Introduction to AI Semiconductors: Mastering Everything with Excel, No EDA Tools Required
Did you give up while learning Verilog and EDA tools? A semiconductor professor who designs NPUs himself will show you how to understand everything from CPUs to NPUs by calculating with just 9 Excel files.
The ability to explain what bottleneck and bandwidth mean in AI semiconductor articles.
The ability to explain the differences between CPU, GPU, and NPU based on their structural reasons.
Answering ability to not get stuck when asked "Explain the difference between GPU and NPU" in an interview
Understanding the differences between DRAM and SRAM through hands-on practice
9 types of AI semiconductor practice Excel files
The ability to manually calculate and find the point where performance plateaus even when increasing the number of arithmetic units.
The explanatory power of why SRAM is not used as main memory despite being 40 times faster.
■ Without a semiconductor simulator, using only Excel
Attempts to understand AI semiconductors usually stop at a high barrier to entry. You have to learn hardware description languages, install EDA tools, secure licenses, and learn how to interpret waveforms. This process alone takes months, and most of your time is spent learning the tools before you even get a feel for the overall architecture.
This lecture offers a different path. It is designed to implement semiconductor behavior as a calculation model on a spreadsheet, allowing students to directly adjust input values and observe the resulting changes.
Just open the Excel file and change the numbers in the yellow cells.
■ Why Excel can replace EDA tools
The essence of semiconductor simulation is to observe how results change when conditions are altered. The practice files for this lecture perform that function within a spreadsheet.
· No installation required
Excel or a compatible spreadsheet program is sufficient.
There is no need for a separate development environment, license, or setup process.
· The entire calculation process is visible
EDA tools output results as waveforms. The intermediate processes are hidden.
In a spreadsheet, all intermediate values are displayed directly in the cells.
You can directly check which formula was used to produce that value.
· Responds immediately
If you change the value in the input field, the related tables and graphs are updated on the spot.
There is no compiling or waiting for execution.
· Remains even after completing the course
Even after the lecture ends, you can continue to use the files as they are.
You can continue your own experiments by inputting new conditions.
■ These are not static charts
The practice files for this lecture are not tables that merely show pre-calculated results. The execution process of the semiconductor is implemented through formulas, so if the input changes, the entire result is recalculated in a chain reaction.
· By unfolding the process of the CPU executing instructions into 27 steps,
Trace each line to see which values are read and accumulated at each step.
· The point where the GPU's arithmetic unit utilization reaches the bandwidth limit
Calculate it using formulas and verify it through graphs.
· The five stages of charge restoration every time the DRAM reads
Compare it with SRAM by going through it step by step.
· The location where calculations are activated in the systolic array
It tracks the process of moving diagonally over time on a cell-by-cell basis.
· While the NPU's 13×10 arithmetic unit array processes 52 tiles
It calculates the number of memory accesses occurring at each step.
This type of learning material is difficult to find on the market. It was created to fill the gap between introductory books that remain at the level of conceptual diagrams and metaphors, and professional texts that presuppose knowledge of hardware description languages.
■ Eight topics lead to a single problem
This lecture is not a list of independent topics. It is designed with a structure that starts from one question and deepens into the next.
What does artificial intelligence calculate?
What kind of burden arises when performing those calculations with a general-purpose processor?
How does the parallel structure alleviate that burden, and where does it reach its limits?
If memory is the constraint, why not just use faster memory?
What is the reason why high-speed memory cannot be fully adopted?
What changes when data is reused?
What benefits and costs arise when this is implemented as a circuit?
The questions remaining at the end of each session serve as the starting point for the next. Ultimately, you will reach a system design perspective that simultaneously considers performance, power, area, cost, and data movement.
■ Learning Method
Each session includes a process of changing input values. The recommended order is as follows.
First, predict the outcome
After changing the values
observe the actual results,
try to explain the reason why.
Entering values is not the goal itself. The goal is to understand the causal relationship between which inputs affect which results. When expectations and reality do not align, your understanding deepens in the process of finding the reason why.
■ Provided Materials
We provide all 9 types of Excel files used in the lecture.
01_Line_Classifier.xlsx Decision process of a linear classifier
02_MNIST_Learning.xlsx Principles of weight learning
03_Training_and_Inference.xlsx Comparison of computational load between training and inference
04_CPU_Operation.xlsx Instruction Execution and Pipelining
05_CPU_GPU_Comparison.xlsx Parallel Architecture and Memory Bottleneck
06_DRAM_SRAM.xlsx Principles of memory hierarchy formation
07_Systolic_Array.xlsx Data Reuse and Execution Schedule
08_NPU.xlsx Structural Characteristics of Dedicated Accelerators
Each file consists of guidance, main body, conditions, tables, and summary sheets. Input cells are distinguished in yellow, and other areas are protected. You can experiment freely without worrying about damaging the formulas.
■ Target Audience
· Those who come across materials related to AI semiconductors but
Those who wish to systematically organize their technical background
· While preparing for employment in the semiconductor industry
Those who wish to organize their understanding of computer architecture and accelerator concepts
· Those who are non-majors but wish to be able to judge technology trends
Those who wish to acquire foundational knowledge
· Before learning hardware description languages
Those who wish to grasp the overall structure first
· Those who need to explain AI semiconductors in an educational setting
■ Prerequisites
Being able to enter numbers into a spreadsheet is sufficient. No programming experience, semiconductor expertise, or separate development environment is required.
■ Scope of the Lecture Model
The calculation models used in this lecture are simplified for educational purposes and are not intended to predict the performance metrics of actual products.
The conditions assumed for each model are specified on each practice screen, and any differences from actual designs are explained separately during the lecture. The goal is not to memorize the numbers themselves, but to understand the structural reasons why such numbers are derived.
■ Lecture Composition
Total 9 sections · 9 lectures · 3 hours 34 minutes total
Section 1 and Section 2 are available for free.
Recommended for these people
Who is this course right for?
A semiconductor job seeker who wants to grasp the overall structure before learning Verilog and EDA tools.
AI developers who want to understand the relationship between model performance and hardware through numerical data
An electronics engineering student who wants to verify concepts learned in computer architecture with actual numbers.
Educators who need hands-on materials to use when explaining AI semiconductors
A semiconductor industry planner who wants to judge technical meetings and documents for themselves.
Need to know before starting?
It is enough if you can enter numbers into the spreadsheet.
No programming experience or specialized knowledge in semiconductors is required.
You will not have to write the formulas yourself. You only need to change the values in the blank cells of the Excel sheet.
"In the AI era, the surest response is to build AI semiconductors yourself."
This program is neither a theory-focused university course nor the standardized curriculum of a contract department tailored to specific specifications. Nor is it limited to simply gaining experience using EDA tools or leveraging libraries, as in conventional fragmented practical seminars.
Starting out as a systems semiconductor researcher in Samsung Electronics’ DS Division, winning the Silver Award (Group Chairman’s Award) at the Samsung Technology Awards, and serving as Director of Samsung’s UK and Germany subsidiaries and head of the System LSI Marketing and Sales Group, this unique, hands-on course was designed and FPGA-validated by a practicing engineering professor with 30 years of experience.
The author also provides EdgeAssembler and EdgeVerilog, free web-based tools he developed for hands-on practice. Use EdgeAssembler to convert RISC-V Assembly into Machine Code and Instruction Encoding and inspect the results, and EdgeVerilog to design and simulate RTL. (Free access: https://www.semiconductorschool.co.kr → click the 'LAUNCH' button)
[ Key differentiators for working professionals & aspiring engineers ]
Architecture design skills—rather than connecting off-the-shelf IP or modifying code, you’ll design the RV32I five-stage pipeline (IF·ID·EX·MEM·WB) directly in RTL, including data hazard forwarding, load-use stalls, and branch handling. Understanding and working directly with the five-stage pipeline is what sets your design skills apart.
From Algorithm to RTL — Implement an MNIST CNN as NPU RTL, starting with a Python model and applying Quantization, then verify that the Python and hardware results match bit-for-bit for the same input. The key is to achieve 100% identical results in Python and Verilog and produce results from actual FPGA operation.
CPU + NPU + RTOS integration on a single FPGA — Put everything on a single device: a CPU you designed running an NPU you designed, with FreeRTOS implemented by EdgeChipLab running on top. Students can use all these publicly available RTL and code resources to create their own semiconductor chips.
Free RTL design and simulation environment — With our in-house developed EdgeVerilog v1.0, you can write Verilog RTL, run simulations, and analyze waveforms on a PC or smartphone without signing up or installing any software. We provide a real simulation environment based on Icarus Verilog and 52 design examples for free, making it easy for anyone to get started with semiconductor design practice.
AI Algorithm → RTL Design → FPGA Implementation → Bit-True Verification → AI SoC
It provides an unparalleled, top-tier practical guide not only for undergraduate and graduate students, but also for working professionals looking to advance into next-generation AI semiconductor/SoC design.
Beyond theory to actual AI hardware implementation
It is difficult to develop system-level AI semiconductor design skills with only scattered, fragmented knowledge.
Students will not stop at using finished black-box IP. The goal is to progressively develop the system design skills to design an AI NPU and a 5-stage Pipeline RISC-V CPU themselves, integrate memory and peripherals, port an RTOS on top of them, and ultimately build their own AI SoC.
Starting with the fundamentals of transistors and digital circuits, this course connects FPGA, CPU, and NPU technologies to hardware for accelerating Transformers and LLMs.
We validate the research results on actual FPGA hardware.
This is not a lecture you only watch. Rather than stopping at theory or simple simulations, you will run AI semiconductor circuits you designed yourself on an actual FPGA and perform Bit-True verification, confirming that the Python model and RTL results match down to the last bit.
The instructor also connects research results, including four KCI-indexed papers on AI semiconductors, FPGAs, and RISC-V, to the educational content, taking technologies validated through research into actual design and FPGA implementation. The source code needed for the exercises—including a self-developed RISC-V CPU, Python code for image training, and Verilog RTL—is made available.
Built around the affordable Arty S7-25 FPGA board and the free Vivado design tool, this course enables learners to reproduce the entire process of designing, implementing, and verifying their own work. It can also be used with Digilent FPGA boards, including the Zybo Z7. Simply modify the XDC pin mapping (board) and clock division parameters (RTL) to match your board. The 'MNIST NPU with FPGA' course only requires modifying the XDC pin mapping, while courses that use an OLED and camera require modifying the clock division parameters (RTL).
→ The divider parameters for the course exercises are set based on the Arty S7’s 100 MHz input clock. Since other boards have different input clocks (for example, the Zybo Z7 has a 125 MHz input clock), you will need to recalculate the divider value so that the target frequency remains the same. If you miss this, the screen or video may not display properly,which can easily be mistaken for a code error, so please check.
Free RISC-V and Verilog Practice Environment — EdgeAssembler · EdgeVerilog
You can now start semiconductor design practice in your web browser without installing any software.
EdgeAssembler and EdgeVerilog, developed by the instructor, are free educational tools for hands-on practice. In EdgeAssembler, you can see how RISC-V Assembly is converted into Machine Code, and in EdgeVerilog, you can design and simulate RTL that processes those instructions. They run directly in your web browser, with no need to install separate EDA software.
Instruction Encoding Display: Select a Machine Code row to display its 32-bit binary representation and illustrate the opcode, rd, funct3, rs1, rs2, and imm fields.
Immediate Reconstruction: Reassemble scattered immediate bits, as in B-type and J-type, and display the offset and target address as well.
TinyRV32I Target: Tests tailored to the range of instructions supported by the textbook CPU
Encoding verification tool: Does not execute hardware; actual behavior is verified through EdgeVerilog’s RTL Simulation.
Key Features of EdgeVerilog
Actual RTL Simulation: Compile and simulate based on the open-source Icarus Verilog 12
Waveform Viewer: Observe signal changes, zoom, display HEX/DEC/BIN, position the cursor at signal changes, navigate to the previous or next change, and expand bus bits.
Example Library: Provides 52 digital circuit design examples, including counters, FSMs, and CPUs.
Project management: Write RTL and testbenches, autosave, and ZIP export/import
Mobile Support: Edit RTL, run simulations, and analyze waveforms on smartphones as well
Free of charge: Freely available for educational and non-commercial purposes.
Using the two tools together, learners can explore everything from introductory Verilog-HDL and FSMs to AXI4 and RISC-V CPU design, and verify everything from instruction encoding to RTL.
EdgeAssembler is a tool for checking instruction encoding, and EdgeVerilog is an educational tool for RTL functional verification. FPGA synthesis, implementation, and timing verification are performed in Vivado.
A seamless full-stack roadmap: from fundamentals to a mini LLM accelerator
Each lecture is part of a journey toward completing one vast system, not an isolated fragment.
Step 1: Fundamentals of AI Theory and Image Processing (Including Machine Learning)
Step 2: AI Accelerator (NPU) Design and Verification (Including Machine Learning)
Step 3: 5-Stage Pipeline RISC-V CPU Design and System Integration
Step 4: Advanced AI SoC Implementation and Expansion into a Mini LLM Acceleration Platform (Ongoing Content Updates)
Objective Verification Metrics
All design deliverables from this course have undergone rigorous validation against global standards and by the academic community.
5-Stage Pipeline RISC-V Architecture Verification: In-house developed RISC-V CPU passed the international foundation’s official compliance test (ACT), and its source code was released (GitHub)
Academic Recognition: Four sole-authored academic papers published in IJIBC, a prestigious KCI Grade A journal.
Global Recognition: Published four global Amazon technology books (reached #1 bestseller status)
Real-World Operation Verification Complete: Core IPs, including the RV32I CPU, NPU, vision system, GPS, Transformer, and AURA-Edge SoC, operate flawlessly in the Arty S7 environment.
Take on the challenge. By the end of this process of loading your own code onto the board and verifying the results, you will have leveled up into a hardware engineer with a completely different perspective.