175: Chapter 175 The Vanished Thirty GB

On the sixty-second day of the twelfth trip to the Wasteland.

The newly added photovoltaic array, the eighty-kilowatt-hour energy storage system in the Outpost, and the independent computing bus in Workshop No. 2 had just passed the continuous load verification.

In the corner, four servers still slumbered in a powered-down state, while only half of the two massive storage devices had been awakened.

In the center of the workbench, the A100 baseline node had been running silently for twelve hours.

The core hot spots, VRAM temperature, and power modules all presented reassuringly stable curves.

This was thanks to the cooling path that Jiang Lin had modified four times.

The billowing heat flow discharged from the servers was precisely guided into the plate heat exchanger on the side of the rack, and then silently dissipated into the wilderness outside the Outpost along the external low-speed air duct.

After the computing node shut down, Jiang Lin encapsulated the complete startup logs into a read-only baseline.

[A100 Baseline Node / Operating Environment Completed]

[Baseline Card: A100 PCIe 40GB]

[Continuous Load: 12h]

[ECC Uncorrectable Errors: 0]

[Thermal Path: Stable]

[Power Supply Priority: Tier-3 Interruptible Load]

This was not the appearance that the 170HX would inevitably present after lifting its restrictions.

There might be differences in board design, power supply capacity, manufacturing screening, and product positioning.

This A100 was only responsible for providing an open, complete, and reproducibly measurable GA100 reference environment.

Jiang Lin pulled up the complete records of the A100 from power-on to the establishment of the computing context, peeling back the layers of the entire startup process into six levels: from Physics connection (L0), identity configuration (L1), security firmware and HBM training (L2), resource table generation (L3), host driver mapping (L4), all the way to the final runtime and task interface (L5).

The information interlocked with each other without a single flaw.

The five groups of HBM completed training, and the Physics resources smoothly entered the device resource table, exposing 40GB of VRAM to the host.

The driver established a complete address space, and the runtime finally grasped the full 40GB of allocatable capacity.

Jiang Lin saved the last set of data and turned around to connect the C170-01 to the same recording system.

A striking red label was affixed to the edge of the board.

[Disassembly Allowed]

[Failure Allowed]

Along with the establishment of the power supply, the board identity read was completed.

The five HBM stacks entered the training process sequentially, and the host completed device enumeration.

The PCIe Physics width was only four lanes, and the current negotiation state stagnated at Gen1.

Jiang Lin enforced link retuning and re-read the link state under device load.

The result still remained at Gen1 × 4.

This difference began to show clues from the device enumeration stage; it might originate from board wiring or startup configuration, but it could not be confused with VRAM capacity for the time being.

The ghost truly worth tracking lurked in another path.

Until the end of L2, the five groups of HBM training behavior presented by the C170-01 to the outside world had an extremely high degree of similarity with the A100.

However, once crossing the threshold of L3, the results shrank abruptly.

[GPU Architecture: GA100]

[Device Identity: CMP 170HX]

[HBM2e: 10240MiB]

[Bus Width: 5120-bit]

[PCIe Current Negotiation: Gen1 × 4]

[Computing Resources: Restricted Configuration]

The five HBM stacks were still recognized, but each group stingily delivered only 2GB of capacity to the host.

Jiang Lin overlapped and compared the two records.

The difference did not appear in the host driver, nor in the CUDA runtime.

When the host first obtained the available resource description of this card, 10GB was already the final answer processed by the underlying layer.

Existing records could only confirm that the first clear divergence in host-visible VRAM appeared between the end of L2 and the submission of the L3 resource description.

As for whether the difference came from fuses, board configuration, or firmware mapping, it still needed to be further dissected.

He wrote down the first set of judgments on the task list.

[PCIe Link Limitation: Independent Issue]

[Physics HBM Capacity: Unconfirmed]

[Fuse / Board Configuration Limitation: Pending Investigation]

[Host Driver / Runtime Limitation: Low Probability]

[Firmware Resource Mapping Limitation: High-Priority Candidate]

Afterwards, he removed the C170-01 from the test rack.

Removed the heatsink and stripped off the shielding cover.

The flash memory on the edge of the board, the startup configuration components, and a row of test points without public definitions were exposed without reservation under the pale light of the workbench.

In the first year, Jiang Lin tried to find a breakthrough on the host side.

He tried modifying open kernel modules, adjusting runtime allocators, reserving addresses for space outside of 10GB, and even replacing the board identity to try to trick the driver.

However, as long as the resource table handed over by the security firmware remained unchanged, the host's field of vision would forever be imprisoned within ten GB.

Modifying the displayed value was merely turning up the numbers on the dashboard; 30 liters of fuel would not magically appear in the fuel tank out of thin air.

As long as he tried to cross the boundary, the security firmware would ruthlessly refuse to load.

In the second year, the battlefield shifted to the board-level isolation bench.

The logic analyzer was connected to the startup configuration link, and the same cold boot was torn into hundreds of thousands of address waveforms.

Jiang Lin was like panning for gold in a vast sea of garbled code, comparing them one by one with the A100.

Excluding interference items such as serial numbers, power supplies, and sensors, the differences were compressed all the way from more than 400,000 to 971 locations, ultimately leaving only five undeciphered unknown regions.

In the spring of the third year, the first Physics intervention paid a heavy price.

Frequent desoldering and soldering caused the pads of the C170-01 to peel off; even though Jiang Lin reluctantly rebuilt the connection with jumper wires, after nineteen cold boots, the device still could not complete the resource mapping stage.

The 13,000-yuan board was declared scrapped, in exchange for a cold conclusion.

Directly replacing the signature configuration failed to bypass the verification.

Continuing to trial and error along this path meant the destruction cost had already exceeded the information value that could be obtained.

A white slip of paper was added next to the red label.

[Value Exhausted]

In the fourth year, C170-09 took over.

Jiang Lin no longer forced his way through, but instead made an interposer board to monitor.

Over 2,600 cold boots, spanning drastic temperature differences from minus 15 degrees to 45 degrees, all attempts to cross the timing boundaries were mercilessly terminated by the security firmware.

This path also came to an end.

Until the winter of the fifth year, MPS-FaultLedger marked a set of anomalies from more than 2,000 failure records.

The anomaly came from C170-29.

Although the first two cold boots failed, it actually completed the five groups of HBM training, yet failed to generate a resource table.

Jiang Lin connected it to the interposer board to reproduce the waveform, and finally caught a low-level maintenance stream extended by 0.8 seconds at the end of L2.

In that incomplete data, behind the training records of the five HBM stacks, each was followed by the same capacity field.

[Single Group Capacity: 8192MiB]

Adding the five groups together totaled 40,960 MiB.

Which was the 40GB nominal in the product specifications.

All five channels completed training.

At least at this stage, the logs did not show an overall missing 30GB.

As for whether there were bad sectors in the hidden space, it could only be confirmed through subsequent full-address read/write and long-term scrubbing.

This 30-GB ghost was quietly hidden in a dark room that the host could never see when the next-layer resource table was generated.

Knowing where the 30GB was and actually using it had an entire set of security links spanning across in between.

Jiang Lin did not have the signature key, but he sharply locked onto that 0.8-second maintenance window.

In the sixth year, 11,000 cold boots.

Jiang Lin discovered that when timing drift occurred in the level shifter on the startup configuration link, the security firmware would briefly enter a recovery path unrecognized by regular drivers.

He therefore temporarily marked this transitional state as [SERVICE_RECOVERY].

This was merely an engineering naming based on behavior and did not mean that the manufacturer had publicly defined a mode with the same name.

Next, he spent a full half-year using the interposer board through controlled timing delays to stretch this extremely harsh window from 0.8 seconds to 6.4 seconds.

6.4 seconds was not enough for the complete driver stack to complete device enumeration, resource mapping, and context establishment.

Jiang Lin therefore bypassed the general driver and wrote a minimal loader himself.

It did not load CUDA, nor did it create a computing context; it only did one thing.

Cornering this card to reveal its true background.

In February of the seventh year, C170-29 returned a complete Physics description in the recovery state for the first time.

PHYSICAL_HBM_STACKS: 5

PHYSICAL_HBM_CAPACITY: 40960MiB

The 40GB shelf was right behind the door, but there was still no road outside the door.

In the autumn of the same year, C170-29 lost all response during long-cycle recovery experiments.

The sample that failed to start twice consecutively in the Real World failed to last until the end, but with its own death, it handed over the location of the keyhole.

In the eighth year, lengthy address mapping.

The more stable C170-09 re-entered the test bench.

The address sketches on the wall had been changed sixteen times. Like a patient decrypter, Jiang Lin pieced together the Physics topology of the hidden space bit by bit through read-and-write feedback of all-zeros, interleaved bits, and pseudo-random sequences.

Continuous address folding, channel interleaving, row-column position mapping...

On the 109th time, the boundary folded back.

On the 376th time, non-zero data was returned outside of 10GB for the first time.

That was a test pattern left over from the HBM training stage.

Proving that his maintenance request had indeed stepped into the territory of the hidden zone.

June of the ninth year.

Jiang Lin wrote four bytes to the first valid address outside the 10GB boundary.

[Write: 0xA5A5A5A5]

[Read back: 0xA5A5A5A5]

These four bytes were a milestone.

Because it proved that the external loader could read and write to this restricted area.

The 10GB boundary began to slowly push outward.

15, 20, 30...

Along with the capacity activation, the radiated heat forced Jiang Lin to remake thermal pads and adjust the air ducts.

Three weeks later, testing restarted.

July 21st of the ninth year, 3:17 AM.

C170-09 completed the first round of full Physics address space writing and reading.

[Detected Physics Capacity: 40960MiB]

[Addressable Capacity: 40960MiB]

[Single-Round Data Errors: 0]

But he did not stop.

One thousand hours of continuous VRAM scrubbing, high-and-low temperature cycling, and random breakpoint recovery.

On the 107th day, two bad sectors were isolated.

The test results were finally encapsulated.

[Physics Capacity: 40960MiB]

[Stable Available Capacity: 39424MiB]

[Isolated Capacity: 1536MiB]

[Uncorrectable Errors: 0]

[Status: Maintenance-Grade VRAM Node]

It was still not an A100, general CUDA programs still could not call it, and the restricted PCIe link still existed.

But for MPS-Agent, computing and storage did not necessarily have to be bound to the same card.

These 39,424 MiB were already able to undertake cold-tier storage for checkpoints, failure ledgers, read-only evidence databases, and low-frequency access data.

However, under the condition that the computing context had not yet been opened and the PCIe link remained restricted, it could not act as a high-bandwidth cache for the A100, let alone directly undertake model sharded computing.

To make the 5120-bit HBM bus truly generate value, the next step must allow the computing units on the board to directly access the hidden address space.

4:02 AM.

The screen of the A100 baseline node on the left side of the workbench lit up with: HBM2: 40960MiB.

On the isolation rack on the right, the screen of the weathered C170-09 displayed: Stable Maintenance Mapping: 39424MiB.

The appearances of the two cards were now worlds apart.

The A100's metal casing was still clean and smooth, while the C170-09 had been disassembled more than a dozen times, with an interposer board hanging on the edge of the board and messy jumper wires soldered onto the test points.

Beneath that faded red label, half of a blue mark left by the mine operation and maintenance personnel back then could still be faintly seen.

A 92,000-yuan A100 provided a complete reference.

C170-09 proved with nine years that out of the 30 GB that had been trimmed away, most had not disappeared from the Physics layer.

A few regions indeed had defects, but the rest of the capacity simply did not enter the product resource table oriented toward the host.

Jiang Lin opened the twelfth transmission task list and calmly typed out the conclusion.

[Device-Side Resource Mapping Limitation Confirmed]

[Limitation Source: Fuse / Board Configuration / Firmware Policy, Not Yet Completely Dissected]

[Single-Sample Hidden Capacity Recovery Completed]

[Current Method Relies on Maintenance Mapping and Does Not Yet Meet Real World Delivery Conditions]

Afterwards, he created a new project directory: MPS-NearMemory_Runtime_v0.1.

The piece of paper pressed next to the A100 transportation box had slumbered in the Outpost for nine years.

The paper's surface lost its original smoothness, and the edges turned slightly yellowish.

The question written back then by Professor Zhou Zhihua remained clear.

[After compression, what cannot be lost?]

Jiang Lin picked up his pen and added a line of characters below.

[After trimming, what is still there?]

Those 30 GB were not recreated.

Most of them had always been there.

It was only until today that they entered a system capable of reading, verifying, and isolating bad sectors for the first time.