King of Technology: A Fresh Start in 1983

Galaxy Supercomputing!

Parallel processing!

This is a challenge from a completely different dimension!

"Memory bandwidth...parallel task scheduling..."

Professor Li Mingyuan frowned. "This requires the processing unit to have extremely high memory access efficiency and low-latency communication capabilities!"

"Our Pangu Card bus driver architecture, while optimizing local memory access, still falls far short when facing multi-node parallelism!"

"The core bottlenecks lie primarily in the on-chip caching architecture and the consistency protocol!"

Zhou Yuguang then hit the nail on the head, "Our current technology doesn't even have a decent L1 cache!"

"Not to mention multi-core shared cache and consistency maintenance! That's an architectural gap!"

The technology gap is looming ahead!

Everyone felt heavy-hearted.

"No matter how big the gap is, we have to find a way to cross it!"

Chen Jianghe's eyes were intense. "Engineer Su called us out not to make us stand by and watch, but to get us to come up with solutions to the problem!"

Even if it's just a prototype, even if it only solves one problem!

He spread out paper and pen: "Now, let's disassemble our 'Pangu' framework! Let's nitpick and find potential even in the smallest details!"

It was still the same heavy explosion-proof door, the cold air, and the solemn atmosphere.

The difference this time was that he was taken directly to a medium-sized conference room.

Several people were already seated in the conference room.

Besides the familiar Chief Engineer Zhang Zhenggao and Su Wanqing, there were several other technical experts with calm demeanor and sharp eyes, who were clearly core members of the supercomputing project.

The atmosphere was solemn and focused.

"Little Chen is here? Have a seat."

Chief Engineer Zhang nodded slightly, his gaze sweeping over the thick stack of documents Chen Jianghe had brought. "Wanqing said you brought some ideas? Tell me about them."

Chen Jianghe took a deep breath, walked to the projector, and projected the hand-drawn concept sketch that Zhou Yuguang had worked on overnight.

He tried his best to explain his design ideas in clear language.

The focus is on the potential of the Pangu architecture in terms of memory access efficiency (dual-channel) and task scheduling (multi-threaded distribution).

And the communication advantages brought by the integrated bus driver IP.

Regarding caching and consistency protocols, Chen Jianghe frankly stated, "This is Tiangong's biggest weakness; we have only proposed an extremely simplified, preliminary idea..."

During the presentation, the conference room was completely silent.

Several experts frowned, taking notes in their notebooks from time to time.

Su Wanqing sat in the corner, her gaze calm and composed, revealing no emotion whatsoever.

After finishing his explanation, Chen Jianghe's palms were sweaty.

"The idea... is very bold."

An elderly expert with gray hair spoke first, "The idea of ​​extending the local experience of graphics coprocessors to general-purpose parallel computing is very inspiring."

"The dual-channel memory access design meets our need for increased bandwidth. However…"

He then changed the subject, saying, "The on-chip cache is too small, and the consistency protocol is too idealistic."

"The dual-thread model of the task scheduling core is woefully inadequate in an environment where supercomputing nodes often have dozens or hundreds of threads."

"The most crucial point is... the power consumption budget and process integration level cannot support your vision."

The doubts came like a bucket of cold water poured over their heads.

"Engineer Liu is right."

Another middle-aged expert commented, "The core problem lies in the fact that the system architecture is too low-level."

"Your design is still at the level of simply piling up functional modules."

"There is a lack of systematic optimization of the nature of parallel computing, data locality, and communication overhead."

"This plan... falls far short of [Galaxy's] needs."

The meeting room fell silent once again.

Chen Jianghe's heart sank to the bottom.

He knew there was a big gap, but he didn't expect to be criticized so harshly.

Just then, Su Wanqing, who had been silent all along, spoke, her voice high-pitched and clear:

"The gap does exist. But their approach has grasped the key pain points: memory access efficiency and communication latency."

"This is the Achilles' heel of parallel computing."

She stood up, walked to the projector, and picked up a red pen…

In Chen Jianghe's concept diagram, the sections on "dual-channel memory access interface" and "bus driver IP" were heavily circled.

"These two points are the core values ​​of the Pangu architecture, and also the directions in which we are most likely to make breakthroughs under our current technological conditions."

She glanced at the experts and said, "I believe that the initial node processors of [Galaxy] do not need to pursue full-featured multi-core processors."

"A coprocessor focused on efficient memory access and low-latency communication..."

"Works in conjunction with the main CPU to perform specific computationally intensive tasks, such as matrix operations, FFT, etc."

"That might be a more realistic approach!"