From 2017 I assembled a TITAN Xp four‑card workstation and have been using it for over 4 years. Recently I performed some upgrades:

• GPU: AORUS GeForce RTX 4090 Xtreme Waterforce 24G
• CPU: AMD Ryzen Threadripper PRO 5975WX
• Motherboard: ASUS Pro WS WRX80E‑SAGE SE WIFI II
• Power supply: ROG Thor 1600W Titanium
• Cooling: upgraded from air cooling to KRAKEN Z73 RGB liquid cooling
• Case: ROG Genesis GR701 E‑ATX case
• Storage: WD BLACK SN850 4TB, Samsung 990 PRO 2TB, Lexar ARES 4TB, Hikvision C4000 4TB

RTX 4090

The RTX 4090 is NVIDIA’s latest consumer‑grade flagship, built on the Ada Lovelace 4 nm process, featuring an AD102 chip with 16 384 CUDA cores, 512 Tensor cores and 128 RT cores, and 24 GB GDDR6X memory (384‑bit bus, 21 Gbps). FP32 and FP16 performance reaches 82.6 TFLOPS, unfortunately it does not support NVLink, and INT8 inference performance hits 661 TOPS. Although the standard power draw is 450 W, the AORUS Waterforce liquid‑cooled version cools well and is quiet.

Installation details

During the build I ran into a few technical details: the case is huge, but the WRX80E motherboard is even larger, so I had to remove a fixed copper post to secure the board. The CPU is square while the water‑block is round, which could affect cooling; testing with cpuburn and gpuburn showed no significant temperature issues. I also kept two TITAN Xp cards as display outputs.

The overall result after installation looks great; the RGB lighting combined with the liquid‑cooling system is very striking. Because the case space is limited, the routing of every cable had to be planned in advance to avoid interfering with other components. Finally I installed Ubuntu 22.04 LTS.

Hyper M.2 x16 expansion card

A major advantage of AMD CPUs is the abundance of PCIe lanes; the Ryzen Threadripper PRO 5975WX provides 128 PCIe lanes (120 available to the user). Even with four RTX 4090 GPUs installed, 56 lanes remain.

ASUS’s WRX80E motherboard offers a Hyper M.2 x16 expansion card, which comes with a large heatsink and an active cooling fan. By creating a RAID 0 array, you can achieve up to 32 GB/s disk throughput. My four M.2 drives:

lsblk -d -o NAME,SIZE,MODEL | grep nvme

# nvme0n1   3.7T HS-SSD-C4000 4096G
# nvme1n1   3.6T WD_BLACK SN850X 4000GB
# nvme2n1   1.8T Samsung SSD 990 PRO 2TB
# nvme3n1   3.7T Lexar SSD ARES 4TB
# nvme4n1 476.9G Samsung SSD 960 PRO 512GB

After inserting the expansion card into the motherboard, set the corresponding PCIe slots to PCIe RAID mode in the BIOS under Onboard Devices Configuration.

btrfs RAID 0

The next step was to build a RAID 0. I first tried the BIOS’s RAIDXpert2 Configuration Utility; for disks of different sizes it only allowed a maximum of 8 TB in RAID 0, and the tool received poor reviews. I then switched to btrfs’s native RAID support:

# Install btrfs tools
sudo apt update
sudo apt install btrfs-progs

# Create RAID 0 directly using btrfs
sudo mkfs.btrfs -d raid0 -m raid1 -L HOME \
  /dev/nvme0n1 /dev/nvme1n1 /dev/nvme2n1 /dev/nvme3n1

# Parameter explanation:
# -d raid0: Use RAID 0 for data
# -m raid1: Use RAID 1 for metadata (safer)

# Mount (only need to mount one device, btrfs will automatically recognize other members)
sudo mkdir -p /home
sudo mount -o noatime,nodiratime,compress=zstd:1,space_cache=v2,\
ssd,discard=async,commit=120 /dev/nvme0n1 /home

# Mount options explanation:
# compress=zstd:1 - Fast compression, improves effective bandwidth
# space_cache=v2 - Improved space cache
# ssd - SSD optimization
# discard=async - Asynchronous TRIM
# commit=120 - Extend commit interval to 120 seconds

btrfs’s dynamic striping mechanism automatically distributes stripes across all available devices, allowing the full 14 TB to be used: the first 8 TB run at four‑disk speed, the remaining 6 TB at three‑disk speed.

# Check device usage
sudo btrfs filesystem show /home

# Label: HOME  uuid: 561ca42e-0811-47f7-900c-d594b5b22033
#         Total devices 4 FS bytes used 144.00KiB
#         devid    1 size 3.73TiB used 1.00GiB path /dev/nvme0n1
#         devid    2 size 3.64TiB used 1.00GiB path /dev/nvme1n1
#         devid    3 size 1.82TiB used 2.01GiB path /dev/nvme2n1
#         devid    4 size 3.73TiB used 2.01GiB path /dev/nvme3n1

sudo btrfs device stats /home

# View detailed space allocation
sudo btrfs filesystem df /home

Add a line to /etc/fstab to auto‑mount the RAID volume:

UUID=$(sudo blkid -s UUID -o value /dev/nvme0n1)
MOUNT_OPTIONS="noatime,nodiratime,compress=zstd:1,space_cache=v2,ssd,discard=async,commit=120"
echo "UUID=$UUID /home btrfs $MOUNT_OPTIONS 0 0" | sudo tee -a /etc/fstab

Using fio to test btrfs RAID 0 performance shows that CoW has a large impact on read/write speed; after disabling CoW with sudo chattr +C the results are: sequential read 28.6 GB/s, sequential write 16.5 GB/s, random read 567 K IOPS, random write 61 K IOPS.

sudo apt install fio

sudo mkdir -p /home/test/
sudo chattr +C /home/test/

sudo fio --name=test --filename=/home/test/file \
  --size=50G --direct=1 --rw=read --bs=1M \
  --iodepth=256 --numjobs=4  --runtime=60 --time_based \
  --group_reporting --ioengine=libaio

sudo fio --name=test --filename=/home/test/file \
  --size=50G --direct=1 --rw=write --bs=1M \
  --iodepth=256 --numjobs=4 --runtime=60 --time_based \
  --group_reporting --ioengine=libaio

sudo fio --name=test --filename=/home/test/file \
  --size=5G --direct=1 --rw=randread --bs=4K \
  --iodepth=256 --numjobs=4  --runtime=60 --time_based \
  --group_reporting --ioengine=libaio

sudo fio --name=test --filename=/home/test/file \
  --size=5G --direct=1 --rw=randwrite --bs=4K \
  --iodepth=256 --numjobs=4  --runtime=60 --time_based \
  --group_reporting --ioengine=libaio