"They independently create texts and images, recognize faces and speech, make medical diagnoses, and reliably predict the weather – self-learning artificial intelligence algorithms are now indispensable in everyday life. They even take over the writing of program code, thus making classic programmers redundant. There is no end in sight to this development. But the intelligence of language models like ChatGPT comes at a price: their enormous energy consumption. And the more intelligent the models become, the more energy-hungry they are, lamented Chinese physicist Lu Fang from Tsinghua University in Beijing in her Breakthrough lecture last Sunday.
The figures she presented are alarming: Every day that approximately 100,000 people worldwide use ChatGPT and produce an average of 50 billion words and ideas, so much electricity is consumed that an electric car could drive around the equator 800 times non-stop, or an average household could be supplied with electricity for 275 years. "That's about 2.9 million kilowatt-hours of electricity."
But can we make artificial intelligence more sustainable, and if so, how? Yes, and the answer is with light, was Fang's response. Photons move effortlessly through space, without resistance and without generating heat, unlike electrons, which dominate digital computers. Light pulses can be modulated, encoded, guided, shaped, amplified, and superimposed. In this way, countless multiplications and additions can be realized – computational operations without which no artificial intelligence can function, for example, in image and speech recognition.
In her lecture, Li Fang presented the result of her research group's work: the Taichi photonic chip. It incorporates countless optical waveguides and a network of tiny interferometers. These optical components weaken, amplify, or extinguish the incoming light pulses and essentially act like analog switches. They control whether a light signal is transmitted or not. The advantage of this concept: networked interferometers are capable of performing almost any mathematical operation, particularly matrix multiplication. This computational operation is fundamental for neural networks, which allow data to be processed in parallel, modeled after biological neurons and synapses. These processors do not require data storage like their electronic counterparts.
Taichi reportedly has one million interconnected artificial neurons and is capable of reliably performing complex tasks. After appropriate training, the photonic processor can produce music clips in the style of Johann Sebastian Bach or generate landscapes in the style of Vincent van Gogh and Edvard Munch. The system has also proven to be extremely efficient and reliable in recognizing and displaying handwritten characters.
Compared to conventional processors, Taichi can compute 100 times faster. "It's like a train journey from Berlin to London taking only six minutes instead of ten hours," says Fang. Taichi is also unbeatable in terms of energy consumption. During training and task execution, the processor consumes only one-thousandth of the energy of the most efficient graphics processors currently available, Lu Fang said in Berlin. "Imagine charging your mobile phone once, and instead of lasting only one day, it now runs for three years thanks to Taichi."
For Fang, optical neural networks are no longer just experimental models. They can now be used in the real world. "Drones can think, autonomous robots can collaborate seamlessly with humans. And autonomous vehicles can perceive and decide what's happening around them at the speed of light. Fang's vision: AI that perceives, adapts, and learns at the speed of light, and does so in an energy-efficient and resource-saving manner. 'We call this idea regenerative green AI,' says Fang.” [1]
The Taichi photonic chip, developed by researchers at Tsinghua University in China, represents a significant breakthrough in optical computing, demonstrating over 1,000 times higher energy efficiency compared to NVIDIA's H100 GPU in specific AI tasks. While it is a "real" advancement that has moved beyond "toy" models to demonstrate actual AI generation and recognition, it is not yet a broad, plug-and-play competitor to NVIDIA for general-purpose computing.
Here is a breakdown of the current status of the Taichi photonic chip:
1. Is the Taichi Application Practical Already?
Proof-of-Concept, Not Commercial Product: Taichi is currently a, "world-first" research breakthrough, not a consumer product you can buy today. However, it is considered a "practical solution" to photonic computing, rather than just a theoretical model, according to a Science journal report.
"Lab" Stage: It is still in the experimental stage and requires bulky peripheral systems (lasers, couplers) that can "take up a whole table," rather than being a single, self-contained chip.
Specific Capabilities: It excels in matrix multiplication and convolutions—the most energy-hungry parts of AI—achieving 92% accuracy in image classification and, as of Aug 2024 (Taichi-II), showing improved speed in training networks with millions of parameters.
The Future: It is designed to act as an AI accelerator, working alongside traditional CPUs/GPUs, rather than replacing them immediately.
2. How Much Does it Cost?
No Commercial Price: Because it is a lab-produced research chip, there is no public market price.
Development Cost: Photonic chips are expensive to produce in small batches. While specific Taichi costs are not public, similar high-end Active Photonic Integrated Circuits (PICs) can cost over $10,000–$300,000 for small, customized batches (per 20-piece sets).
Lower Operating Cost: While the initial hardware may be costly to build, its main economic advantage is drastically reduced energy consumption, consuming nearly zero energy for calculations.
3. Is it Real Competition for Nvidia Chips?
Yes, in Specific Areas: For specialized AI tasks, especially large-scale inference and image generation (AIGC), Taichi is a major competitor, offering 100x to 1,000x better energy efficiency and higher speeds than NVIDIA’s H100.
No, Not as a General Replacement: Photonic chips are specialized for specific,, "narrowly defined" workloads, whereas Nvidia GPUs are highly flexible, general-purpose chips. Taichi is unlikely to replace Nvidia for general-purpose programming in the near future.
Advantage in Power/Heat: Because Taichi uses light, it generates far less heat than electric chips, addressing the primary bottleneck (overheating) in large data centers.
Competition Context: It is seen as China's "strategic answer" to US chip controls, allowing them to potentially leapfrog in AI acceleration without needing the latest advanced photolithography tools required for traditional 3nm/5nm chips.
Summary Table: Taichi vs. Nvidia H100
Feature Taichi Photonic Chip Nvidia H100 GPU
Technology Light (Photonics) Electricity (Electrons)
Status Research/Lab Prototype Commercial/Mass Production
Energy Efficiency ~160 TOPS/W (Reported 1000x+ better) High, but lower than optical
Best For Specific AI Inference/AIGC General-purpose AI training/inference
Heat/Power Minimal heat, very low power High heat, high power consumption
Note: Taichi-II, a newer iteration, has shown even greater efficiency,, "expediting the training of optical networks containing millions of parameters by an order of magnitude".
What do they need to make Taichi-II able to do General-purpose AI training/inference?
To enable Taichi-II for general-purpose AI training and inference, it requires
substantial scaling of its optical chiplet architecture to handle massive, complex workloads. Key needs include implementing Fully Forward Mode (FFM) learning on-chip for direct training, creating modular, interconnected chiplet systems for large-scale, high-bandwidth computing, and improving on-chip memory for storing large models.
Here are the specific requirements based on the development of Taichi-II:
- Massive Scalability & Modular Architecture: To move beyond specialized tasks to general-purpose AGI, the Taichi chiplets must be scaled up, with multiple modules operating together to create a more powerful, integrated computing system.
- Fully Forward Mode (FFM) Learning [2]: Unlike traditional AI which uses backpropagation (requiring high energy and electronic computation), Taichi-II utilizes FFM. This allows the computer-intensive training process to occur directly on the optical chip, facilitating efficient, high-speed learning.
- High-Speed Interconnects: To achieve general-purpose capabilities, the optical chips require robust, high-bandwidth connections between modules to handle data movement, ensuring that the speed of light computing is not bottlenecked by electrical interfaces.
- On-Chip Memory & Reconfigurability: Taichi-II uses diffractive components for calculation and interferometric (Mach-Zehnder) elements for reconfigurability. Enhanced on-chip memory is needed to hold large models and data, reducing reliance on slow, external storage.
- High-Speed Data Conversion: While optical computing is fast, efficient and high-speed electrical-to-optical (and vice versa) converters are necessary for the input/output of data.
Taichi-II's strength lies in using light for computation, offering superior energy efficiency (160 TOPS/W160) compared to traditional GPU-based systems.
1. Lichtblick für die grüne KI: Der photonische Chip Taichi verarbeitet Lichtpulse nach Art des Gehirns: ungeheuer effizient. Frankfurter Allgemeine Zeitung; Frankfurt. 13 Nov 2025: B4 MANFRED LINDINGER
2. Fully Forward Mode (FFM) Learning is an onsite training architecture for optical neural networks (ONNs) that implements the compute-intensive training process directly on physical optical systems. Developed by researchers at Tsinghua University and demonstrated on the Taichi-II chip, FFM overcomes the limitations of traditional "in silico" (digital computer) emulation, which often fails due to the complexity of modeling physical imperfections in optical hardware.
Core Principles
- Onsite Gradient Descent: Unlike standard back-propagation which requires complex digital modeling of light fields, FFM conducts machine learning operations on-site by leveraging spatial symmetry and Lorentz reciprocity. (Lorentz reciprocity is a fundamental principle in electromagnetics and physics stating that the relationship between an oscillating source and its resulting field is unchanged when the positions of the source and the observer are interchanged. Simply put, if you can see someone's eyes, they can see yours—the path for light or electromagnetic waves is generally identical in both directions.)
- Elimination of Back-propagation: By using forward light propagation to calculate gradients based on measured output error fields, it removes the need for a separate backward pass, allowing optical parameters to be self-designed directly on the physical system.
- High Efficiency: It achieves state-of-the-art energy efficiency, performing operations with light intensity as weak as sub-photon per pixel (approximately 5.40×10185.40 cross 10 to the 18th power operations-per-second-per-watt).
Key Performance Advantages (as of 2026)
- Large-scale Scaling: FFM supports the training of deep optical neural networks (ONNs) with millions of parameters, reaching accuracies equivalent to ideal digital models that previously could not be matched by physical systems.
- Speed: It facilitates learning processes that are orders-of-magnitude faster than traditional methods, enabling real-time adaptations.
- Imaging Resolution: It enables focusing through scattering media at the diffraction limit and allows for parallel imaging of hidden objects outside the direct line of sight at kilohertz frame rates.
- Model-free Discovery: It can automatically search for non-Hermitian exceptional points in topological photonics without requiring an analytical physical model. Non-Hermitian exceptional points (EPs) in topological photonics are specialized degeneracies in open optical systems where both eigenvalues and eigenvectors coalesce, breaking traditional Hermitian, orthogonal, and real-eigenvalue rules. Appearing in active photonic systems with gain and loss (e.g., ring resonators, waveguides), EPs enable unprecedented sensitivity, topological edge states, and unidirectional light transport, forming a new paradigm for manipulating light.
Primary Applications
- Deep Optical Neural Networks (ONNs): Achieving high classification accuracy (e.g., on MNIST and Fashion-MNIST datasets) that rivals digital counterparts.
- Non-Line-of-Sight (NLOS) Imaging: Recovering images of objects hidden around corners with high photon efficiency.
- Integrated Photonic Circuits (PICs): Implementing self-training on silicon-on-insulator
chips for tasks like iris flower classification.
Estimate how long will take for the Chinese with Optical Neural Networks to outcompete NVIDIA?
Timeline Projection: If the rapid development of photonic AI platforms continues, commercial viability or niche adoption in Chinese data centers could occur by 2026–2028, but fully outcompeting NVIDIA’s entire ecosystem across all AI tasks will likely take much longer.
While optical AI provides a potential "shortcut," it is not expected to completely displace NVIDIA's, market-leading GPUs in the immediate short term (1-2 years), but rather challenge them in specialized, power-intensive inference tasks.
Komentarų nėra:
Rašyti komentarą