176: Chapter 176 Breaking the Barrier in Fourteen Microseconds

[Stable Available Capacity: 39424MiB]

Thirty-nine thousand four hundred and twenty-four MiB.

This figure translates to 38.5 GiB, conclusively proving that there is stable storage space beyond 10 GB inside C170-09, but it proved only this single fact.

In the recovery state, the gates of the computing context remained tightly shut.

The host could only read and write to that space through maintenance mapping; spanning between the board and the server was still that narrow PCIe link.

Sending data in took time.

Reading it back out meant going through it all over again.

The 5120-bit bus between the GA100 and the five groups of HBM did not truly participate in computation under this working mode.

It was like a huge warehouse built at the end of a narrow corridor; using it to save checkpoints might not even be more cost-effective than two sets of rack-mounted storage, and using it to train models was completely meaningless.

Jiang Lin did not check the option to obtain long-term running computing nodes on the twelfth transfer task list.

He simply typed a line of annotation calmly after the newly created project directory.

[Current Status: Addressable, Non-computable.]

Jiang Lin called up the two sets of resource charts for the A100 and the 170HX.

The A100's path was smooth and completed in one go.

Physics resources entered the device resource table, generated computing partitions, the driver created the context, and finally the runtime submitted the kernel.

Yet the 170HX was like two parallel lines forcibly cleaved apart.

When booted normally, it could compute, but could only see a cramped 10GB of VRAM.

Upon entering the recovery state, it sat upon a massive 40GB warehouse, yet its hands were bound, unable to compute.

On one side were hands, on the other side was a warehouse.

In the manufacturer's rigid rules, there was no room for both to be effective simultaneously.

In the winter of the tenth year, Jiang Lin began searching for the intersection of the two paths within the abyss of code.

He had tried out-of-bounds access, modifying resource tables, forcibly writing device page tables, and even attempting the mapping of virtual address holes.

The results obtained were without exception: rejection, error reporting, termination, and returning all zeros.

The price was heavy.

After the sixty-eighth test, C170-21 no longer accepted new hidden mapping requests.

The normal 10GB mode could still boot, and the original computing load was not affected.

Jiang Lin withdrew it from the hidden mapping experiment and transferred it to the 10GB routine node group.

[C170-21]

[Basic Computing Functions: Retained]

[Hidden Mapping Experiment: Terminated]

[Failure Reason: Hidden mapping state unrecoverable]

It was not scrapped; it had merely lost the qualification to continue answering this question.

The remaining cards could no longer be consumed in this blind-man-feeling-the-elephant manner.

Jiang Lin stopped the testing and spent three months dissecting the sixty-eight failures layer by layer.

Finally, in the fourth group of records returning null values, he captured a trace of subtle inconsistency.

Except for the closing timing of the maintenance mapping, the remaining conditions of the seventeenth and forty-ninth tests were completely identical.

The same virtual addresses.

The same page table format.

The same computing kernel.

This meant that the request had continued downward and stuck in some deep logic.

The only difference between the two experiments was the closing time of the maintenance read-write window.

In the forty-ninth test, after the computing context was established, the maintenance mapping was kept open for an additional fourteen microseconds.

Fourteen microseconds.

So short that human nerve endings could not perceive it, yet enough for a maintenance request that had already entered the device to forcibly occupy an unreleased mapping slot before the product isolation layer closed.

Jiang Lin reopened C170-09.

This time, he no longer fought against the manufacturer's rules, but went with the flow.

He let the normal computing context be fully established, while using an external interposer board to precisely seize the slot before the maintenance window closed, submitting a diagnostic mapping request complying with the maintenance interface format.

All he had to do was make this temporary window, which should have been released upon configuration completion, close a little later.

Extending the window to twenty-seven microseconds caused the device to error out.

Extending it to forty-one microseconds caused the device to reboot.

By the one hundred and eighty-ninth test, the window extension time finally stabilized at fifty-six microseconds.

The computing kernel ran to the third instruction and stopped at the read operation.

Jiang Lin did not record it as a success.

This could only prove that the failure location had advanced from the context establishment stage to the actual read stage.

Whether it had truly entered the physical space beyond 10 GB still required an independently verifiable return value.

March 17, Year 11.

Within 10GB, Jiang Lin wrote a set of consecutive integers.

Beyond 10GB, he wrote another set of data through that fifty-six-microsecond shadow window.

Subsequently, he submitted a summation kernel of a dozen lines.

The kernel, like a stalker, started from the visible space, silently crossed the 10GB boundary, and slipped into the temporary window.

The counter on the screen began to jump slowly.

[1024]

[2048]

[4096]

No shutdown, no return of all zeros.

[Input Elements: 8388608]

[Checksum: 35184367894528]

Jiang Lin called up the host-side independent calculation result; they matched completely.

Repeated one hundred times, one thousand times, changing data, changing window positions.

On the 7,421st time, the result remained as solid as a rock.

The iron wall written into the product configuration by the manufacturer had not vanished; Jiang Lin had merely forcibly wedged open a 256MB gap before the iron wall closed.

To make these 38.5 GiB truly participate in tasks, all computations had to be chopped up, passing through this gap like ants moving house.

He rewrote the runtime's three-tier architecture: Shadow_Aperture maintaining the window, Tile_Ledger recording data chunk ledgers, and Kernel_Gate responsible for boundary and rollback checks.

The first 32 GiB matrix block computation task would originally have been directly judged as out-of-memory by the general framework.

Jiang Lin chopped it into 128 blocks, each 256 MiB.

Data was first moved into the hidden space, mapped to the window during computation, leaving the result after calculation, before cutting the next block.

The first run took 9 hours and 47 minutes.

Iterating to the thirteenth version with the prefetching mechanism added, the time was compressed to 56 minutes.

In the autumn of the twelfth year, the stable running time dropped to 21 minutes.

Although the speed was far inferior to the A100, it was finally no longer just a warehouse that could only move things out through a crack in the door.

Jiang Lin updated the project status.

[Addressable / Computable / Incompatible with General Runtime / Requires Chunks and Checkpoints]

Opening up a single card only represented that this engineering path was feasible.

In the thirteenth year, the re-testing of the entire batch began.

The remaining thirty cards displayed riddled-with-holes personalities in testing.

Some threw massive errors under high temperature, some carried fixed physical bad sectors, and some failed frequently upon low-temperature startup.

Past mining farms did not need to distinguish these; as long as 10GB could run hash algorithms, all was well.

But in the Outpost, Jiang Lin had to re-understand the limits, temperature curves, and bad sector distributions of every single card, just like understanding a comrade-in-arms' scars.

The long polishing lasted until the end of the seventeenth year.

The stable computable capacity of the twenty-seven hidden mapping cards was distributed between 34 GiB and 39 GiB.

[Original Boards: 32]

[Destructive Loss: 2]

[Normal 10GB Mode Available: 30]

[Hidden Mapping Available: 27]

[Hidden Mapping Unavailable: 3]

[Hidden Space Stable Computable Capacity: Approx. 1TiB]

[Arbitrary Single Node Failure Recoverable: Passed]

Jiang Lin opened the twelfth transfer task list.

[Obtain Long-term Running Computing Nodes: Completed]

At noon in the summer of the eighteenth year, howling winds swept across the Wasteland, with twenty-four scratched old photovoltaic panels and two Wasteland wind turbines exerting effort together.

When the margins of photovoltaics, wind power, and energy storage all crossed the safety line, computing permissions were dispatched.

The fans of the four servers roared and sped up sequentially.

Thirty 170HX cards completed normal initialization.

Among them, twenty-seven established shadow windows, while the other three retained only the 10GB routine computing mode.

[NODE-A: Eight hidden mapping cards]

[NODE-B: Eight hidden mapping cards]

[NODE-C: Six hidden mapping cards + two routine cards]

[NODE-D: Five hidden mapping cards + one routine card]

[Cluster Status: READY]

Full load lasted only eleven minutes; after the wind speed dropped, the cluster began to shrink.

These eleven minutes did not complete any earth-shattering calculations, but it announced a miracle.

That batch of retired equipment packaged and cleared out by the mining farm reassembled on the Wasteland after the demise of human civilization into a heterogeneous near-storage computing cluster with a 1TiB-class schedulable HBM capacity.

This was not a complete large VRAM.

Every card had its own bad sectors, temperature boundaries, window sizes, and task ledgers.

MPS-Scheduler was responsible for making them look like the same system.

And that noble A100 now existed only as a standard validation node.

Different frequencies didn't matter, carrying bad sectors didn't matter; as long as task semantics were uniform and failures could roll back, this Wasteland franken-machine could operate.

After laying a solid hardware foundation, the next step was to let MPS-Agent enter the real computing environment.

Jiang Lin did not rashly connect GPT-NeoX-20B directly to the decision-making layer.

Testing proved that large models would conflate similar faults, resurrect old conclusions that had already been overthrown, and exaggerate special cases under specific conditions into general rules.

Its writing was extremely deceptive, yet putting it into underlying logic would trigger disaster.

Jiang Lin gazed at the note pressed under the glass plate for over a decade.

[After compression, what cannot be lost?]

For hardware, compressing 40GB to 10GB lost addresses.

For AI models, with millions of words compressed into parameters, what could not be lost?

He wrote down four iron rules beneath the paper.

[Source]

[Boundary]

[Counterexample]

[Uncertainty]

Subsequently, MPS-EvidenceGate was established.

The model was reduced to a proposer; every extracted summary had to be accompanied by the original time, applicable conditions, and known counterexamples.

Isolated evidence without a source was intercepted across the board.

It did not thereby transform into a conversational intelligent agent.

The model would still propose erroneous correlations and omit truly important conditions.

The change occurred in only one thing.

Content lacking a source, missing boundaries, or bypassing counterexamples could no longer be quietly written into MPS-Memory.

The model could make mistakes.

Mistakes had to leave a position.

From the nineteenth year to the twenty-second year, MPS-EvidenceGate was rewritten twenty-seven times in total.

Initially, it could only process structured fault logs.

Later, it gradually connected papers, code commits, manual annotations of test images along with structured detection results, proof versions, and transcribed handwritten records.

Proximity in time no longer automatically equaled causal correlation.

The latest summary could not overwrite old conclusions.

Paths already punctured by counterexamples could not be resurrected from the failure ledger just because the model regenerated a piece of fluent text.

By the winter of the twenty-second year, the real-world adaptation version of MPS-Agent passed the current known answer regression set for the first time.

[MPS-Agent α Real-World Adaptation: Controlled Input Version Completed]

March of the twenty-third year.

Computing power began processing the unclosed items left behind by the eleventh Wasteland expedition.

Northeast Seventy-Three was not a newly appeared direction.

PMCU-17, L-NE-01, and L-NE-02 had long formed the ARC-NE-01 candidate arc segment on the map, and the northern collapse zone was also marked as a broad orange zone of the outer-ring cooling trunk line.

The problem was that that orange zone still covered dozens of square kilometers.

The candidate arc segment could not be closed.

There were no answers for the signal source, propagation path, and burial depth.

Constrained by PMCU-17's original logs and the existing outer-ring model, MPS-Diagnose expanded the parameter space into over 80,000 groups of pseudo-environments.

Cooling medium state.

Node clock bias.

Propagation speed.

Broken link location.

Relay turnaround.

MPS-Scheduler was responsible for rotating and executing these reproduction tasks according to computing power, power supply, and thermal load boundaries.

MPS-EvidenceGate, on the other hand, compressed each set of candidate results back into the seven original maintenance events. Any model unable to simultaneously explain three evidence points was eliminated.

That arc which had been unclosable for decades shrank drastically on the map, finally shrinking into a dark red area over 800 meters wide in the northeast direction.

[Candidate Central Area: North side of Northeast Seventy-Three collapse zone]

[Horizontal Error: ±420m]

[Estimated Depth]

[Judgment Nature: Outer-ring common signal source]

Jiang Lin opened the TM-7 inner-ring access prerequisite list.

He did not cross off the first item "entrance localization unclosed," but calmly added a stroke behind it.

[Common signal source candidate area established.]

Subsequently, he called up the status records of G-Explorer-C.

The backup joints were checked, the communication relay nodes were in stock, and the environmental record box was unsealed.

Those expedition supplies that accompanied him across the years, after lying dormant for many years, finally welcomed their true destiny.

In Workroom No. 2, the roar of the mining card fans was subsiding group by group.

That heterogeneous cluster with a 1TiB-class schedulable HBM capacity did not directly crash open the door of truth for him.

It merely violently compressed a boundless Wasteland into a small red patch on the map.

Jiang Lin turned around, took a wrench from the mottled tool wall, and walked steadily toward G-Explorer-C.

Computation had narrowed the road to the extreme.

The remaining distance needed to be walked by the machine itself.