HP Z840 Hardware Overview

The HP Z840 was a very powerful workstation for its time when it was first released in 2014 and it still is competitive with new workstations in 2023. It has two sockets for Xeon E5 CPUs, support for up to 2 TB of ECC memory, 3 PCIe x16 slots for GPUs, and a 1125W PSU.

In 2023 the prices for HP Z840 workstations in the used market have decreased significantly and which makes it a great option for starting a home lab. Most people start self hosting with a Raspberry Pi or other small SBCs but I think they would be better off using an old workstation. It will consume a lot more power than small SBCs but ere are some advantages that make it worth the energy costs:

• Experiment with LLMs locally
• Use custom ML models for image recognition in surveillance systems
• Plenty of capacity to virtualize with Proxmox
• Run media servers with transcoding and multiple concurrent users
• Run a high performance TrueNas server with sufficient RAM
• Replace expensive PCs with cheap SBCs connected to powerful VMs for programming, data analytics, or gaming
• Use server grade parts like ECC memory and SAS HDDs

My Hardware

Some of the parts that I am using to support higher performance are:

• 2x Xeon E5-2690v4 CPUs
• 1tb M.2 NVMe drive for booting
• 5 SAS HDDs and 2 SATA SSDs (added 3 bays on top of the 4 included)
• 1 Tesla P40 GPU (24GB VRAM)
• 1 GTX 980Ti GPU (6GB VRAM)

Modifications for GPUs

There are 2 issues with using these GPUs. First, the included power cables are 6-Pin instead of the common 8-Pin PCIe cables that most GPUs use. Second, the Tesla P40 was designed for cooling in a rack mounted server and it does not have any fans.

For the Tesla GPU, I was able to find a vendor who sells them modified with fans instead the standard heat sink. These fans take a lot of space but it may still be possible to use the middle x16 slot without a riser. I may add a NVMe adapter there in the future for 4 addition drives.

The power issues are very confusing if you read information online. Even ChatGPT will not be able to provide the correct information because so much wrong information was provided on various forum posts.

There are plenty of 6-Pin to 8-Pin adapters available but if you add up the wattage there does not seem to be enough to support multiple GPUs. A standard 6-Pin cable only provides 75W and the PCIe lane provides another 75W so you seem to only have 150W per lane while these GPUs require 225W. If you carefully review the documentation from HP you will find that these 6-Pin cable are not normal and they actually provide up to 216W. This means that there is plenty available and there is no need to try to add in more adapters for Molex or Sata power.

Another thing to be cautious of is that the Tesla GPUs must use an EPS 8-Pin cable like a CPU would use instead of the 8-Pin PCIe cable that I used for the GTX 980Ti card. The polarity is reversed between these connector so if you use the wrong one you will destroy the GPU. The connector is designed to not fit into the wrong type people have done it.

The EPS adapters for this are less common and some warn to never use in a GPU but I have had no issues. I did have to trim the clip to make it fit in the narrow space on the Tesla P40. This issue has also led to confusion online which is why I think it's important to share my results.