120: Chapter 120 ICM Online Report Invitation

Single-day Net Profit Settlement — [ + 18,742.91 ]

[ Current Equity: 371224.70 ]

This figure was not worth mentioning in the eyes of financial tycoons, and was not even enough for the cost of a business dinner.

But for a personal micro-account with a starting capital of only around 300,000, a profit growth of over 5% in a single day was already a very fierce plunder.

Jiang Lin knew very well in his heart that this was not because the MPS-Kernel possessed the ability to foresee the future and predicted the movements of the market main force for him.

The true qualitative change brought about by this technological upgrade lay in the fact that the Physics capacity upper limit of this quantitative system began to be substantially raised upward.

The money in the account was merely an appearance projected by the technological advantage into the Real World.

What truly underwent profound changes was the ambition that this system could carry in the future.

It could be said that capacity was the cornerstone for carrying massive funds.

After eating dinner, Jiang Lin accompanied his parents to watch TV and chat, rested for a while, and continued to replace the second batch of kernels.

median7_window

top3_of_8

cross_section_rank

rolling_quantile_update

For this replacement, Jiang Lin no longer needed to handwrite line by line like a traditional programmer, nor did he need to debug those headache-inducing assembly instructions or C++ underlying code.

The brilliance of MPS-Kernel v0.2 began to exert its true engineering value.

In the vast mathematical space, the semantic equivalence class searcher, like a tireless miner, rapidly generated hundreds or thousands of candidate code snippets.

The proof chain mechanism was like a ruthless sieve, instantly obliterating all erroneous code with minor logical loopholes and boundary overflow risks.

The robust cost backend, on the other hand, was responsible for the final adjudication, filtering out those pseudo-champions that scored extremely prettily in a single benchmark test but had extremely unstable time consumption under complex cache changes.

In the end, what logically entered the quantitative trading system was not a piece of incomprehensible magical assembly code in the traditional sense at all.

Instead, they were stacks of evidence cards attached with complete mathematical logic.

Every adopted kernel possessed a seamless mathematical proof state, clear microarchitecture application assumptions, measured time-consuming distribution histograms under various extreme pressures, and that forever-guaranteed return path.

While methodically connecting these evidence cards into the system, Jiang Lin watched as the performance bottlenecks of the entire trading system continuously moved upstream like a tide on the dashboard.

At the very beginning, the slowest blockages of system operation were mathematical calculations like rank and median.

When this part was completely flattened by MPS, the new shortcomings were immediately exposed.

The memory merging and cache line layout of Tick data became glaring.

When he casually optimized the data structure, what became slow was the thread scheduling loss during multi-strategy concurrent playback.

Immediately afterward, the bottleneck shifted to the disk I/O read and write rate of the solid-state drive.

Then came the Physics bandwidth of the memory channels.

Finally, it got stuck on the lock contention for state synchronization among multiple Physics cores of the CPU.

This system was like a complex mechanical device forcibly peeled away layer by layer.

Whenever he used a scalpel to excise an obstruction and peel away a layer of shell, the next more deeply hidden Physics bottleneck would immediately and cruelly be exposed to the air.

This feeling of constantly touching the Physics limits, constantly falling into bottlenecks, and constantly breaking through them was something Jiang Lin was far too familiar with.

During the long years of the ninth Wasteland, that behemoth, the MPS-Kernel, grew from crude code into an industrial-grade tool chain amidst the extreme scarcity of resources.

In the world of engineering, there was never any so-called end point.

There was only the next layer of limits forever waiting to be conquered.

It did not take long before the first version of the complete quantitative pipeline acceleration stress test report was generated on the terminal.

[ MPS-TimeLine v0.3 Comprehensive Performance Evaluation ]

[ Core hot path underlying replacement: Cumulative 17 places ]

[ Full-scale shadow comparison semantic difference: 0 ]

[ Historical concurrent backtest data throughput: Upgraded to 2.8x ]

[ Multi-variety / multi-timeframe parameter matrix scanning speed: Upgraded to 3.4x ]

[ Risk control engine p99 tail latency extreme value: Cliff-like drop of 37.6% ]

[ Order book micro-abnormal pulse detection defense window: Shortened to 54% of the original ]

[ Live trading interface access status: Core strategy enabled on a small scale normalized basis ]

Jiang Lin leaned his body against the back of the chair, closed his eyes, and forcefully rubbed his sore and swollen brow.

This was merely replacing the outermost layer of basic microkernels on the ordinary commercial CPU side.

Deep within the MPS system, that much larger and more structurally complex heterogeneous search space, he did not dare to touch at all.

The more complete global optimization capability of MPS-Kernel v0.2, due to computing power limitations, could only be forced to remain in a dormant state.

If he wanted to continue pushing into the deep waters of technology and let MPS truly erupt with the industrial-grade suppression power it had displayed in the Wasteland, he needed exponentially growing computing power support.

The current hardware conditions were far too weak.

The workstation in front of him could write code, view logs, and remotely manage external nodes via SSH, but in Physics terms, it fundamentally lacked the realistic conditions to carry multiple heavy GPUs for all-weather concurrent computing.

The whole machine power supply, wires, sockets, and PDUs could not withstand that kind of long-term high load.

The narrow cooling fins were unable to suppress hundreds of watts of violent heat.

The pitifully few PCIe channel bandwidths would suffocate the graphics cards to death alive.

The system memory capacity was simply not worth mentioning in the face of the huge state matrix.

And the stability of home-grade electronic components under 7x24-hour full-load bombardment was even more fragile as if made of paper.

Jiang Lin opened his eyes, straightened his body, and opened a brand-new blank spreadsheet.

Title: [ Computing Power Procurement and Deployment Plan ]

Column 1: Consumer-grade high-density GPU workstation.

Column 2: Data center-grade enterprise GPU array.

Column 3: Public cloud / supercomputing center computing power leasing.

Column 4: Long-term planned self-built server room nodes.

Column 5: Risk boundaries of Physics compliance and funds.

Jiang Lin first cast his gaze towards the most grounded consumer-grade graphics card market.

With the roaring collapse of the cryptocurrency mining wave, the prices of high-end graphics cards that were once hyped to outrageous levels finally returned to a relatively rational range like receding seawater.

N-cards two flagship models, RTX 3090 and RTX 3090 Ti, were both equipped with 24GB of GDDR6X video memory.

For the state space search tasks Jiang Lin wanted to run, the Physics ceiling of the video memory capacity of the 3090 Ti had not broadened by even the slightest bit.

That extra tiny bit of single-card core frequency burst was like a cup of water trying to put out a cartload of burning wood in front of the massive scheduling tasks, fundamentally unable to solve the structural problems he was facing.

The bottlenecks encountered by the MPS-Kernel were never that the game frame rates were not high enough, but rather a strict operational research problem.

Under a fixed unit fund budget, power supply limits, and limited indoor heat dissipation and exhaust pressure, how many high-density concurrent scheduling tasks could the system simultaneously spread out and maintain?

[ 3090 Ti: The core single-card peak is indeed higher, but the Physics flaw of the video memory remains 24GB, and the power consumption curve and core thermal density are seriously high, making it extremely easy to trigger thermal throttling. Determined to be unsuitable as a basic unit for building a multi-card high-density array under the current tight budget. ]

[ 3090: The core video memory is both 24GB, but the procurement cost per unit computing power and the Physics chassis deployment density are more excellent. It natively supports underlying CUDA multi-card collaboration, P2P direct memory access, and dual-card NVLink bridging. Determined to be able to serve as the core building foundation for the first-stage short-term computing power pool. ]

Jiang Lin directly filled in the strategic objectives of the first stage — [4 × RTX 3090 Deep Learning Workstation Node].

After planning the local nodes, he switched browser tabs and opened the official documentation page for the NVIDIA A100 computing card.

The screen was flooded with enterprise-level server rack solutions.

It was filled with obscure abbreviations of various authorized system integrators.

It was filled with channel quotations that required layers of email approvals to obtain.

It was filled with lengthy delivery times calculated in months or even quarters.

It was filled with tedious after-sales SLA maintenance contracts.

It was filled with stringent standard 4U / 8U rack installation specifications.

It was filled with professional server room environments requiring constant temperature, humidity, and antistatic measures.

It was filled with specially customized piped liquid cooling or high-pressure air-cooled cooling modules requiring powerful air conditioning.

It was filled with exaggerated three-phase power supply standards.

As well as various compliance statements related to procurement, warranty, cross-border supply, and enterprise customer qualifications.

The A100 was undoubtedly pure heavy infrastructure for data centers.

For an individual, such strategic-level equipment was not completely isolated from the world.

As long as one had sufficient premium funds and patience, it was still possible to acquire a single card through certain marginal channels.

But if one wanted to rapidly purchase, install, and deploy an A100 system capable of unleashing its true performance in an extremely short time as an unbacked individual, the hidden cost would far exceed the exorbitant price tag of the computing card itself.

Buying a single card and plugging it into an ordinary motherboard was a waste of divine gifts.

What truly endowed the A100 with cluster-level power was not a single card alone, but the NVLink / NVSwitch interconnection within complete machine platforms like HGX / DGX, and the infrastructure jointly formed by server motherboards, PCIe / NVLink topologies, server room power supply and cooling, driver stacks, and scheduling systems.

The A100 architecture was not suitable as a procurement solution currently pursuing rapid implementation.

Temporarily frozen.

Afterward, he casually opened the documentation page for the H100, which represented an even more distant future.

The microarchitecture was extremely gorgeous.

The officially announced theoretical performance parameters were gorgeous enough to make one's heart race.

However, the Physics problems lying before reality were equally glaringly clear.

There was no stable spot supply in the entire market.

Procurement channels were even more closed and burdensome than those for the A100.

The price was even higher, reaching an astronomical figure that made him despair at the current stage.

For Jiang Lin right now, it was not an object that could be incorporated into an actual procurement plan at all.

It was merely a lighthouse standing far away on the route of technological development, serving as a distant reference point to measure his current computing power gap.

Jiang Lin did not waste too much ink on this column.

He was very clear that his current scale did not yet have the ability to solve problems of that magnitude.

Writing down any more plans regarding supercomputing centers now would merely be meaningless self-intoxication.

His attention and focus contracted again, returning to the spreadsheet filled with red lines before him.

Four-card 3090 workstation.

This was the only computing power node that could realistically be implemented first after countless compromises and calculations.

In the third column of the spreadsheet, the cloud computing power leasing column, he did not cross it out with his pen immediately.

From an engineering perspective, the public cloud was undoubtedly an excellent resource pool.

Its elastic expansion capability was extremely strong, scaling up or down on demand at any time.

For truly major clients, GPU resources far exceeding personal workstations could be spun up within minutes to hours.

During certain specific time periods, as long as one was willing to pay high pay-per-use bills, one could even temporarily rent a computing card cluster array far more formidable than the local workstation he planned to build.

But in this world connected by fiber optic cables, any external server not within one's own absolute physical control range was a dangerous dark room.

External computing power could at best be regarded as a soulless, cheap coolie, used to process peripheral edge verification tasks that were completely stripped of core business logic, deeply desensitized, and impossible to piece together reversely.

For example, rendering public robot test videos that might be needed in the future, transcoding tasks could be moved to the cloud without hesitation.

If it was just running some generalization theory experiments using public datasets, it could also go to the cloud.

...

However, that complete quantitative trading strategy concerning the life and death of funds.

That complete MPS-Kernel engine poured with countless efforts from the Wasteland.

That complete underlying logic proof chain representing the limit of computing power.

That complete search state diagram containing the evolutionary direction of the system.

Not even a single byte of these true core secrets should travel through the network lines of the public cloud.

Time quietly elapsed.

Jiang Lin no longer hesitated and began placing the order.

Four RTX 3090s.

Workstation-grade motherboard.

Multi-PCIe lane workstation CPU.

256GB ECC memory.

NVMe array.

Redundant power supply.

Open-frame rack.

Industrial-grade UPS.

PDU.

Industrial fans.

Temperature, current, and card-drop monitoring modules.

10-Gigabit network card.

Two sets of dual-card NVLink bridges.

The payment page redirected.

The phone vibrated slightly.

The balance in the technical service fee account plummeted by a huge chunk instantly.

Jiang Lin was just about to shut down when an email popped up — "Invitation for You to Deliver a Special Online Report at ICM2022".

...

Mr. Jiang Lin:

On behalf of the relevant members of the Organizing Committee and Scientific Program Committee of the International Congress of Mathematicians 2022, we formally invite you to deliver a special online report during this congress.

Your recent preprint on the construction and verification of a single aperiodic monotile has attracted widespread attention in fields such as tiling theory, discrete geometry, symbolic dynamical systems, and computational verification.

This construction has currently been referred to as Jiangs Brick in related discussions.

Although the scientific program arrangements for this congress were determined long ago, the committee believes that your results possess special immediate academic value.

Therefore, we hope to arrange a specially added online report for you.

The tentative title is: "Jiangs Brick: Local Forcing, Finite-State Verification, and the Construction of a Single Aperiodic Monotile."

We suggest a report duration of forty-five minutes, followed by twenty minutes of Q&A and discussion.

The meeting will be held in an online format and open to registered ICM attendees.

With your permission, the recording of the report will also be archived as part of the congress online materials.

The committee particularly hopes that you can discuss the following issues in your report.

1. The design principles of the monotile and its boundary forcing mechanism.

2. The finite-state verification framework used to rule out periodic tiling.

3. The Layer-7 boundary blockage problem in recent discussions and how it is handled.

4. The relationship between this construction and Penrose tiling, as well as subsequent single-tile candidate constructions.

5. Whether your verification method can potentially be reused for other problems in discrete geometry.

We understand that you are currently still in a special and early academic stage.

Based on this, if you are willing, the committee can arrange a senior expert in tiling theory or discrete geometry to act as the moderator for this report and assist you with the technical preparations required for the online report.

Please let us know at your convenience whether you are willing to accept this invitation.

If you accept, we hope you can provide a brief abstract to confirm the final report title and inform us whether you agree to the congress saving the recording of the report.

We would be deeply honored if your report could be included in the schedule of this congress.

With sincere regards.

Professor Elena Marković

On behalf of the ICM 2022 Organizing Committee

International Congress of Mathematicians 2022

Professor David R. Holt

On behalf of the Special Lectures Coordination Group

Discrete Geometry and Aperiodic Order Direction