I Flashed The Matrix VBIOS And Now I Train Models All Day

Yesterday I wrote about how AI failed to help me find the InfoROM for VBIOS flashing. It could not do it. I had to do it myself. I spent the night reading forums. Reading modding guides. Reading warnings that I should not be doing this.

I have an Astral ROG RTX 5090 OC LC. Not the Matrix card. The Matrix is the three thousand dollar limited edition with the gold finger connector. I have the regular Astral LC with one sixteen-pin connector. The Matrix VBIOS can be flashed to the Astral with a resistor mod. I did the mod. I flashed the BIOS. It worked.

The AI could not find the InfoROM. I found it. The difference between AI and a person with a hex editor and too much time.

The New Numbers

Here is what I am running now. These numbers are real. These numbers are stable. These numbers make my power supply nervous.

Specification	Stock 5090	Astral LC OC	My Modded VBIOS
Max Core Clock	2407 MHz	2610 MHz	3480 MHz
Max VRAM Clock	28 Gbps	28 Gbps	17501 MHz (70 Gbps)
Max Wattage	450W	600W	800W
Power Connector	1x 16-pin	1x 16-pin	1x 16-pin (Risky)
Cooling	Air / AIO	AIO Liquid	AIO Liquid

How I Actually Did It

The AI failed at step four because it could not find the InfoROM. The InfoROM was not missing. It was in a different location on the Astral compared to reference cards. The AI did not know this. I had to find a Matrix VBIOS dump. I had to compare the structures. I had to find the power table offset.

Here is what actually worked:

                        # The actual process that worked

                        1. Dump original VBIOS with GPU-Z

                        2. Download Matrix 1200W VBIOS from modding forum

                        3. Flash Matrix VBIOS (ignore warnings)

                        4. Use NVFlash to set power limit to 800W

                        5. Set core clock offset to +350 MHz

                        6. Set memory clock to 17501 MHz

                        7. Save as new VBIOS

                        8. Flash to primary slot

                        # Yes I kept the backup BIOS ready.

The Matrix VBIOS has a 1200W power table. I flashed it. Then I lowered the wattage to 800W in software. This is safer than running the full 1200W. The sixteen-pin connector maxes at 600W officially but can handle more with good cables. I have good cables. I am not an idiot. Mostly.

Stability Report

✓ Memory Clock: 100% Stable

17501 MHz on GDDR7. No artifacts. No errors. No crashes. I ran memory tests for six hours. Everything passed. The GDDR7 on the 5090 can handle this speed with good cooling. The liquid cooler keeps the VRAM cool enough.

✓ Temperature: 60C Under Full Load

Sixty degrees Celsius. One hundred percent of the time. Under full training load. The AIO liquid cooler is doing its job. The radiator is warm. The room is warmer. The GPU is happy.

I am not melting. I am not artifacting. I am not crashing. This is the most stable my GPU has ever been. The higher power limit means the card does not throttle. It runs at consistent clocks. Training is predictable.

What This Means For Training

I will be training AI models for probably the rest of today. The Sonnet model is at 0.4 percent. With the new clocks it should finish faster. Maybe 200 hours instead of 261. Maybe less. I will know soon.

The memory bandwidth increase is significant. At 70 Gbps effective the data moves faster. The batch sizes can be larger. The iterations per second are higher. The loss curve goes down quicker.

I did not overclock for gaming. I overclocked for training. There is a difference. One is fun. The other is necessary.

The Card Situation

I do not have the Matrix card. The Matrix is the forty hundred dollar limited edition with the GC-HPWR gold finger connector. It has special power delivery. It can handle 1000W+ safely. I have the Astral LC. It has one sixteen-pin connector. It is rated for 600W. I am running 800W.

This is not recommended. This is not safe. This is what I am doing anyway. The connector has not melted. The cable is not hot. I monitor temperatures constantly. If anything gets warm I shut down.

Why The AI Failed

AI is trained on public data. VBIOS modding for RTX 5090 is not well documented. The Matrix VBIOS flash to Astral cards was only discovered in January 2026. The training data does not have this information. The AI could not help.

I had to use forums. I had to read modding threads. I had to trust strangers on the internet. This is how hardware modding works. This is how it has always worked. AI cannot replace community knowledge. Not yet.

Training Plans

Today is training day. No blogs. No debugging. No VBIOS modding. Just training. The Sonnet model needs to finish. The Opus model is waiting. The Haiku model is giving fish answers and needs companionship.

I will check the loss curve every few hours. I will monitor temperatures. I will watch the power draw. I will not touch anything. I will let the GPU do its job. I will do mine.

Final Thoughts

The VBIOS is flashed. The clocks are locked. The memory is stable. The temperature is controlled. The training is running. I am tired. I am happy. I am ready to not touch my GPU settings for at least a week.

If you are thinking about doing this, do not do this. If you are going to do this anyway, have a backup BIOS. Have good cables. Have a fire extinguisher. Have a plan. And maybe have a friend who knows what they are doing.