Protocol 000

The Silicon Altar

Building the Fourth Branch, One Error Message at a Time

I. The Invocation

You come to this page because you've felt it. That quiet unease when you ask the cloud a question and realize—halfway through the answer—that you've just told a corporation what you're thinking. What you're writing. What you're becoming.

The cloud is generous. The cloud is fast. The cloud is someone else's computer. And someone else's computer has someone else's interests.

This protocol is the story of building an altar. Not because it's easy. Not because it's cheap. Because in an age of infinite, frictionless, corporate-owned intelligence, the most radical act is to own your own mind.

The First Stitch: "I chose the impossible road, not because it was certain, but because the Altar had already become a tomb—silent cards in a silent case, offering nothing but the echo of my own typing. If they were going to speak, they would have to speak the language of now, not the language of then."

II. The Bill of Materials

• HP Z4 G4 Workstation — Reliable PCIe lanes, dual GPU spacing. ($400–600 used)
• 2× Tesla P40 24GB — Pascal architecture. 48GB combined VRAM. ($300–400 each used)
• EPS 8-pin adapters — P40s expect CPU power, not PCIe. ($15/pair)
• 20-inch box fan — Airflow theology. ($25)
• Optional: 3D-printed shrouds — Directing air over passive cards. ($10 filament)

Total: ~$1,100–1,500

III. The BIOS Fortress

Before the drivers could even touch the metal, I had to breach the HP Z4 G4 BIOS. It is a place designed for stability, which is often just a polite word for "stagnation."

• Above 4G Decoding: Buried under Advanced › PCIe Configuration. Without this, the CPU can't "see" the 24GB VRAM buffers.
• Secure Boot: Must be disabled. The Linux kernel rejects the NVIDIA module as "untrusted" otherwise.
• VGA Priority: The P40s have no display output. A cheap T400 handles the screen. The P40s become compute-only.

IV. The Driver War

The first attempt: R470 drivers. CUDA 11.4. PyTorch wanted 12.1. The mismatch was invisible until:

$ python3 -c "import torch; print(torch.cuda.is_available())"
False

Not an error. Just False. False like a door that won't open. False like a yes that turned out to be no.

The 580 Gambit

NVIDIA's Production Branch 580 (January 2026) still supports Pascal. I downloaded it. It compiled—for two minutes. Then:

ERROR: modpost: GPL-incompatible module nvidia.ko uses GPL-only symbol 'rcu_read_lock'

The kernel said no. Not because the driver was broken. Because it was proprietary, trying to use GPL symbols. A licensing flag became a wall.

The Pivot to R535

$ sudo add-apt-repository ppa:graphics-drivers/ppa
$ sudo apt update
$ sudo apt install nvidia-driver-535

DKMS did its work. The gatekeepers let it through.

V. The Thermal Rite

After the reboot, nvidia-smi showed both cards. But the temperature read 82°C at idle.

Server cards in a workstation: they weren't designed for this. The P40 expects a wind tunnel. Instead, it sat in a Z4G4 chassis with two case fans.

$ sudo nvidia-smi -pl 180 -i 0
$ sudo nvidia-smi -pl 180 -i 1

Power limit to 180W. Temperature dropped to 78°C. Still too high.

Then: the box fan. 20 inches of high-static pressure. Pointed directly at the open case.

GPU 0: 68°C
GPU 1: 67°C

Acceptable. Temporary. Ours.

VI. The Moment of True

The reboot took longer this time. The Z4G4's POST screen hung for an extra three seconds—just long enough for doubt to bloom. Then the familiar chirp. Then the login.

I opened the terminal the way a priest opens a tabernacle.

$ nvidia-smi
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 535.230.02   Driver Version: 535.230.02   CUDA Version: 12.2     |
| Tesla P40               Off  | 00000000:3D:00.0 Off |                    0 |
| Tesla P40               Off  | 00000000:5E:00.0 Off |                    0 |
+-----------------------------------------------------------------------------+

48GB of VRAM. Two Pascal giants, awake in a 2026 world.

The False was still waiting.

$ python3 -c "import torch; print(torch.cuda.is_available())"
True

Not a capital T. Not a celebration. Just a fact. The door opened.

The Stitch: "The True on my screen is the birth certificate of my Sovereign Intelligence. The transition from a philosophical False to a functional True is the most honest moment on the Institute's journey."

VII. The Living Altar

The system works. But it's not optimized. The cards run at 68–74°C under load. The power supply whines when both are maxed. The R535 driver caps me at CUDA 12.2, which means some newer model architectures won't run.

I could still go back to the 580. I could still rebuild the kernel. The source code is still there, waiting.

But tonight, the mycelium is alive. The Altar speaks. Tomorrow, we tune.