radiothebook Chapter 205 Iterative Upgrades
Chapter 205 Iterative Upgrades
Clinical data is accumulating faster than expected.
Postoperative follow-up data from the first batch of ten patients gradually returned: signal calibration logs, motor function scores, and daily activity records—each item was being accumulated on Shen Yiming's server. He spent three weeks organizing this data into a structured dataset, and then said to Zuo Cheng, "This data is enough for us to do a major version iteration."
Zuo Cheng asked, "Where do you plan to start?"
Shen Yiming opened his laptop, displaying an architecture diagram of the NX-30 chip. He said, "Currently, the biggest bottlenecks are in two areas. One is the number of channels; 1024 sampling points only cover the core area of the motor cortex, failing to capture activity in the peripheral areas. The other is the decoding latency; it takes fifty milliseconds from signal acquisition to outputting control commands, which is significant for delicate finger operations."
He said, "I set three goals for the team: double the number of channels to 2048, reduce the decoding latency to below 20 milliseconds, and shrink the chip size by another 40%."
Zuo Cheng looked at the diagram he had drawn and said, "We can achieve the first two goals through algorithm optimization and data accumulation, but the third goal requires a redesign of the chip architecture."
Shen Yiming said, "That's right. The current architecture of the NX-30 separates data acquisition and signal processing into two steps, with a conversion overhead layer in between. If this layer can be removed, the size can be directly reduced."
Zuo Cheng asked, "How do we remove it?"
Shen Yiming said, "The core idea is to have the acquisition unit have built-in preprocessing logic, so the output is a usable signal directly, without the need for further conversion. However, this will increase the chip's power consumption, and we will have to redesign the thermal interface."
Zuo Cheng asked, "How low can the power consumption be reduced?"
Shen Yiming said, "The goal is to reduce the power from 50 milliwatts to 35 milliwatts while doubling the number of channels."
Zuo Cheng thought for a moment and said, "I've seen similar architectures before."
Shen Yiming looked up at him.
Zuo Cheng didn't explain. He opened the technology tree panel in his mind. The branch in the direction of the brain-computer interface had already grown four leaves, and the fifth was taking shape. On the system panel, the accumulation rate of the eighth branch had exceeded sixty percent.
He used his mind to select the cross-branch fusion entry point, connecting the neural network compilation and optimization capabilities in the AI field with the brain-computer interface field.
System notification: Cross-branch fusion requirement detected. Objective: Preprocessing logic built into the acquisition unit. Fusion direction: Neural network compiler optimization plus brain-computer signal decoding. Estimated cost: 5 points; fusion success rate: 92%.
Zuocheng confirmed.
On the panel, two streams of light of different colors converge at one point, forming a new technological node. The system prompts: Fusion successful. "Integrated Neural Acquisition and Decoding Architecture" achieved. Transmission latency reduced from fifty milliseconds to twenty milliseconds, size reduced by thirty-eight percent, and power consumption controllable.
He turned off the panel and said to Shen Yiming, "If you can integrate signal acquisition and preprocessing at the chip level, using heterogeneous computing to integrate the three steps of signal amplification, filtering, and digitization into one module, you can reduce power consumption."
Shen Yiming drew a few lines on the paper and said, "You're talking about chip-level heterogeneous integration. If this path is successful, the size can be reduced significantly, but the architectural complexity will be much higher."
He said, "If we can really integrate the three steps into a single module, along with your approach, the latency should be reduced to twenty milliseconds."
Zuo Cheng said, "Twenty isn't enough; the target is fifteen."
Shen Yiming said, "It's a bit radical."
(Please remember 1110.com for the fastest chapter updates)
Zuo Cheng said, "We were ahead in the first generation, and we will crush them in the second."
Shen Yiming and his team embarked on a new round of research and development. Tang Ning was responsible for the reconstruction of the signal acquisition layer, while Chen Minghui came from Xingchen Technology to be stationed on-site, specifically to redesign the chip layout. The new architecture was no longer the traditional acquisition and forwarding, but instead, some parameters of the neural network model were directly written into the chip's firmware layer, allowing preprocessing decisions to be completed at the hardware level.
There are many specific technical issues to solve. How to set the signal amplification factor, how to adjust the filter parameters, how many bits of precision to use in the firmware layer calculations—each item needs to be tested repeatedly.
Six weeks later, the first prototype chip was produced.
Test data began flowing out in the afternoon. All 2048 sampling channels were successfully connected, and the signal quality did not decrease despite the doubling of channels. The decoding accuracy reached 95.3% on the first run, and Shen Yiming stared at the screen in a daze for a few seconds.
The second run, 96.1%.
The third run, 96.8 percent.
Shen Yiming sent a message to Zuo Cheng with just four words: "968."
Zuo Cheng went to the lab the next day. The new NX-40 chip was nearly 40% smaller than the NX-30, and about the thickness of a coin. 2048 microelectrodes were arranged in a matrix, with the spacing between each electrode being denser than the previous generation. On the test bench, the new chip was running.
Shen Yiming said, "The latency has dropped to 14.7 milliseconds. The number of channels has doubled, and the latency has decreased by nearly three times."
Zuo Cheng read the report and said, "This isn't an improvement; this is a complete overhaul."
Shen Yiming said, "There's something else I want to tell you. This change in chip architecture has opened up a new possibility. Previously, the NX-30 was a unidirectional acquisition system, only able to read signals from nerve endings. In the new architecture, each channel is bidirectional, meaning it can not only read but also, theoretically, write."
Zuo Cheng asked, "What does 'writing' mean?"
Shen Yiming explained, "It involves sending electrical signals to the nervous system. If precise programming can be achieved, patients will not only be able to control their exoskeleton but also receive tactile feedback. When they pick up objects, they will be able to feel the hardness and temperature of those objects."
Zuo Cheng was silent for a few seconds. He said, "This is more important than decoding."
Shen Yiming said, "I know. That's why I haven't told anyone yet. The security and ethical issues involved in write functionality are too great; we can't do it rashly."
Zuo Cheng said, "First, we need to complete the safety verification, clearly defining the minimum safe current, frequency range, and tissue tolerance for writing. Once the standards are set, then we can implement the writing function."
Shen Yiming said, "Okay."
Two months later, the paper was published.
The peer review process for *Nature Neuroscience* took three months, with two of the three reviewers giving it an "outstanding" rating. The paper is titled "High-Density Brain-Computer Interface Chip Architecture for Bidirectional Neural Interfaces," with Shen Yiming as the first author and Zuo Cheng as the corresponding author.
On the day of publication, Shen Yiming's phone rang all day. More than a dozen laboratories worldwide sent collaboration invitations, and five international academic journals requested interviews. A professor at Harvard Neuroscience Center publicly commented, "The bidirectional architecture proposed in this paper is likely one of the most important technological directions in the field of brain-computer interfaces over the past five years."
Zuo Cheng put down his phone after reading the comments.
He told Shen Yiming, "This paper was published a bit too early. The two-way interface is just our first step. The real work is much more extensive than that."
Shen Yiming asked, "What is it?"
Zuo Cheng said, "Rest for a few days. When you come back, I'll talk to you about consciousness."