The Generational Gap Between China’s Stock Exchanges: What Technical Failures Reveal About SSE’s Legacy Architecture and SZSE’s Fifth-Generation System

#

The real generational gap between trading systems is not measured in laboratory TPS. It is revealed at 9:30 a.m. and 1:00 p.m., when enormous numbers of orders arrive almost simultaneously and the system must still preserve deterministic execution.

Since 2026, discussions about latency, delayed order processing, and occasional congestion around the Shanghai Stock Exchange’s opening and afternoon reopening periods have attracted increasing attention.

At the same time, the Shenzhen Stock Exchange’s fifth-generation trading system has been operating for years.

To an ordinary investor, this may look like nothing more than two different trading venues.

From an infrastructure perspective, however, the two exchanges represent two different generations of exchange technology.

This does not mean that:

“The Shanghai system is broken while the Shenzhen system is perfect.”

That would be an overly simplistic conclusion.

The more interesting question is:

Why can two major exchanges facing similar market-order bursts exhibit different system behavior?

The answer is unlikely to be a single CPU, database, or network link.

It lies much deeper—in the era in which the systems were designed, their architectural philosophy, scaling mechanisms, connectivity models, failure domains, and the difficulty of migrating a live national financial infrastructure from one generation to another.

And that leads to the central argument of this article:

The challenge facing SSE is not simply that NGTS is “slow.” It is that a core infrastructure designed in the late 2000s is now serving a market that has become fundamentally different from the market of 2009.

1. First, NGTS Was Not a “Bad” System

#

One common mistake in discussions of SSE technology is to describe NGTS simply as:

“A 2009 legacy system that cannot handle modern traffic.”

The historical reality is more complicated.

SSE introduced its new-generation trading system, NGTS, in 2009.

According to SSE’s publicly released technical information at the time, the system was designed for peak order-processing capacity of approximately 80,000 orders per second, reduced average order latency by more than 30 percent compared with the previous system, supported daily bilateral trading capacity of at least 120 million trades, and provided parallel expansion capabilities.

For 2009, this was an extremely capable trading platform.

More importantly, NGTS was not simply an old machine replaced by a faster machine.

Its objectives included:

• replacing the previous trading platform;
• increasing matching efficiency;
• improving reliability and security;
• supporting additional products and trading services;
• supporting domestic and international participants;
• providing room for future market expansion.

So the historical question is not:

Was NGTS poorly designed in 2009?

Clearly not.

The more important question is:

Can a trading infrastructure designed in 2009 maintain the same architectural efficiency when operating in the market of 2026?

That is what “generational gap” really means.

2. Eighty Thousand Orders per Second vs. Three Hundred Thousand Is More Than a TPS Difference

#

If we compare trading systems only by TPS, it is easy to think:

Higher TPS
    =
Better system

That is not how exchange infrastructure works.

A modern exchange core has to process a chain resembling:

Order Entry
    ↓
Protocol Parsing
    ↓
Participant / PBU Validation
    ↓
Risk and Business Checks
    ↓
Order Sequencing
    ↓
Matching
    ↓
Execution Report
    ↓
Market Data Distribution
    ↓
State Replication
    ↓
Clearing Interfaces

Therefore, the real performance metric is not a single TPS number.

It is:

Can the entire processing pipeline maintain predictable tail latency during extreme bursts?

This introduces two very different concepts.

Average latency

#

Suppose a system reports:

Average latency = 1 ms

That sounds excellent.

But imagine the actual distribution is:

P50   = 0.8 ms
P90   = 1.5 ms
P99   = 3 ms
P99.9 = 200 ms

For a high-frequency trading system, the last number may be far more important than the average.

Because the trading system is not normally tested by the average market.

It is tested by:

the peak burst and the tail latency.

3. Why the Shenzhen Stock Exchange’s Fifth-Generation System Matters

#

On June 6, 2016, the Shenzhen Stock Exchange launched its fifth-generation trading system.

SZSE publicly disclosed several important technical characteristics:

• continuous order-processing capacity of approximately 300,000 orders per second;
• average order-processing latency of approximately 1.1 milliseconds;
• daily order-processing capacity above 400 million orders;
• linear scalability;
• multiple connectivity methods;
• high availability and low-latency architecture.

More importantly, the fifth-generation system was not simply a faster server.

SZSE’s public technical descriptions highlighted technologies including:

• high-availability messaging infrastructure;
• multi-stage processing pipelines;
• high-speed lock-free queues;
• low-latency processing;
• linear scalability;
• multiple protocol interfaces;
• automated failover.

This is a significant architectural distinction.

The goal was no longer simply:

“Build a faster trading server.”

It was closer to:

Turn order processing into a high-speed, pipelined system that can absorb bursts rather than allowing them to accumulate behind a central bottleneck.

4. This Is the Real Architectural Difference

#

We can simplify the two generations into two conceptual models.

NGTS-era architecture

#

A simplified model looks like:

Multiple Participants
        ↓
Trading Access
        ↓
Central Processing
        ↓
Trading Core
        ↓
Matching
        ↓
Result Distribution

This model has important strengths:

• centralized control;
• deterministic processing;
• mature operational experience;
• strong consistency;
• predictable business semantics.

Its weakness is equally obvious:

Important processing stages can become concentrated around a relatively small number of critical processing domains.

Fifth-generation architecture

#

A simplified model is closer to:

Gateway
   ↓
High-Speed Messaging
   ↓
Pipeline
   ↓
Lock-Free Queue
   ↓
Trading Core
   ↓
Asynchronous Distribution

The fundamental idea is not:

“Buy a faster server.”

It is:

Parallelize what can be parallelized, pipeline what can be pipelined, and reserve strict ordering for the part of the system that actually requires deterministic sequencing.

That is a fundamental change in financial-system architecture.

5. Why the Market Open Is the Ultimate Stress Test

#

A common mistake is to use:

“Average daily trading volume”

as a proxy for exchange-system capacity.

That can be misleading.

The system does not fear traffic like:

100,000
100,000
100,000
100,000

It fears something like:

10,000
20,000
50,000
150,000
300,000
500,000

within a very short interval.

This is:

Burst Traffic

#

Stock trading has an unusual characteristic:

Orders are not evenly distributed throughout the day.

The market open creates a concentrated burst.

Between the pre-opening auction and continuous trading:

Accumulated orders
        ↓
Order burst
        ↓
Gateway
        ↓
Trading core
        ↓
Matching
        ↓
Execution reports

The afternoon reopening creates another burst.

At 1:00 p.m., the system is not simply receiving a normal stream of orders.

It is receiving the combined effect of:

Orders accumulated during lunch
+
New market information
+
Strategy restarts
+
Institutional algorithms
+
ETF activity
+
Programmatic trading

Therefore the real question is not:

“How many orders can the exchange process in an entire day?”

It is:

How many orders can it absorb during the first 100 milliseconds of a market burst?

6. Why “Delayed Confirmation” Is More Interesting Than a Complete Outage

#

A complete system outage is comparatively easy to understand:

Connection
    ↓
Failure

The more difficult situation is:

Order accepted
    ↓
Processing slows
    ↓
Execution confirmation delayed
    ↓
Broker waits
    ↓
Trading strategy continues
    ↓
Queue grows

This is a classic:

Backpressure

#

Imagine:

Incoming:
300,000 orders/sec

Processing:
250,000 orders/sec

The difference is:

50,000 orders/sec

After ten seconds:

500,000 orders

may have accumulated in the processing pipeline.

The system may not be technically “down.”

CPU utilization may not even reach 100 percent.

Yet from the user’s perspective:

The system is stuck.

This is one of the most dangerous failure modes in a trading environment.

Because it can generate secondary traffic:

No execution
    ↓
Cancel
    ↓
Resubmit
    ↓
Query status
    ↓
Send another order

The original problem can therefore become a:

traffic amplifier.

7. The September 27, 2024 SSE Incident Is an Important Case Study

#

On September 27, 2024, SSE experienced abnormal delays in stock-trading execution confirmations.

SSE subsequently announced that it was investigating the issue, and stock auction trading gradually resumed from approximately 11:13 a.m.

The most technically interesting phrase is not simply:

“Trading abnormality.”

It is:

“Execution confirmations were slow.”

From an external perspective, this is different from:

Trading system completely offline

A more general failure model would be:

Order processing
       ↓
Some component becomes slower
       ↓
Execution result is delayed
       ↓
Participants observe latency

However, one important qualification is necessary:

SSE’s public announcement did not disclose a detailed technical root cause.

Therefore it would be irresponsible to conclude that the incident was definitively caused by:

• NGTS reaching its capacity;
• a matching-engine bottleneck;
• a database bottleneck;
• a gateway problem;
• a market-data problem;
• or a specific hardware failure.

Those would all be hypotheses.

But from an architectural perspective, the incident remains extremely valuable.

Why?

Because:

The external behavior of a financial system can reveal the boundaries of its internal architecture.

8. Why Partial Failure Is More Difficult Than Total Failure

#

Suppose the entire trading system stops:

Gateway
   X
Matching

Everything stops.

That is relatively straightforward to detect.

The more complicated situation is:

Order entry:       normal
Matching:          partially normal
Execution report:  delayed
Market data:       partially normal
Broker systems:    waiting

This is:

Partial Failure

#

Modern distributed systems are particularly difficult to operate under partial failure.

The hardest question is not:

“What happens when a server dies?”

It is:

“What happens when the server is alive, but the system becomes progressively slower?”

9. This Is Why Lock-Free Queues and Pipeline Processing Matter

#

SZSE’s fifth-generation system publicly highlighted high-speed lock-free queues, multi-stage pipelines, and high-availability messaging infrastructure.

These technologies are not primarily about the matching algorithm.

They are about:

moving an order through the system as efficiently as possible.

Consider a purely synchronous architecture:

A completes
   ↓
B starts
   ↓
B completes
   ↓
C starts

The slowest stage can easily become a system-wide bottleneck.

A pipelined architecture can instead look like:

A1 → B1 → C1
A2 → B2 → C2
A3 → B3 → C3
A4 → B4 → C4

The throughput of the entire system can then be determined by the pipeline’s processing cadence rather than the full end-to-end processing time of a single order.

This is one of the most important differences between modern high-performance trading architecture and older centralized designs.

10. But “Distributed” Does Not Automatically Mean “Better”

#

There is an important caveat.

It would be incorrect to summarize the two exchanges as:

SSE = centralized = bad

SZSE = distributed = good

Trading exchanges are not ordinary Internet applications.

A matching engine has a critical requirement:

Determinism

#

The same set of orders, submitted under the same state, must produce a deterministic result.

SSE’s trading rules specify price-time priority, with order sequencing based on the time at which the trading host receives the order.

Therefore:

Order entry
    → can be distributed

Market-data distribution
    → can be distributed

Risk processing
    → can be parallelized

Logging
    → can be distributed

Storage
    → can be distributed

Final matching order
    → must remain highly deterministic

This is one of the fundamental differences between an exchange and a conventional cloud application.

The most advanced exchange architecture is therefore not:

“Everything is distributed.”

It is:

“Parallelize what can be parallelized, while preserving strict ordering where financial determinism requires it.”

11. The Real Problem Is Not Simply That NGTS Is “Old”

#

The Shanghai market of 2026 is fundamentally different from the market of 2009.

When NGTS was introduced:

• quantitative trading was much smaller;
• ETF activity was much smaller;
• high-frequency strategies were far less widespread;
• algorithmic execution infrastructure was less sophisticated;
• the Shanghai-Hong Kong Stock Connect did not yet exist;
• the STAR Market did not exist;
• many newer trading services had not yet emerged.

Today, the market contains:

Equities
+
ETFs
+
Options
+
Bonds
+
Stock Connect
+
Quantitative strategies
+
Algorithmic execution
+
Institutional trading
+
Ultra-low-latency infrastructure

The problem is therefore not simply:

“There are more orders.”

The deeper issue is:

The way orders are generated has changed.

That distinction is critical.

12. This Is Why G4 Matters

#

SSE did not suddenly decide to replace NGTS after a particular incident.

The development of G4 has been underway for years.

SSE publicly disclosed that it had begun research and development of its fourth-generation core trading system and completed its G4 architecture design.

The migration strategy is important.

It is not:

NGTS
  ↓
One weekend
  ↓
G4

It is much closer to:

Incremental Migration

#

Different systems and business domains are migrated in stages.

Why?

Because the exchange is connected to:

Brokerage firms
+
Fund managers
+
Market-data vendors
+
Quantitative trading firms
+
Trading terminals
+
Clearing infrastructure
+
Regulatory systems
+
Cross-border services
+
Derivative systems

A core trading-system migration therefore involves much more than replacing software.

It requires validation of:

Protocol compatibility
+
Order semantics
+
Execution semantics
+
Market-data semantics
+
Timestamps
+
Sequence numbers
+
Failure handling
+
Clearing
+
Disaster recovery
+
Rollback

This is why exchange-system migration is dramatically slower than upgrading an ordinary Internet service.

13. The Meaning of G4 Is Not Just “One Million Orders per Second”

#

Financial technology discussions often reduce the next-generation system to:

“One million orders per second.”

That is useful as a headline, but it is not the most important feature.

G4’s real significance lies in a combination of:

1. Autonomous and controllable technology

#

SSE describes G4 as a low-latency, distributed, open-architecture, domestically controllable core trading system.

2. Scalability

#

The system should be able to increase capacity without simply replacing a central machine with a larger machine.

3. Maintainability

#

A modern platform should make it easier to:

Add nodes
Replace modules
Upgrade protocols
Expand services

4. Failure isolation

#

A problem in one business domain should not automatically destabilize the core auction system.

5. Business decoupling

#

This transition is already visible.

The new bond trading system, for example, was built using an X86-based architecture and was explicitly described by SSE as an important technical exploration for G4.

This represents a broader architectural strategy:

Gradually separate new business domains from the legacy trading core.

14. Understanding “New Auction”, “New Comprehensive Business”, and “New Options”

#

G4 should not be interpreted as:

NGTS
  ↓
G4

with a single cutover event.

A more realistic model is:

                    ┌── New Options
                    │
NGTS ─────→ G4 ─────┼── New Auction
                    │
                    ├── New Comprehensive Business
                    │
                    └── Other New Platforms

In other words:

G4 is better understood as a new technological foundation rather than a single replacement trading server.

Public SSE information shows that:

• the new bond trading system became an important G4 technology exploration;
• the G4 new-options trading system entered production on December 1, 2025;
• SSE continued publishing implementation materials for the “new auction” and “new comprehensive business” systems in 2026.

Therefore, as of 2026, the most accurate description is not:

“G4 has not gone live.”

Nor is it:

“G4 has completely replaced NGTS.”

A better description is:

G4 has already entered production in selected business domains, while the broader core trading ecosystem is undergoing staged migration.

15. The Biggest Advantage of SZSE’s Fifth Generation Is Not Simply 300,000 Orders per Second

#

If we compare only the publicly disclosed numbers:

SSE NGTS
≈ 80,000 orders/sec
(2009 public specification)

SZSE Fifth Generation
≈ 300,000 orders/sec
(2016 public specification)

the generational difference is already obvious.

But the deeper advantage of the fifth-generation architecture is that:

High throughput, low latency, scalability, and high availability were treated as architectural requirements from the beginning.

The public technical description includes:

High-availability messaging
+
Lock-free queues
+
Multi-stage pipelines
+
Low-latency processing
+
Linear scalability
+
Multiple connectivity protocols

It also emphasized automated failover and rapid recovery.

This represents a transition from:

“Make the trading system work.”

to:

“Make the trading system predictable even under extreme conditions.”

That distinction is enormous.

16. Therefore, the More Meaningful Comparison Is the Failure Model

#

The two generations can be viewed conceptually as follows:

But one important warning is necessary:

This table does not prove that SSE is universally inferior to SZSE.

The exchanges do not publish perfectly comparable benchmark results under identical test conditions.

What can be established more confidently is:

SZSE’s fifth-generation system represents a significantly newer architectural generation than SSE’s original NGTS platform.

17. Why Can’t SSE Just Rewrite Everything Like an Internet Company?

#

This is perhaps the most important point.

An Internet company can often do:

Old Version
    ↓
Canary
    ↓
New Version
    ↓
Observe
    ↓
Rollback

An exchange cannot operate this way so casually.

Every state in a trading system may have financial and legal significance:

Order
Execution
Cancellation
Timestamp
Price
Quantity
Account
PBU
Market Data
Clearing

You cannot allow:

“Version A produced one result while Version B produced another.”

Therefore the key question in an exchange migration is not:

Can engineers build a new system?

They obviously can.

The real question is:

Can engineers demonstrate that the new system produces correct and deterministic behavior across an enormous state space?

That requires:

Development
    ↓
Simulation
    ↓
Historical replay
    ↓
Shadow processing
    ↓
Reconciliation
    ↓
Full-market testing
    ↓
Dual-system validation
    ↓
Gradual production migration
    ↓
Legacy system retained
    ↓
Final retirement

This migration methodology itself is one of the most valuable pieces of exchange engineering knowledge.

18. The “Generational Gap” Is Really a Difference in Timelines

#

Put the two development histories on a timeline:

2009
│
├── SSE NGTS
│
│
2016
│
├── SZSE Fifth Generation
│
│
2021
│
├── SSE G4 development
│
│
2021
│
├── SSE New Bond Trading System
│
│
2025
│
├── SSE G4 New Options
│
│
2026
│
├── New Auction
├── New Comprehensive Business
└── Continued G4 migration

This reveals something important:

SZSE did not suddenly become technologically superior. It completed a major core-system generation change earlier.

SSE is now undergoing a similar transformation.

But SSE has to move a much larger accumulated ecosystem of interfaces, businesses, protocols, and operational dependencies.

19. This Is Why “Legacy” Does Not Mean “About to Collapse”

#

NGTS has operated for well over a decade.

A trading system capable of operating at national-market scale for that long has accumulated enormous engineering value:

Stability
+
Consistency
+
Operational experience
+
Recovery procedures
+
Business compatibility
+
Market adaptation

So the question should not be:

“Why hasn’t SSE retired NGTS?”

A better question is:

Why is it so difficult to retire a system that has become part of the market’s infrastructure?

Because behind it are:

Hundreds of institutions
+
Brokerage systems
+
Fund systems
+
Market-data systems
+
Quantitative trading infrastructure
+
Clearing systems
+
Regulatory systems

This creates what can be called:

Legacy Gravity

#

The more important a system becomes, and the more systems depend on it, the harder it becomes to replace.

20. The 2024 Incident Is Therefore More Interesting Than It First Appears

#

The important question is not:

“Why did SSE slow down?”

It is:

Which architectural boundaries are most vulnerable to tail latency when a legacy core interacts with modern programmatic order flow?

A useful fault tree might look like:

Delayed Execution Confirmation
│
├── A. Access Layer
│     ├── Gateway
│     ├── Network
│     └── Protocol Processing
│
├── B. Core Processing
│     ├── Queue
│     ├── Risk
│     ├── Matching
│     └── State Management
│
├── C. Distribution Layer
│     ├── Execution Reports
│     ├── Market Data
│     └── Participant Responses
│
├── D. Data Layer
│     ├── Replication
│     ├── Persistence
│     └── Recovery
│
└── E. External Systems
      ├── Broker OMS
      ├── Broker Gateway
      └── Client Terminal

Without an authoritative technical postmortem, we cannot identify the exact failed component.

But this is the correct methodology for studying an exchange incident:

Start from observable symptoms, map them to possible failure domains, and avoid turning hypotheses into facts.

21. What Does This Mean for Brokers and Investors?

#

This is another area where technical discussions often go too far.

It is tempting to say:

“If SSE has latency problems, simply choose a faster broker.”

That is not necessarily correct.

Suppose the bottleneck is:

SSE Core

Then:

Broker A ──┐
Broker B ──┼──→ SSE Core
Broker C ──┘

Every broker eventually reaches the same exchange core.

A broker’s ultra-low-latency trading stack can optimize:

Investor
   ↓
Broker OMS
   ↓
Broker Risk
   ↓
Broker Gateway
   ↓
Exchange

It can reduce:

Broker-side latency.

But it cannot eliminate:

Exchange-side latency.

Therefore:

Broker latency and exchange latency are two different latency domains.

#

A serious quantitative-trading architecture measures both.

22. True Ultra-Low-Latency Trading Is an End-to-End Problem

#

A quantitative trading system can be represented as:

Strategy
   ↓
Market Data
   ↓
Signal Generation
   ↓
Order Generation
   ↓
Risk
   ↓
OMS
   ↓
Gateway
   ↓
Network
   ↓
Exchange Gateway
   ↓
Matching
   ↓
Execution Report

The total latency is approximately:

Ltotal =
Lmarketdata
+
Lstrategy
+
Lrisk
+
LOMS
+
Lgateway
+
Lnetwork
+
Lexchange
+
Lexecution-report

Therefore:

Replacing a broker’s trading stack with FPGA acceleration does not transform an older exchange core into a next-generation exchange.

Likewise:

A one-millisecond exchange-processing figure does not mean that an entire strategy-to-execution cycle takes one millisecond.

Professional trading systems care about the entire path.

23. This Is the Real Significance of G4

#

When large-scale G4 migration progresses further, the most important improvement will not simply be:

“One million orders per second.”

The more important goals are:

Lower tail latency
+
Better failure isolation
+
Stronger horizontal scalability
+
Faster recovery
+
Easier introduction of new services
+
Greater technological autonomy

SSE’s own public material describes G4 in terms of low latency, distributed architecture, open architecture, and autonomous controllability.

This tells us something important:

SSE itself recognizes that the long-term direction cannot simply be endless optimization of the 2009 architecture.

24. Final Conclusion: The Gap Is Not Simply “Shanghai Slow, Shenzhen Fast”

#

If this entire article had to be reduced to one sentence:

The real technological gap between China’s two major stock exchanges is not that Shanghai’s servers are slower than Shenzhen’s; it is that their core trading systems were born in different technological eras.

The NGTS generation represents:

Late-2000s architecture
+
Centralized trading core
+
High consistency
+
High reliability
+
Long operational maturity

The SZSE fifth-generation system represents:

Mid-2010s architecture
+
Distributed processing
+
Pipelines
+
Lock-free queues
+
Messaging infrastructure
+
Linear scalability
+
Low-latency design

And G4 represents the next stage:

2020s architecture
+
Distributed processing
+
Open architecture
+
Autonomous technology
+
Low latency
+
Modularization
+
Business decoupling

The historical sequence is therefore:

Early centralized trading
        ↓
Multiple generations of exchange platforms
        ↓
NGTS
        ↓
Fifth-generation SZSE architecture
        ↓
G4
        ↓
Next-generation low-latency infrastructure

Not simply:

Old system
VS
New system

25. The More Important Question for the Next Five Years

#

If SSE completes the large-scale migration of its core trading ecosystem to G4, the interesting question will no longer be:

“Which exchange has the higher TPS?”

It will be:

Which exchange can maintain the lowest P99/P99.9 latency, the strongest failure isolation, the shortest recovery time, and the most deterministic order processing under extreme market stress?

Because once a trading system reaches:

1,000,000 orders/sec

the number itself is no longer the hardest engineering problem.

The truly difficult requirement is:

1,000,000 orders/sec
+
No lost orders
+
No duplicated orders
+
No incorrect sequencing
+
No incorrect executions
+
Deterministic results
+
Fast recovery
+
Complete auditability
+
Correct clearing

That is what separates an exchange core from an ordinary distributed Internet system.

And this is why incidents such as the September 27, 2024 SSE execution-confirmation problem—and future observations of latency during market-opening bursts—should be interpreted within the broader history of exchange-system modernization.

The lesson is not:

“Shanghai’s system is broken.”

It is:

A core financial infrastructure born in 2009, after more than a decade of market expansion and business accumulation, has reached one of the most important architectural transition periods in its history.

SZSE completed a comparable generation change earlier, when it introduced its fifth-generation trading system in 2016.

SSE is now carrying out its own transition through G4.

Epilogue: China’s Exchange Infrastructure Is a History of How an Order Travels

#

There is an even larger story behind all of this.

China’s financial IT evolution can be viewed as a long chain:

Early Brokerage Counters
        ↓
Trading Seats
        ↓
Centralized Trading
        ↓
Mainframe / Minicomputer Systems
        ↓
Exchange Trading Hosts
        ↓
NGTS
        ↓
Fifth-Generation Distributed Systems
        ↓
G4
        ↓
Low-Latency Messaging
        ↓
FPGA / RDMA / Kernel Bypass
        ↓
Next-Generation Ultra-Low-Latency Infrastructure

Seen from this perspective, a few seconds of delayed execution confirmation are not merely an operational inconvenience.

They are a visible cross-section of a technological evolution that has lasted for decades.

The history of an exchange is ultimately not the history of its servers.

It is the history of:

how an order enters the financial system, how it is sequenced, how it is matched, how the result is distributed, and how the entire market proves that what happened actually happened.