# Z840 Power Consumption Study

A before/after power study documenting the Z840's evolution from a dual-socket
Xeon workstation into a dedicated local AI inference box. Measured and documented
2026-05-18 for posterity — this was the early days of running local AI at home
on recycled enterprise iron, before it became normal.

---

## The Machine

HP Z840 Workstation. Bought second-hand. Repurposed as a local AI inference box
running Ollama, Open WebUI, and Hermes monitoring. 128GB ECC DDR4 — retained
because RAM prices made starting fresh impractical.

See [[z840]] for full hardware spec and service documentation.

---

## Phase 1 — Baseline (2026-05-18, pre-upgrade)

**CPUs installed:** 2× Intel Xeon E5-2698 v3 @ 2.30GHz
- 16 cores / 32 threads per socket
- TDP: 135W each — **270W combined TDP**
- Architecture: Haswell-EP (v3), LGA 2011-3

**GPU installed:** NVIDIA Quadro M4000 (original workstation card, placeholder)
- 8GB GDDR5
- Not suitable for modern AI inference

**RAM:** 128GB DDR4 ECC Registered @ 2133 MT/s (8× 16GB DIMMs across both sockets)

### Measured idle power draw (RAPL, system at rest, load avg ~0.00)

| Component | Idle Draw |
|-----------|-----------|
| CPU socket 0 (E5-2698 v3) | 21.3W |
| CPU socket 1 (E5-2698 v3) | 19.8W |
| DRAM | ~0.6W |
| Quadro M4000 (nvidia-smi) | 42.3W |
| **Total measured** | **~84W** |

> Note: RAPL measures CPU package power only. Wall draw will be higher due to
> motherboard, storage, fans, and PSU inefficiency. Estimated wall idle: **150–180W**

### AI inference capability at this stage

- Ollama running but limited by Quadro M4000 (8GB VRAM, no modern tensor cores)
- Models constrained to ~7B parameter quantised
- No tensor parallelism across cards

---

## Phase 2 — CPU Upgrade (arriving 2026-05-19)

**CPUs incoming:** 2× Intel Xeon E5-2620 v4 @ 2.10GHz
- 8 cores / 16 threads per socket
- TDP: 85W each — **170W combined TDP** (vs 270W before — **saving 100W TDP**)
- Architecture: Broadwell-EP (v4), LGA 2011-3
- Cost: £24 for the pair (eBay)

**Why this makes sense:** The GPU does the AI heavy lifting. CPU cores are largely
idle during inference — they're just feeding data and managing I/O. Dropping from
16-core to 8-core per socket costs nothing in real-world AI performance.

### Expected power improvement

| Scenario | Before (v3) | After (v4) | Saving |
|----------|-------------|------------|--------|
| CPU TDP (both sockets) | 270W | 170W | 100W |
| Estimated CPU idle draw | ~41W | ~25W | ~16W |
| Estimated wall idle | ~150–180W | ~130–160W | ~20W |

> Measured figures to be recorded after swap — see Phase 2 update below.

---

## Phase 3 — GPU Upgrade (cards arriving, date TBC)

**GPUs incoming:** 2× NVIDIA GeForce RTX 3060
- 12GB GDDR6 each — **24GB total VRAM across both cards**
- TDP: 170W each — 340W combined at full load
- Architecture: Ampere — proper tensor cores, CUDA compute 8.6
- Cost: TBC

### AI capability step-change

| Capability | Quadro M4000 | 2× RTX 3060 |
|------------|--------------|--------------|
| VRAM | 8GB | 24GB (12GB×2) |
| Max model size (quantised) | ~7B | ~70B (split across cards) |
| Inference speed | Slow | Fast |
| Stable Diffusion | Marginal | Excellent |
| Tensor cores | No | Yes |

### Power at full AI load (estimated)

| Component | Full Load Draw |
|-----------|----------------|
| 2× E5-2620 v4 | ~120W |
| 2× RTX 3060 | ~300W |
| Motherboard/RAM/fans | ~50W |
| **Estimated wall draw** | **~500–550W** |

Z840 PSU is 1125W — plenty of headroom.

---

## Phase 4 — Power Mitigation Scripts (planned)

Once GPUs are installed, a set of scripts and Hermes integrations will manage
power dynamically to reduce idle draw:

### GPU power capping (nvidia-smi)

At idle, cap each 3060 to ~50W. On job start, restore full power limit:

```bash
# /usr/local/bin/gpu-power-idle.sh
nvidia-smi -i 0 -pl 50
nvidia-smi -i 1 -pl 50

# /usr/local/bin/gpu-power-full.sh
nvidia-smi -i 0 -pl 170
nvidia-smi -i 1 -pl 170
```

### CPU core parking (single socket at idle)

When no heavy job is queued, offline all cores on socket 1:

```bash
# /usr/local/bin/cpu-park-socket1.sh
for cpu in $(cat /sys/devices/system/node/node1/cpulist | tr ',' ' '); do
  echo 0 > /sys/devices/system/cpu/cpu${cpu}/online
done

# /usr/local/bin/cpu-unpark-socket1.sh
for cpu in $(cat /sys/devices/system/node/node1/cpulist | tr ',' ' '); do
  echo 1 > /sys/devices/system/cpu/cpu${cpu}/online
done
```

### Hermes integration (planned)

- Add GPU wattage to Hermes Z840 stats endpoint (`nvidia-smi --query-gpu=power.draw`)
- Alert if combined GPU draw exceeds 400W unexpectedly
- Dashboard panel showing real-time power consumption
- Auto-cap GPUs after 30 minutes of idle inference

---

## Phase 2 Results — POST CPU SWAP (measured 2026-05-19, ~2 hours post-install, 34°C ambient)

```
CPU socket 0 (E5-2620 v4): 15.2W
CPU socket 1 (E5-2620 v4): 14.0W
DRAM socket 0: 1.9W
DRAM socket 1: 0.9W
GPU (Quadro M4000): 42.3W
Total measured (RAPL + nvidia-smi): ~74W
Estimated wall draw: ~130–150W
Saving vs Phase 1 (CPU package only): ~12W
```

> RAPL sampled at idle (load avg ~0.00), 5-second window per domain.
> DRAM figures slightly higher than Phase 1 sample — within normal variance.
> CPU package power: 41.1W → 29.2W combined (-29% reduction).

## Phase 3 Results — POST GPU SWAP (to be filled in after cards arrive)

```
CPU socket 0 (E5-2620 v4): ___W
CPU socket 1 (E5-2620 v4): ___W
2× RTX 3060 at idle: ___W
Total measured idle: ___W
Estimated wall idle: ___W

2× RTX 3060 at full inference load: ___W
Total measured load: ___W
Estimated wall load: ___W
Saving vs Phase 1 idle: ___W
Monthly cost at UK avg (28p/kWh, idle 18h/day + load 6h/day): £___
```

---

## Cost Context (UK electricity, May 2026)

UK average electricity price: ~28p/kWh

| Scenario | Draw | Daily cost | Monthly cost |
|----------|------|-----------|--------------|
| Phase 1 idle 24/7 | ~165W | £1.11 | £33.80 |
| Phase 4 idle 18h + load 6h | ~160W idle + ~500W load | £0.97 | £29.50 |
| Phase 4 idle only (nights off) | ~160W × 12h | £0.54 | £16.40 |

> Moral: the CPUs are not the enemy. The GPUs at full load are. Smart scheduling
> and power capping matter more than the CPU swap — but every watt counts.