizon), and/or AI models.
Minisforum MS-A2
As an introduction, the Minisforum MS-A2 is a mini-computer that packs a punch. The ones I purchased have the following specs:
AMD Ryzen 9 9955HX Processor
128GB of RAM
M.2/U.2 NVME Support (plenty of M.2 slots)
Dual 10Gbps SFP+ NIC (supports VMware vSphere)
Built-in PCIe x16 slot that supports bifurcation
Numerous other goodies that I’m not listing
These have performed well in my homelab, and run VMware vSphere reliably.
Note that my preference for the MS-A2 systems is because they are running the AMD processor which handles VMware ESX without modification or disablement of cores which is required on some Intel newer generation systems.
Installing VMware ESX was simple.
In my systems, I installed a dual OCuLink PCIe adapter, and enabled bifurcation in the BIOS. This allows us to connect the MS-A2 to external devices using an OCuLink cable.
Dual OCuLink PCIe Adapter Card
The card installs nicely, see below for images of the card installed internally, and ports on the rear of the Minisforum MS-A2.
As you can see, it’s a nice clean install.
Minisforum DEG1 OCuLink eGPU Dock
On the other end of the OCuLink cable, I connect the Minisforum DEG1 eGPU dock. Below is a picture of the unit.
You can use an ATX Power supply to power both the DEG1 dock, along with your preferred GPU.
NVIDIA RTX 4500 Blackwell Server Edition Datacenter GPU
In this specific test, I decided to use the new NVIDIA RTX Pro 4500 Blackwell Server Edition Datacenter GPU. This card is built off the NVIDIA Blackwell GPU architecture, has 32GB of GDDR7 memory, consumes 165W, and has a PCIe 5 x16 interface.
Please note:
Due to OCuLink being used in this PoC, only 4 lanes will be supplied to the GPU using this configuration.
I used a 3D printed air baffle and fans I purchased on eBay to cool the car (as the card is passively cooled).
Installing the card was fairly easy:
Again, pretty clean and nice looking install.
Moving the DEG1 over to my MS-A2 in my homelab, in excitement I quickly connected it.
Minisforum MS-A2 connected to DEG1 OCuLink Dock, with NVIDIA RTX 4500 Blackwell Datacenter GPU
After powering it on and waiting for VMware vSphere ESX to boot, I checked the Graphics Devices section in on the ESX host (I already had the latest NVIDIA vGPU 20.1 drivers in the ESX lifecycle image, so the drivers were present on the host).
Confirming that the card was detected and host vGPU drivers loaded, I proceeded to attach a vGPU profile (4Q) to a test VM, disable the vGPU framerate limiter, and powered on the VM. I then enabled SSH, connected to the host via SSH, and ran “nvidia-smi” to query the GPU status.
Running “nvidia-smi” using the RTX 4500 Blackwell SE GPU
Testing
I went ahead and used Omnissa Horizon to connect to the VDI VM using Blast. This resulted in a high quality hardware encoded/decoded session using NVENC AV1 as the automatically negotiated encoder.
While running through some normal applications, the temperature fluctuated on the card between 33-38C. Note that the fans I’m using are 12V fans, however they were underpowered during this testing and running at 5V.
Loading up Heaven Benchmark, running a test at 1080p using high setting, I was able to achieve 699fps continuously. During this period, the card showed “GPU-Util” in the 90 percentile, and temperatures raised to 80-83C.
Overall, for a quick proof of concept, this was a great test and I’m super happy all the components worked out together.
For over a year and a half I have been working on building a custom NVMe Storage Server for my homelab. I wanted to build a high speed storage system similar to a NAS or SAN, backed with NVMe drives that provides iSCSI, NFS, and SMB Windows File Shares to my network.
The computers accessing the NVMe Storage Server would include VMware ESXi hosts, Raspberry Pi SBCs, and of course Windows Computers and Workstations.
The focus of this project is on high throughput (in the GB/sec) and IOPS.
The current plan for the storage environment is for video editing, as well as VDI VM storage. This can and will change as the project progresses.
The History
More and more businesses are using all-flash NVMe and SSD based storage systems, so I figured there’s no reason why I can’t have build and have my own budget custom all NVMe flash NAS.
This is the story of how I built my own NVMe based Storage Server.
The first version of the NVMe Storage Server consisted of the IO-PEX40152 card with 4 x 2TB Sabrent Rocket 4 NVMe drives inside of an HPE Proliant DL360p Gen8 Server. The server was running ESXi with TrueNAS virtualized, and the PCIe card passed through to the TrueNAS VM.
The results were great, the performance was amazing, and both servers had access to the NFS export via 2 x 10Gb SFP+ networking.
There were three main problems with this setup:
Virtualized – Once a month I had an ESXi PSOD. This was either due to overheating of the IO-PEX40152 card because of modifications I made, or bugs with the DL360p servers and PCIe passthrough.
NFS instead of iSCSI – Because TrueNAS was virtualized inside of the host that was using it for storage, I had to use NFS since the host virtualizing TrueNAS would also be accessing the data on the TrueNAS VM. When shutting down the host, you need to shut down TrueNAS first. NFS disconnects are handled way healthier than iSCSI disconnects (which can cause corruption even if no files are being used).
CPU Cores maxed on data transfer – When doing initial testing, I was maxing out the CPU cores assigned to the TrueNAS VM because the data transfers were so high. I needed a CPU and setup that was better fit.
Version 1 went great, but you can see some things needed to be changed. I decided to go with a dedicated server, not virtualize TrueNAS, and go for a newer CPU with a higher Ghz speed.
And so, version 2 was born (built). Keep reading and scrolling for pictures!
The Hardware
On version 2 of the project, the hardware includes:
While the ML310e Gen8 v2 server is a cheap low entry server, it’s been a fantastic team member of my homelab.
HPE Dual 10G Port 560SFP+ adapters can be found brand new in unsealed boxes on eBay at very attractive prices. Using HPE Parts inside of HPE Servers, avoids the fans from spinning up fast.
The ML310e Gen8 v2 has some issues with passing through PCIe cards to ESXi. Works perfect when not passing through.
The new NVMe Storage Server
I decided to repurpose an HPE Proliant ML310e Gen8 v2 Server. This server was originally acting as my Nvidia Grid K1 VDI server, because it supported large PCIe cards. With the addition of my new AMD S7150 x2 hacked in/on to one of my DL360p Gen8’s, I no longer needed the GRID card in this server and decided to repurpose it.
HPe ML310e Gen8 v2 with NVMe Storage
I installed the IOCREST IO-PEX40152 card in to the PCIe 16x slot, with 4 x 2TB Sabrent Rocket 4 NVME drives.
IOCREST IO-PEX40152 with GLOTRENDS M.2 NVMe SSD Heatsink on Sabrent Rocket 4 NVME
While the server has a PCIe 16x wide slot, it only has an 8x bus going to the slot. This means we will have half the capable speed vs the true 16x slot. This however does not pose a problem because we’ll be maxing out the 10Gb NICs long before we max out the 8x bus speed.
HPE ML310e Gen8 v2 with IOCREST IO-PEX40152HPE ML310e Gen8 v2 with IOCREST IO-PEX40152
I also installed an HPE Dual Port 560SFP+ NIC in to the second slot. This will allow a total of 2 x 10Gb network connections from the server to the Ubiquiti UniFi US-16-XG 10Gb network switch, the backbone of my network.
HPE ML310e Gen8 v2 with HPE 560SFP+ and 10Gig DACHPE ML310e Gen8 v2 with HPE 560SFP+ and 10Gig DAC
The Server also have 4 x Hot Swappable HD bays on the front. When configured in HBA mode (via the BIOS), these are accessible by TrueNAS and can be used. I plan on populating these with 4 x 4TB HPE MDL SATA Hot Swappable drives to act as a replication destination for the NVMe pool and/or slower magnetic long-term storage.
HPE ML310e Gen8 v2 with Hotswap Drive bays
I may also try to give WD RED Pro drives a try, but I’m not sure if they will cause the fans to speed up on the server.
TrueNAS Installation and Configuration
For the initial Proof-Of-Concept for version 2, I decided to be quick and dirty and install it to a USB stick. I also waited until I installed TrueNAS on to the USB stick and completed basic configuration before installing the Quad NVMe PCIe card and 10Gb NIC. I’m using a USB 3.0 port on the back of the server for speed, as I can’t verify if the port on the motherboard is USB 2 or USB 3.
TrueNAS USB Stick on HPE ML310e Gen8 v2
TrueNAS installation worked without any problems whatsoever on the ML310e. I configured the basic IP, time, accounts, and other generic settings. I then proceeded to install the PCIe cards (storage and networking).
TrueNAS Installed on NVMe Storage Server
All NVMe drives were recognized, along with the 2 HDDs I had in the front Hot-swap bays (sitting on an HP B120i Controller configured in HBA mode).
TrueNAS NVMe Disks
The 560SFP+ NIC also was detected without any issues and available to configure.
I created a striped pool (no redundancy) of all 4 x 2TB NVMe drives. This gave us around 8TB of usable high speed NVMe storage. I also created some datasets and a zVOL for iSCSI.
NVMe TrueNAS Storage Pool with Datasets and zVol
I chose to go with the defaults for compression to start with. I will be testing throughput and achievable speeds in the future. You should always test this in every and all custom environments as the results will always vary.
Network Configuration
Initial configuration was done via the 1Gb NIC connection to my main LAN network. I had to change this as the 10Gb NIC will be directly connected to the network backbone and needs to access the LAN and Storage VLANs.
I went ahead and configured a VLAN Interface on VLAN 220 for the Storage network. Connections for iSCSI and NFS will be made on this network as all my ESXi servers have vmknics configured on this VLAN for storage. I also made sure to configure an MTU of 9000 for jumbo frames (packets) to increase performance. Remember that all hosts must have the same MTU to communicate.
10Gb NIC on Storage VLAN
Next up, I had to create another VLAN interface for the LAN network. This would be used for management, as well as to provide Windows File Share (SMB/Samba) access to the workstations on the network. We leave the MTU on this adapter as 1500 since that’s what my LAN network is using.
10Gb NIC on LAN VLAN
As a note, I had to delete the configuration for the existing management settings (don’t worry, it doesn’t take effect until you hit test) and configure the VLAN interface for my LANs VLAN and IP. I tested the settings, confirmed it was good, and it was all setup.
At this point, only the 10Gb NIC is now being used so I went ahead and disconnected the 1Gb network cable.
Sharing Setup and Configuration
It’s now time to configure the sharing protocols that will be used. As mentioned before, I plan on deploying iSCSI, NFS, and Windows File Shares (SMB/Samba).
iSCSI and NFS Configuration
Normally, for a VMware ESXi virtualization environment, I would always usually prefer iSCSI based storage, however I also wanted to configure NFS to test throughput of both with NVMe flash storage.
Earlier, I created the datasets for all my my NFS exports and a zVOL volume for iSCSI.
Note, that in order to take advantage of the VMware VAAI storage directives (enhancements), you must use a zVOL to present an iSCSI target to an ESXi host.
For NFS, you can simply create a dataset and then export it.
For iSCSI, you need to create a zVol and then configure the iSCSI Target settings and make it available.
SMB (Windows File Shares)
I needed to create a Windows File Share for file based storage from Windows computers. I plan on using the Windows File Share for high-speed storage of files for video editing.
Using the dataset I created earlier, I configured a Windows Share, user accounts, and tested accessing it. Works perfect!
Connecting the host
Connecting the ESXi hosts to the iSCSI targets and the NFS exports is done in the exact same way that you would with any other storage system, so I won’t be including details on that in this post.
We can clearly see the iSCSI target and NFS exports on the ESXi host.
TrueNAS NVMe iSCSI Target on VMware ESXi Host
NVMe iSCSI and NFS ESXi Datastores
To access Windows File Shares, we log on and map the network share like you would normally with any file server.
Testing
For testing, I moved (using Storage vMotion) my main VDI desktop to the new NVMe based iSCSI Target LUN on the NVMe Storage Server. After testing iSCSI, I then used Storage vMotion again to move it to the NFS datastore. Please see below for the NVMe storage server speed test results.
CrystalDiskMark testing an IOCREST IO-PEX40152 and Sabrent Rocket 4 NVME SSD
CrystalDiskMark testing IOPS on an IOCREST IO-PEX40152 and Sabrent Rocket 4 NVME SSD
Note, that when I performed these tests, my CPU was maxed out and limiting the actual throughput. Even then, these are some fairly impressive speeds. Also, these tests were directly testing each NVMe drive individually.
Moving on to the NVMe Storage Server, I decided to test iSCSI NVMe throughput and NFS NVMe throughput.
I opened up CrystalDiskMark and started a generic test, running a 16GB test file a total of 6 times on my VDI VM sitting on the iSCSI NVMe LUN.
NVMe Storage Server iSCSI Benchmark with CrystalDiskMark
You can see some impressive speeds maxing out the 10Gb NIC with crazy performance of the NVME storage:
1196MB/sec READ
1145.28MB/sec WRITE (Maxing out the 10GB NIC)
62,725.10 IOPS READ
42,203.13 IOPS WRITE
Additionally, here’s a screenshot of the ix0 NIC on the TrueNAS system during the speed test benchmark: 1.12 GiB/s.
TrueNAS NVME Maxing out 10Gig NIC
And remember this is with compression. I’m really excited to see how I can further tweak and optimize this, and also what increases will come with configuring iSCSI MPIO. I’m also going to try to increase the IOPS to get them closer to what each individual NVMe drive can do.
Now on to NFS, the results were horrible when moving the VM to the NFS Export.
NVMe Storage Server NFS Benchmark with CrystalDiskMark
You can see that the read speed was impressive, but the write speed was not. This is partly due to how writes are handled with NFS exports.
Clearly iSCSI is the best performing method for ESXi host connectivity to a TrueNAS based NVMe Storage Server. This works perfect because we’ll get the VAAI features (like being able to reclaim space).
iSCSI MPIO Speed Test
This is more of an update… I was finally able to connect, configure, and utilize the 2nd 10Gbe port on the 560SFP+ NIC. In my setup, both hosts and the TrueNAS storage server all have 2 connections to the switch, with 2 VLANs and 2 subnets dedicated to storage. Check out the before/after speed tests with enabling iSCSI MPIO.
TrueNAS NVME iSCSI MPIO BeforeTrueNAS NVME iSCSI MPIO AfterBefore and After enabling iSCSI MPIO on TrueNAS with NVME Storage
As you can see I was able to essentially double my read speeds (again maxing out the networking layer), however you’ll notice that the write speeds maxed out at 1598MB/sec. I believe we’ve reached a limitation of the CPU, PCIe bus, or something else inside of the server. Note, that this is not a limitation of the Sabrent Rocket 4 NVME drives, or the IOCREST NVME PCIe card.
Moving Forward
I’ve had this configuration running for around a week now with absolutely no issues, no crashes, and it’s been very stable.
Using a VDI VM on NVMe backed storage is lightning fast and I love the experience.
I plan on running like this for a little while to continue to test the stability of the environment before making more changes and expanding the configuration and usage.
Future Plans (and Configuration)
Drive Bays
I plan to populate the 4 hot-swappable drive bays with HPE 4TB MDL drives. Configured with RaidZ1, this should give me around 12TB usable storage. I can use this for file storage, backups, replication, and more.
NVMe Replication
This design was focused on creating non-redundant extremely fast storage. Because I’m limited to a total of 4 NVMe disks in this design, I chose not to use RaidZ and striped the data. If one NVMe drive is lost, all data is lost.
I don’t plan on storing anything important, and at this point the storage is only being used for VDI VMs (which are backed up), and Video editing.
If I can populate the front drive bays, I can replicate the NVMe storage to the traditional HDD storage on a frequent basis to protect against failure to some level or degree.
Version 3 of the NVMe Storage Server
More NVMe and Bigger NVMe – I want more storage! I want to test different levels of RaidZ, and connect to the backbone at even faster speeds.
NVME Drives with PLP (Power Loss Prevention) for data security and protection.
Dual Power Supply
Let me know your thoughts and ideas on this setup!
So you want to add NVMe storage capability to your HPE Proliant DL360p Gen8 (or other Proliant Gen8 server) and don’t know where to start? Well, I was in the same situation until recently. However, after much research, a little bit of spending, I now have 8TB of NVMe storage in my HPE DL360p Gen8 Server thanks to the IOCREST IO-PEX40152.
Unsupported you say? Well, there are some of us who like to live life dangerously, there is also those of us with really cool homelabs. I like to think I’m the latter.
PLEASE NOTE: This is not a supported configuration. You’re doing this at your own risk. Also, note that consumer/prosumer NVME SSDs do not have PLP (Power Loss Prevention) technology. You should always use supported configurations and enterprise grade NVME SSDs in production environments.
Update – May 2nd 2021: Make sure you check out my other post where I install the IOCREST IO-PEX40152 in an HPE ML310e Gen8 v2 server for Version 2 of my NVMe Storage Server.
Update – June 21 2022: I’ve received numerous comments, chats, and questions about whether you can boot your server or computer using this method. Please note that this is all dependent on your server/computer, the BIOS/EFI, and capabilities of the system. In my specific scenario, I did not test booting since I was using the NVME drives purely as additional storage.
DISCLAIMER: If you attempt what I did in this post, you are doing it at your own risk. I won’t be held liable for any damages or issues.
NVMe Storage Server – Use Cases
There’s a number of reasons why you’d want to do this. Some of them include:
Server Storage
VMware Storage
VMware vSAN
Virtualized Storage (SDS as example)
VDI
Flash Cache
Special applications (database, high IO)
Adding NVMe capability
Well, after all that research I mentioned at the beginning of the post, I installed an IOCREST IO-PEX40152 inside of an HPE Proliant DL360p Gen8 to add NVMe capabilities to the server.
IOCREST IO-PEX40152 with 4 x 2TB Sabrent Rocket 4 NVME
At first I was concerned about dimensions as technically the card did fit, but technically it didn’t. I bought it anyways, along with 4 X 2TB Sabrent Rocket 4 NVMe SSDs.
The end result?
HPE DL360p Gen8 with NVME SSD
IMPORTANT: Due to the airflow of the server, I highly recommend disconnecting and removing the fan built in to the IO-PEX40152. The DL360p server will create more than enough airflow and could cause the fan to spin up, generate electricity, and damage the card and NVME SSD.
Also, do not attempt to install the case cover, additional modification is required (see below).
The Fit
Installing the card inside of the PCIe riser was easy, but snug. The metal heatsink actually comes in to contact with the metal on the PCIe riser.
IO-PEX40152 installed on DL360p PCIe Riser
You’ll notice how the card just barely fits inside of the 1U server. Some effort needs to be put in to get it installed properly.
HPE DL360p Gen8 with IO-PEX40152 Installed
There are ribbon cables (and plastic fittings) directly where the end of the card goes, so you need to gently push these down and push cables to the side where there’s a small amount of thin room available.
We can’t put the case back on… Yet!
Unfortunately, just when I thought I was in the clear, I realized the case of the server cannot be installed. The metal bracket and locking mechanism on the case cover needs the space where a portion of the heatsink goes. Attempting to install this will cause it to hit the card.
HPE DL360p Gen8 Case Locking Mechanism
The above photo shows the locking mechanism protruding out of the case cover. This will hit the card (with the IOCREST IO-PEX40152 heatsink installed). If the heatsink is removed, the case might gently touch the card in it’s unlocked and recessed position, but from my measurements clears the card when locked fully and fully closed.
I had to come up with a temporary fix while I figure out what to do. Flip the lid and weight it down.
HPE DL360p Gen8 case cover upside down
For stability and other tests, I simply put the case cover on upside down and weighed it down with weights. Cooling is working great and even under high load I haven’t seen the SSD’s go above 38 Celsius.
The plan moving forward was to remove the IO-PEX40152 heatsink, and install individual heatsinks on the NVME SSD as well as the PEX PCIe switch chip. This should clear up enough room for the case cover to be installed properly.
The fix
I went on to Amazon and purchased the following items:
IOCREST IO-PEX40152 with GLOTRENDS M.2 NVMe SSD Heatsink on Sabrent Rocket 4 NVME
And now we install it in the DL360p Gen8 PCIe riser and install it in to the server.
You’ll notice it’s a nice fit! I had to compress some of the heat conductive goo on the PFX chip heatsink as the heatsink was slightly too high by 1/16th of an inch. After doing this it fit nicely.
Also, note the one of the cable/ribbon connectors by the SAS connections. I re-routed on of the cables between the SAS connectors they could be folded and lay under the card instead of pushing straight up in to the end of the card.
As I mentioned above, the locking mechanism on the case cover may come in to contact with the bottom of the IOCREST card when it’s in the unlocked and recessed position. With this setup, do not unlock the case or open the case when the server is running/plugged in as it may short the board. I have confirmed when it’s closed and locked, it clears the card. To avoid “accidents” I may come up with a non-conductive cover for the chips it hits (to the left of the fan connector on the card in the image).
And with that, we’ve closed the case on this project…
HPE DL360p Gen8 Case Closed
One interesting thing to note is that the NVME SSD are running around 4-6 Celsius cooler post-modification with custom heatsinks than with the stock heatsink. I believe this is due to the awesome airflow achieved in the Proliant DL360 servers.
Conclusion
I’ve been running this configuration for 6 days now stress-testing and it’s been working great. With the server running VMware ESXi 6.5 U3, I am able to passthrough the individual NVME SSD to virtual machines. Best of all, installing this card did not cause the fans to spin up which is often the case when using non-HPE PCIe cards.
This is the perfect mod to add NVME storage to your server, or even try out technology like VMware vSAN. I have a number of cool projects coming up using this that I’m excited to share.
Lately, I’ve been playing with video editing and encoding as a new hobby. It requires a powerful system for the production process for both editing, and encoding. While CPU power isn’t necessarily important, the CPU instruction sets and your GPU play a key part with editing and encoding.
For the last few weeks, I’ve been attempting to use my desktop rig with a couple of Nvidia GeForce cards and I’ve been struggling to be able to edit in real time, as well as encode completed video productions in a reasonable amount of time.
As part of my demo and homelab, I have an HPE ML310e Gen8 v2, with an Nvidia Grid K1 card. This setup is backed with an HPE MSA 2040 SAN for main storage and a Synology DS1813+ for backup and video storage.
It’s a beautiful desktop replacement as it’s silent, provides a desktop backed with a Intel Xeon Processor, and a high performance GPU.
VDI Video Editing and Encoding
For software, I’ve been using Cyberlink’s PowerDirector software. I’ve used this in the past for screen recording, as well as editing videos I use for demos.
On my physical desktop rig, even with two GPUs it struggles to allow me to preview in realtime the edits I’ve done on a project. The preview window is jolty with loss frames, and it’s hard to know what you’re doing. Also, when producing and encoding a finalized video project, it would take forever to complete even a small 5 minute video at 1080p.
When I first loaded this up on my VDI environment, the software instantly detected the Nvidia GRID card, and asked me if it could use it. From that point on the preview window was fluid, transitions and add-ins were rendered on the fly during previewing, and the final production encoding was literally over 20 times faster using 1080p. Keep in mind this VM only has one Nvidia K180q profile attached to it, so I’m only using less than 25% of the cards full capability.
Other benefits to video editing and encoding on VDI
There’s some other benefits that can be realized when doing video editing and encoding inside of a VDI environment:
Ability to connect remotely and work anywhere
Ability to work anywhere with a high performance system
High speed video storage on demand (since it’s all remote)
It can become part of your normal backup solution
This is just another great use case scenario for VDI. Whether it’s for the video professional, or a large organization.
This website uses cookies to improve your experience. We'll assume you're ok with this, but you can opt-out if you wish.
Do you accept the use of cookies and accept our privacy policy? AcceptRejectCookie and Privacy Policy
Privacy & Cookies Policy
Privacy Overview
This website uses cookies to improve your experience while you navigate through the website. Out of these cookies, the cookies that are categorized as necessary are stored on your browser as they are essential for the working of basic functionalities of the website. We also use third-party cookies that help us analyze and understand how you use this website. These cookies will be stored in your browser only with your consent. You also have the option to opt-out of these cookies. But opting out of some of these cookies may have an effect on your browsing experience.
