Reborn in 2010, the AI ​​in my mind is too powerful.

Chen Yiran originally didn't want to look at academic papers and reports. It wasn't that these papers and reports were worthless, but for him, if there was any really exciting news, he could just look at the papers online. With Jasmine around, he only needed to open the page-turning script and then clear his mind and stare at the screen.

However, plans don't always go as expected. On the first afternoon, there was a special session on GPU graphics cards. At that time, all three major players would be present: Nvidia, AMD, and Intel.

That's right, Intel is now also a contender in the graphics card arena, and a very prominent one at that.

In 2008, they made a high-profile announcement of Project Larrabee, claiming that they would build the world's first x86-based GPU, directly challenge NVIDIA and AMD, and were very confident that they would win. They made a lot of noise and hype.

Intel invested billions of dollars and touted Hot Chips 2008 and Hot Chips 2009 for two consecutive years, so this year's event obviously also attracted attention.

Their idea is also wonderful: dozens of small x86 cores are tied together to run parallel computing, compatible with all x86 code, without needing to learn CUDA or OpenCL, which greatly simplifies the workflow.

Even without Jasmine's reminder, Chen Yiran knew they had definitely failed, because his impression of Intel graphics cards was from 2020.

However, Chen Yiran only thought that it might be a sales problem or a cost problem that prevented them from defeating Nvidia and AMD.

Although everyone around was talking about Intel's announcement in May that the schedule might be postponed, indicating that the project wasn't progressing well, they still believed that at least there would be some output, just perhaps not ideal, requiring more time for adjustments. Three years had passed; this meeting should at least showcase some results.

No one expected Intel to pull off such a big move this time.

The afternoon session on GPU graphics cards officially began, with NVIDIA starting first. Chief Scientist Bill Daly gave an in-depth analysis of the Fermi architecture.

Then they piled up a bunch of data and started emphasizing that this wasn't a graphics card, but a computing platform.

Although everyone noticed that Bill Daly deliberately avoided mentioning power consumption throughout the presentation, given that Nvidia is currently the only player in the GPU computing field, this performance can still be considered quite successful.

AMD was the second to take the stage, and their performance was even more impressive, as they started making grand promises while relentlessly attacking their competitors.

First up is the Fusion APU, which is highly innovative.

The CPU and GPU are integrated on the same chip and share a memory controller. Because data does not need to be exchanged through PCIe, graphics and computing can be switched seamlessly.

This was already very exciting, and then AMD sent out the second slide.

A review of the Evergreen architecture reveals that while its performance is excellent, they focused almost entirely on power consumption control, practically directly challenging Nvidia.

The atmosphere inside the venue became very lively, proving that watching a spectacle is human nature.

Everyone was already enjoying Intel's performance, and then the stage was empty for ten minutes, with no one coming up.

The host then went on stage looking embarrassed and said that Intel had canceled the presentation due to "internal adjustments".

The situation quickly spiraled out of control, with countless onlookers refusing to leave and instead discussing whether Intel had backed down after seeing the success of the other two companies.

Or have x86 GPUs completely failed? They can't even create a simple PowerPoint presentation anymore, not a single thing remains?

Moreover, why not cancel in advance? Canceling at the last minute is even more embarrassing.

Chen Yiran was also stunned. He knew that Intel was no longer as powerful as it used to be, but he didn't expect it to be so unreliable now.

There was no other way but to leave, and he started wandering around. He flipped through the schedule in the corridor; the data center and server CPU were both for the next day, and the industry reports in the main hall of the memorial hall no longer contained anything he wanted to see.

Retreating directly would be too early, so Hannah invited him to dinner and a movie that evening. He had originally planned to stay until five o'clock before leaving.

Seeing the posters at the entrance of the side hall, Chen Yiran felt that she could only look at academic-related things.

Since we're already here...

The side hall was filled with academic papers and reports. Unlike news from industry, many of the topics discussed in these papers and reports were difficult to implement.

Chen Yiran hadn't originally planned to listen to the academic paper presentations. It wasn't that these presentations were useless; in fact, every major company would send someone to follow them. This is because the ideas presented in these academic papers often outline the planning direction for the next five, ten, or even several decades.

But for him, if a paper gets a huge response, all he needs to do is download it, open the automatic page-turning script, and stare at the computer screen with his mind blank.

Among a pile of schedules, Chen Yiran saw an interesting term: Dark Silicon.

This was a paper by a young researcher at UCSD named Nathan Goulding. It was just the time for the presentation to start, and Chen Yiran went straight to the corresponding classroom.

There were already sixty or seventy people inside, mostly professors and graduate students, with a few industry professionals who had been drawn in by the title to listen to the student's thesis presentation.

Nathan Goulding dropped a bombshell right from the start: "Denard's scaling law will completely fail within five years."

This statement was not news in the chip industry in August 2010; those involved in the manufacturing process had already been aware of it around 2006.

But when it comes from an academic researcher addressing the entire chip design community, it sounds completely different; it always feels like a declaration.

What he went on to say left the sixty or seventy people in the audience speechless, and everyone could sense that the conclusions of this paper would influence the development of chips for decades to come.

Dennard's scaling law is a core principle in the semiconductor industry, stating that power density remains constant while transistors shrink. Proposed by IBM scientist Robert Dennard in 1974, it supported the rapid development of the semiconductor industry for over 30 years and is considered one of the three major laws of the semiconductor industry, along with Moore's Law and Amdahl's Law.

Before 65 nanometers, each new process node could pack more transistors into the same power budget.

However, the situation changed after 65 nanometers. Due to the extremely small size, the gate oxide layer in the transistor has been compressed into a very thin layer. At this scale, electrons can directly tunnel through quantum tunnels, which causes the power consumption in the off state to become out of control.

This results in a chip having enough transistors, but not enough heat dissipation capacity to allow them to operate at full speed simultaneously.

After all, the power consumption density of each new generation of chips is increasing, but the heat dissipation capability of the packaging has hardly changed.

The consequence was that as each generation of transistors was manufactured, an increasingly larger proportion could not be powered simultaneously. Goulding named this concept "dark silicon."

One slide in his PowerPoint presentation showed a prediction curve, with the horizontal axis representing the process node from 45 nanometers to 8 nanometers, and the vertical axis representing the proportion of transistors that can work simultaneously.

The curve plummeted from over 60 percent, and Goulding concluded that if nothing changes, by the 8-nanometer era, less than 20 percent of transistors on a chip will be lit at the same time.

Chen Yiran felt that there must be something wrong with this data, because there must be a solution, otherwise people would not have continued to pursue 5 nanometers or even 3 nanometers.

Jasmine then stepped in to explain.

[Goulding's predictive model uses the most conservative Dennard scaling collapse assumption; it doesn't consider the mitigating effects of adaptive voltage regulation and dynamic power gating. The actual dark silicon ratio at 8nm is around 20%, not below 20%.]

Chen Yiran was speechless: That's not right. Twenty percent is also terrible. If I understand correctly, dark silicon means it can't be lit up, right? It's useless, right? Is that possible? Future chips can't only be usable in two or three percent, can they?

The trend is correct. Around 2006, Dennard scaling went from being a theoretical bottleneck to a physical one. The proportion of transistors that can operate at full speed with each new generation of technology is decreasing, and this fact is indisputable.

At this point, Goulding came to the end and offered a solution, a concept: instead of making increasingly larger general-purpose chips, he would create a bunch of circuits that only do one thing and put them together.

General-purpose CPU cores had the worst energy efficiency during the Dark Silicon era, making dedicated accelerators the most cost-effective option.

Chen Yiran didn't know if Goulding was right. He left the memorial without lingering at Stanford and even canceled his date with Hannah to go home.

After arriving home, Chen Yiran excitedly turned on her computer and began communicating with Jasmine.

"I have an idea, although I don't really understand it. It's just a suggestion, but I'd like to hear your opinion and see if it's right."

【Say it already! (๑•̀ㅂ•́)و✧】

"Traditional placement only considers area and line length, right? If I understand correctly, can we add a new index, which is temperature?"

This temperature calculation allows for the distribution of high-power cells and the placement of cooler cells in the center, resulting in a more even temperature distribution across the entire chip and preventing overheating. Does this have some effect on reducing dark silicon? My limited understanding is that dark silicon is simply due to poor heat dissipation, right?

【_( ﾟДﾟ)ﾉ Great idea! And it's very simple and easy to do. Grey can directly calculate the three-dimensional temperature field and then perform global optimization to flatten the temperature gradient across the entire wafer.】

"Can you write it quickly? If you can finish it in a month, I can use it as a topic at Nvidia's conference. This is a blue ocean market; the other three major EDA companies don't have any related technology at all."

[Yes, and I've also thought of a good idea. Chips have more than three threshold voltage libraries: low Vt, standard Vt, and high Vt. Low Vt transistors are fast but have high leakage current, while high Vt transistors have low leakage current but are slow. We can configure this and use grey to calculate partitions, using low Vt to speed up critical paths and high Vt to reduce leakage current on non-critical paths. This reduces dark silicon at the source, using transistors with low leakage current in areas where it's not needed from the beginning.]

"Isn't the workload too heavy? Can you finish these two in a month?"

No problem, employees might need to work overtime.

"Alright, then let's go into seclusion and work on this."

Chen Yiran first opened the shopping website, preparing to order high-calorie foods such as chocolate and cream, in preparation for the long-awaited high-intensity use of jasmine.

Man proposes, God disposes. I had just finished an hour of intense work when TED called.

The company structure I prepared beforehand is finally finished. I have to go to the law firm tomorrow.