Building an MNIST NPU with FPGA — From RTL to Bit-True Verification

When people decide to build a CPU, they usually start by looking for a circuit diagram. However, the moment they unfold the schematic, unanswerable questions arise: Why must this block be here? Why are there two ports for reading registers? The answers all lie within the instructions. Once you know what an instruction requires, the necessary circuitry is determined. It doesn't work the other way around. There are already many RISC-V lectures available. However, most of them simply display values on a slide, say "This is how it works," and move on. Students have no way to verify if those values are correct. This course provides a Python-based instruction execution simulator. When you type `run ADD`, the machine code is assembled on the spot, the 32 bits are laid out by field, and you can see exactly how the registers change on your screen. Every number in the textbook is a result of this tool's execution, and it can be reproduced identically on your screen. Commercial debuggers require licenses, and simulators often have installation barriers. This tool only requires Python. You don't need a board, a license, or a complex setup. This is a process of executing all 40 RV32I instructions without skipping a single one, discovering for yourself what the circuit needs to do.

When you study machine learning, you learn how to improve accuracy. But the moment you try to deploy that model onto an actual chip, you are left with questions no one ever taught you to ask: How many multiplications does this model require? What does it need to remember, and how much? Can the target chip specifications handle these demands? Machine learning textbooks focus on accuracy, while semiconductor textbooks focus on circuits, yet resources bridging the two are scarce. I repeatedly encountered this gap while implementing RISC-V CPUs and NPUs on FPGAs—facing situations where the model with the highest accuracy simply wouldn't fit on the board. In this course, you will build six algorithms from scratch in Python, measuring the number of multiplications, comparisons, memory usage, and accuracy for each. You will then translate these metrics into actual FPGA board resources. While knowing the number of multiplications alone doesn't give you a sense of scale, expressing it as a percentage of the board's resources makes it easy to assess. After completing nine labs and compiling the six models into a single table, an unexpected picture emerges. We spent nine sessions counting multiplications, only to find that multiplications weren't the actual bottleneck. From there, we move on to quantizing floating-point numbers to integers, managing bit widths, and even establishing criteria for grading circuits. An FPGA board is not required. The course can be completed using Python alone, with the board serving merely as a reference for translating the measured metrics.

To directly design AI semiconductors such as NPUs, Transformer accelerators, and mini LLMs, you must understand the theoretical reasons behind the operations occurring within them. This course distills the essential theories required for hardware design—from a single neuron to CNNs, Transformers, and LLMs—into a compact three-part series. Rather than memorizing formulas, you will gain an intuition for "why things work the way they do," as a current professor with 30 years of experience in the semiconductor industry bridges the gap between software theory and hardware implementation.

This course is for those who have learned Verilog syntax but are stuck wondering, "So, what should I build?" With just a single FPGA board, you will directly design four types of hardware: a digital clock, an ultrasonic distance sensor, an OLED display, and real-time camera video. The difficulty progresses from output to input/control, display, and finally video, allowing you to master digital design through hands-on experience with visible results. An active professor with 30 years of experience in the semiconductor industry teaches only content that has been verified on actual FPGAs.

Many people have learned FPGA and Verilog but get stuck at the question, "So, how do I actually build an AI accelerator?" This course resolves that curiosity. Using just a single entry-level FPGA and free tools, you will start from a hand-designed 3×3 kernel, train a CNN yourself, quantize the trained weights to INT8 to load onto the hardware, and verify that the results are Bit-True—not deviating by even a single bit from the Python model. As a professor with 30 years of experience in the semiconductor industry, I teach only verified content as 'working circuits' on actual FPGAs. By gaining the experience of designing and proving it yourself rather than just learning theory, you will acquire practical skills that are recognized in the fabless design and AI semiconductor fields.

Many people have learned FPGA and Verilog but get stuck at the question, "So, how do I actually build an AI accelerator?" This course resolves that curiosity. Using just a single entry-level FPGA and free tools, you will start from a hand-designed 3×3 kernel, train a CNN yourself, quantize the trained weights to INT8 to load onto the hardware, and verify that the results are Bit-True—not deviating by even a single bit from the Python model. As a professor with 30 years of experience in the semiconductor industry, I teach only verified content as 'working circuits' on actual FPGAs. By gaining the experience of designing and proving it yourself rather than just learning theory, you will acquire practical skills that are recognized in the fabless design and AI semiconductor fields.

Are you feeling overwhelmed by practical application and interview preparation? Based on my industry experience, I will help you firmly grasp everything from CNN to DETR, with a strong focus on code.

To directly design AI semiconductors such as NPUs, Transformer accelerators, and mini LLMs, you must understand the theoretical reasons behind the operations occurring within them. This course distills the essential theories required for hardware design—from a single neuron to CNNs, Transformers, and LLMs—into a compact three-part series. Rather than memorizing formulas, you will gain an intuition for "why things work the way they do," as a current professor with 30 years of experience in the semiconductor industry bridges the gap between software theory and hardware implementation.

This course is for those who have learned Verilog syntax but are stuck wondering, "So, what should I build?" With just a single FPGA board, you will directly design four types of hardware: a digital clock, an ultrasonic distance sensor, an OLED display, and real-time camera video. The difficulty progresses from output to input/control, display, and finally video, allowing you to master digital design through hands-on experience with visible results. An active professor with 30 years of experience in the semiconductor industry teaches only content that has been verified on actual FPGAs.

We will summarize the theoretical concepts of MNIST, the most fundamental dataset in computer vision, and learn how to implement it using Torchvision's API as well as how to implement a custom dataset class directly.

If Season 1 was about checking your skills and solidifying the basics, Season 2 is intensive practical training focused solely on 'passing and career transitions.' Beyond simple syntax or gate-level design, you will directly design top-tier, high-difficulty modules (CDC, Synchronous/Asynchronous FIFO, AMBA APB, SPI, I2C, FIR filters, etc.) that frequently appear in coding tests and actual workflows at major corporations and foreign companies. Do you need solid design capabilities that won't waver even under a recruiter's relentless follow-up questions? This course contains all the practical know-how and ultimate coding test problems from ETA, who successfully moved from a regional university to a global semiconductor firm. In this final season, complete your career transition free pass.

When people decide to build a CPU, they usually start by looking for a circuit diagram. However, the moment they unfold the schematic, unanswerable questions arise: Why must this block be here? Why are there two ports for reading registers? The answers all lie within the instructions. Once you know what an instruction requires, the necessary circuitry is determined. It doesn't work the other way around. There are already many RISC-V lectures available. However, most of them simply display values on a slide, say "This is how it works," and move on. Students have no way to verify if those values are correct. This course provides a Python-based instruction execution simulator. When you type `run ADD`, the machine code is assembled on the spot, the 32 bits are laid out by field, and you can see exactly how the registers change on your screen. Every number in the textbook is a result of this tool's execution, and it can be reproduced identically on your screen. Commercial debuggers require licenses, and simulators often have installation barriers. This tool only requires Python. You don't need a board, a license, or a complex setup. This is a process of executing all 40 RV32I instructions without skipping a single one, discovering for yourself what the circuit needs to do.

This is a lecture that studies Vision Transformer, one of the latest deep learning technologies, and implements a paper using Pytorch. Come experience the new future of the vision field with me!

This is a comprehensive machine learning project course that learns various useful machine learning regression and classification projects along with theory using Python and TensorFlow 2 based on Boston Marathon big data.

When you study machine learning, you learn how to improve accuracy. But the moment you try to deploy that model onto an actual chip, you are left with questions no one ever taught you to ask: How many multiplications does this model require? What does it need to remember, and how much? Can the target chip specifications handle these demands? Machine learning textbooks focus on accuracy, while semiconductor textbooks focus on circuits, yet resources bridging the two are scarce. I repeatedly encountered this gap while implementing RISC-V CPUs and NPUs on FPGAs—facing situations where the model with the highest accuracy simply wouldn't fit on the board. In this course, you will build six algorithms from scratch in Python, measuring the number of multiplications, comparisons, memory usage, and accuracy for each. You will then translate these metrics into actual FPGA board resources. While knowing the number of multiplications alone doesn't give you a sense of scale, expressing it as a percentage of the board's resources makes it easy to assess. After completing nine labs and compiling the six models into a single table, an unexpected picture emerges. We spent nine sessions counting multiplications, only to find that multiplications weren't the actual bottleneck. From there, we move on to quantizing floating-point numbers to integers, managing bit widths, and even establishing criteria for grading circuits. An FPGA board is not required. The course can be completed using Python alone, with the board serving merely as a reference for translating the measured metrics.

From core theories of deep learning and CNN to various CNN model implementation methods, and practical deep learning development know-how through real-world problems, If you want to become a deep learning CNN technology expert based on Pytorch, join us in this lecture :)

It goes without saying that AI artificial intelligence is the trend. But what can you use machine learning and deep learning for? Now, let's learn useful Internet of Things IoT Computer Vision projects that are used in daily life.

This lecture implements a DDR Memory Arbiter.

Let's create a web app that uploads an image and uses OpenCV to process it and detect faces.

UART is the most basic and widely utilized serial communication method, making it a very suitable learning topic for beginners new to FPGA design. In this course, we will thoroughly understand the communication principles and frame structure of UART, and based on this, we will proceed with directly implementing transmit/receive logic through RTL (Register Transfer Level) design. By implementing UART on an FPGA, you can systematically develop your digital circuit design skills and gain experience verifying its operation at the actual hardware level through simulation and verification. Furthermore, by learning the method utilizing the MicroBlaze soft processor and Xilinx UART IP, you will gain a broad understanding of processor-based system design flow and interface configuration. This will enable you to acquire comprehensive development capabilities encompassing both logic design and processor design.

This lecture implements a DDR Memory Arbiter.

Have you been getting stuck on linear algebra formulas every time you try to study AI? Instead of memorizing formulas, this lecture shows you visually how vectors and matrices actually move in space. By the end of the course, you will be able to interpret the mathematics behind neural networks, PCA, and SVD on your own.

AI Artificial Intelligence, it goes without saying that mobile is the trend. But how do you learn machine learning and deep learning and use them on mobile? Now, let's learn useful mobile Computer Vision projects that are used in daily life.

If you want to properly learn OCR technology that's truly used in practice, this one course is all you need! Aiming for over 98% accuracy even with unstructured documents and complex layouts, based on the latest SOTA models and real-world know-how, we build enterprise-level OCR projects together.

This is a lecture covering LLM theory and practical examples based on <Large Language Models, Just the Essentials!> (Insight, 2025).