Post

Did I Just Strike Gold Found Two Amfeltec Pcie Carrier Boards With 4X 1Tb Samsung 960 Pros In A 10 Flea Market Junk Box

Did I Just Strike Gold Found Two Amfeltec Pcie Carrier Boards With 4X 1Tb Samsung 960 Pros In A 10 Flea Market Junk Box

Did I Just Strike Gold? Found Two Amfeltec PCIe Carrier Boards With 4X 1TB Samsung 960 Pros In A $10 Flea Market Junk Box

Introduction

The thrill of discovering high-performance hardware in unexpected places is every sysadmin’s hidden fantasy. When a €10 box of “junk” yielded two Amfeltec carrier boards loaded with eight 1TB Samsung 960 Pro NVMe SSDs, it became more than just luck – it became a masterclass in repurposing enterprise-grade hardware for modern DevOps workflows.

For homelab enthusiasts and self-hosted infrastructure architects, such finds represent critical opportunities to:

  • Deploy high-IOPS storage without cloud costs
  • Experiment with PCIe bifurcation and hardware passthrough
  • Build durable storage arrays using MLC NAND – a rarity in today’s TLC/QLC-dominated market

This guide explores how to transform reclaimed enterprise storage into a hyper-converged powerhouse, covering:

  1. Technical analysis of Amfeltec’s PCIe carrier architecture
  2. Maximizing endurance on aged-but-legendary Samsung 960 Pro SSDs
  3. Implementation patterns for Kubernetes persistent storage
  4. Real-world performance tuning for NVMe arrays

Understanding the Technology

Amfeltec PCIe Carrier Boards (SKU-086-34)

These PCIe 3.0 x16 carrier boards implement hardware multiplexing to host four M.2 NVMe SSDs per card. Unlike software-based solutions, Amfeltec’s design:

  • Uses PCIe packet switching to bypass OS-level limitations
  • Requires no special driver support (appears as four discrete NVMe devices)
  • Supports bifurcation-unaware motherboards

Key Specifications:
| Parameter | Value |
|——————–|————————|
| PCIe Version | Gen 3.0 (8GT/s) |
| Host Interface | x16 (split to 4x x4) |
| SSD Support | M.2 22110/2280 NVMe |
| Power Delivery | 12V via PCIe + SATA |

Samsung 960 Pro 1TB NVMe SSDs

These 2016-era flagships remain prized for:

  • MLC NAND: 3,000 P/E cycles (vs. 1,000 in modern TLC)
  • V-NAND Architecture: 48-layer 3D stacking
  • Sustained Performance: 3,500/2,100 MB/s seq. R/W

Endurance Metrics:

1
2
3
4
5
# Check remaining lifespan via SMART  
sudo smartctl -a /dev/nvme0n1 | grep "Percentage Used"  
Percentage Used:         13%  
Data Units Read:         15,246,123  
Data Units Written:      22,561,495  

Prerequisites

Hardware Requirements

  • PCIe 3.0 x16 slot (x8 electrically acceptable with performance penalty)
  • Supplemental SATA power connectors (12V/2A per board minimum)
  • Adequate cooling (4x NVMe SSDs can dissipate 15W+ under load)

Software Requirements

  • Linux 4.15+ kernel (for NVMe-oF support)
  • nvme-cli v1.8+
  • mdadm or ZFS 0.8+ for software RAID

Verification Checklist:

  1. Confirm PCIe lane allocation:
    1
    2
    
    lspci -vv -s $PCI_ADDRESS | grep LnkSta  
    LnkSta: Speed 8GT/s, Width x16  
    
  2. Validate SSD health:
    1
    
    sudo nvme smart-log /dev/nvme0n1  
    

Installation & Configuration

Hardware Initialization

  1. Power the carrier board properly:
    • Connect both PCIe slot and SATA power
    • Verify 12V rail stability:
      1
      2
      3
      
      sudo dmesg | grep "nvme nvme"  
      [    7.123456] nvme nvme0: pci $PCI_ADDRESS  
      [    7.123789] nvme nvme0: 4/0/0 default/read/poll queues  
      
  2. Configure PCIe bifurcation if available (not required but improves performance):
    1
    
    BIOS Settings → PCIe Configuration → x16 Slot → 4x4x4x4  
    

Filesystem Optimization

For mixed read/write workloads typical in DevOps environments:

1
2
3
4
5
# XFS with optimal stripe alignment  
mkfs.xfs -d su=64k,sw=4 -l version=2,su=64k /dev/md0  

# Mount options for NVMe array  
/dev/md0 /mnt/array xfs defaults,noatime,nodiratime,discard,logbufs=8,logbsize=256k 0 2  

ZFS Alternative Configuration:

1
2
3
zpool create -o ashift=12 tank mirror nvme0n1 nvme1n1 mirror nvme2n1 nvme3n1  
zfs set compression=zstd-9 tank  
zfs set atime=off tank  

Performance Tuning

Kernel-Level Optimizations

1
2
3
4
# /etc/sysctl.d/99-nvme-optimize.conf  
vm.dirty_ratio = 10  
vm.dirty_background_ratio = 5  
block/nvme_core.io_timeout=300  

IRQ Balancing

1
2
3
4
# Assign IRQs to specific cores  
for i in $(grep nvme /proc/interrupts | awk '{print $1}' | sed 's/://'); do  
    echo 3 > /proc/irq/$i/smp_affinity_list  
done  

Usage Patterns for DevOps

Kubernetes Persistent Storage

1
2
3
4
5
6
7
8
# storage-class-nvme.yaml  
apiVersion: storage.k8s.io/v1  
kind: StorageClass  
metadata:  
  name: nvme-raid0  
provisioner: kubernetes.io/no-provisioner  
volumeBindingMode: WaitForFirstConsumer  
allowVolumeExpansion: true  

Ansible Provisioning Playbook

1
2
3
4
5
6
7
8
9
10
11
12
- name: Configure NVMe array  
  hosts: storage_nodes  
  tasks:  
    - name: Update initramfs for NVMe  
      command: update-initramfs -u  
      when: ansible_kernel == '5.4.0-162-generic'  

    - name: Enable IOMMU for passthrough  
      lineinfile:  
        path: /etc/default/grub  
        regexp: '^GRUB_CMDLINE_LINUX='  
        line: 'GRUB_CMDLINE_LINUX="intel_iommu=on iommu=pt"'  

Troubleshooting

Common Issues and Solutions

Problem: NVMe devices not detected

1
dmesg | grep -i "Controller reset"  

Solution: Power cycle board with SATA connector attached

Problem: Performance degradation during sustained writes

1
2
iostat -x 1  
# Check %util and await metrics  

Solution: Implement write throttling:

1
echo 100 > /sys/block/nvme0n1/queue/write_back_dev_threshold  

Conclusion

Salvaged enterprise hardware like Amfeltec carrier boards paired with Samsung’s MLC-based NVMe drives offers exceptional value for:

  • Building high-density Kubernetes storage nodes
  • Creating low-latency testbeds for distributed systems
  • Developing cost-effective CI/CD infrastructure

Next Steps:

  1. Benchmark with fio using real DevOps workloads
  2. Implement SMART monitoring with Prometheus NVMe exporter
  3. Explore hardware passthrough to KVM/QEMU VMs

For further reading:

The true value in such finds lies not just in the hardware’s sticker price, but in the engineering education gained through implementing enterprise-grade solutions on reclaimed infrastructure.

This post is licensed under CC BY 4.0 by the author.