We are proud to share that Takuto Ando, a master’s student (M1) at the Computing Architecture Lab, has received the IEEE CEDA All Japan Joint Chapter SASIMI Young Researcher Award at the 26th SASIMI 2025 international workshop, held in Nara, Japan, on October 9–10, 2025.
The Workshop on Synthesis and System Integration of Mixed Information technologies (SASIMI) is an international forum for industry–academia exchange in system design, semiconductor design automation (EDA), and design methodologies. The Young Researcher Award honors outstanding poster presentations by researchers under 35.
Mr. Ando’s presentation, titled “A Detailed Analysis of LLM Execution on IMAX3 and Initial Evaluation of IMAX4 Prototype for Server Environment,” was highly acclaimed. His research proposes the IMAX4 prototype, an advanced accelerator based on the power-efficient CGLA architecture, achieving significant reductions in host-side processing load for large language model (LLM) tasks such as LLaMA3 8B.
This recognition underscores the strength of LENZO’s ongoing collaboration with academic partners in advancing next-generation computing architectures.
On Sep 19, 2025 NAIST Computing Architecture Lab's Youtube channel published a explainer video about the latest developments of the underlying technology LENZO is using. From the description; "This video is additional episode to the hungry engineer series. I present this video to all hungry engineers who are tired of CPU, GPU, FPGA, tensor core, AI core, who want some challenging one with no black box inside, and who want to improve by themselves."
On August 27th, Y Kobayashi wrote an article on XenoSpectrum about Lenzo, the original article is available only in Japanese, and can be found here. Below follows a translation to English:
Could this signal a breakthrough in the stagnation of semiconductor design that has lasted for 30 years? This article delves into the core of this technology and its vision.
Built with Expertise: Lenzo envisions the next generation of computing
Japan's semiconductor industry has long since lost its former glory. However, Lenzo, a certified startup from the Nara Institute of Science and Technology (NAIST), has stepped forward with technology that has the potential to overturn conventional wisdom in the industry.
CEO Kenshin Fujiwara previously worked at Sony Computer Entertainment (now Sony Interactive Entertainment) developing the CPU/GPU for PlayStation 2 and 3. Co-founder and chief architect Professor Yasuhiko Nakajima is a true authority on computer science, having been involved in the development of supercomputer processors at Fujitsu.
What they are proposing to the world is a completely new semiconductor architecture called CGLA (Coarse-Grained Linear Array) .
What is "CGLA"? A brief introduction to a new approach
To understand the innovation of CGLA , we must first look at the fundamental challenges facing modern computers.
Why is a new architecture needed now?
Most of the processors (CPU/GPU) in the PCs and smartphones we use every day are based on a basic design known as the von Neumann type, which separates the "arithmetic unit" that performs calculations from the "memory" that stores data.
While this structure is highly versatile, it suffers from a few weaknesses. Every time an operation is performed, a huge amount of data must be transferred between the processing unit and memory, and this transfer itself consumes a large portion of the power, creating a bottleneck in processing speed. This basic design has remained virtually unchanged for the past 30 years. Improvements in processor performance have been achieved primarily through the miniaturization of semiconductors, but this is also approaching its physical limits. In today's world, where the amount of data handled is exploding, as seen in AI and blockchain, this "von Neumann bottleneck" is a major hindrance to technological evolution.
Beyond the limits of CGRA and Systolic Array
To overcome this barrier, various architectures have been proposed, most notably the "Coarse-Grained Reconfigurable Array (CGRA)" and the "Systolic Array ".
CGRA: Its distinguishing feature is its flexibility, allowing its circuit configuration to be dynamically rewritten. However, this flexibility also comes at the cost of a fatal drawback: it takes several hours to several days to reconfigure (compile) the circuit.
Systolic Array: Used in Google's TPUs and other applications, it demonstrates high performance by specializing in AI matrix operations. However, because it is too specialized for a specific process, it lacks flexibility and is unsuitable for non-AI calculations such as blockchain.
Lenzo's CGLA was developed to solve these problems. They have succeeded in combining the reconfiguration flexibility of CGRA, the high speed of Systolic Array, and the concept of near memory computing, which minimizes data transfer to and from memory, into a single architecture.
The heart of CGLA: three innovative ideas
The superiority of CGLA is theoretically supported in an academic paper published by the Nakajima Laboratory at NAIST entitled "CGLA: Coarse-Grained Linear Array for Multi-Hash Acceleration in Blockchain Mining ". At its core are the following three ideas:
Self-updating data method: Blockchain mining involves a continuous brute force search for a value called a "nonce." In conventional methods, the computation unit must query the host PC for the next nonce value each time it attempts to find it, which creates a bottleneck in communication. CGLA implements a function that allows the computation unit (PE: Processing Element) to automatically update the nonce value itself. This dramatically reduces communication with the host PC, allowing the computation unit to focus on calculations.
Scalable linear PE array: The extremely simple structure of arranging the PEs, which are the processing units, in a one-dimensional linear fashion is adopted. This not only simplifies design but also provides high scalability, making it easy to scale performance by increasing the number of PEs as needed.
Dedicated ALU design: The arithmetic unit (ALU) built into each PE is optimized for hash function calculations. It is designed to be able to perform multiple calculations in one clock cycle, boasting extremely high calculation efficiency.
Data reveals the power of CGLA
CGLA isn't just a theoretical concept - Lenzo has already run a prototype on the TySOM-3A FPGA (a semiconductor whose circuit configuration can be reprogrammed after manufacturing) board, and has provided data to prove its capabilities.
High Hash rate that leaves the competition behind
In terms of hash rate, a performance indicator for cryptocurrency mining, CGLA surpasses the market leaders:
Product name
Process Node
TH/s (Terahashes per second)
Bitmain Antminer S21 Pro
5nm
234
Lenzo Core
16nm
196.6
Lenzo Core
7nm
619.29
(Source: Lenzo)
It is noteworthy that while Bitmain's latest mining machine, the Antminer S21 Pro, achieves 234 TH/s using the cutting-edge 5nm process, Lenzo Core achieves 619.29 TH/s , approximately 2.6 times that figure, using the previous generation 7nm process.
Even the older 16nm process shows performance approaching that of the latest 5nm machine. This demonstrates the overwhelming superiority of the architecture itself, rather than the superiority of the process technology.
Up to 8.7x faster than GPUs and up to 44.5x faster than conventional CGRAs
Even more impressive than its performance is its energy efficiency. According to the aforementioned research paper, in an ASIC simulation using a 45nm process, CGLA outperformed existing technologies by an order of magnitude.
Relative to GPUs: 2.8x to 8.7x more energy efficient than NVIDIA's high-end GPUs .
Compared to conventional CGRAs: Compared to previously studied CGRA architectures, we have achieved up to 17.8 times the throughput and up to 44.5 times the energy efficiency.
This is a crucial competitive advantage in the mining industry, where the cost of electricity for computation directly impacts profitability: Lenzo's architecture can perform more computations with less power.
Crypto assets are just the beginning
Despite CGLA's impressive performance, Lenzo isn't just aiming to dominate the mining market. Their ultimate goal is to disrupt the AI computing market.
Why "Crypto First"?
The biggest barrier to market entry for new semiconductor architectures is the lack of a software ecosystem. In particular, in the AI field, the development environment "CUDA" that NVIDIA has built over many years has become the de facto standard, and new entrants have been unable to break this stronghold.
So Lenzo has come up with a very clever two-step strategy.
Entering a CUDA-independent market: Cryptocurrency mining does not rely on specific software frameworks like CUDA, and revenue is directly linked to the computational performance of the hardware. CGLA's overwhelming performance and efficiency make it ideal for establishing an early market.
Exploiting the supply-demand imbalance: With the rise in cryptocurrency prices, the demand for mining machines is constantly outstripping supply. This market gap presents a huge opportunity for new players like Lenzo.
Securing self-funding (self-mining): By conducting mining using high-performance mining machines developed in-house, Lenzo can secure a stable source of revenue before we start selling products. This will be reinvested in the development of AI chips and the construction of a software ecosystem, which is the biggest challenge.
The strategy is to avoid a direct confrontation with NVIDIA, first establish a foothold and secure funds in a market where it is easier to compete, and then, once ready, take on the AI market, which is Lenzo's main target.
Groundwork for AI servers: LLM is already underway
Lenzo's efforts towards AI is not just a concept. The company has already built a prototype AI server equipped with 64 Lenzo cores in an environment that combines Intel servers and FPGAs. They have successfully run Meta's large-scale language model (LLM), Llama, on this server.
In a publicly released demo video, an FPGA equipped with a Lenzo Core running at just 140 MHz demonstrated inference performance approaching one-third of an Intel Xeon processor (3.1 GHz), despite operating at only 1/22nd of the clock speed. It’s a clear sign that CGLA has the potential to become a game-changer in AI inference processing as well.
(Source: Lenzo)
Can Lenzo rekindle Japan’s semiconductor ambitions?
Top engineers who once helped put Japan at the forefront of global technology — from PlayStation to supercomputers — are coming together again to tackle today’s computing challenges from the architecture level.
Lenzo’s story feels to the author like more than just a startup journey; it carries a broader significance.
According to CEO Mr. Fujiwara, he has long felt concerned that, since the PlayStation 4 era, the industry has moved toward off-the-shelf architectures, losing momentum around original chip architecture development.
“In the past 30 years, we’ve seen very little true innovation in chip architecture. With our unique CGLA architecture, which can support both AI and crypto workloads, we’ve achieved something only we could.”
His words reflect a strong desire to help Japan return to the forefront of technological innovation after decades of stagnation. Of course, the road ahead won’t be easy. One of the biggest hurdles will likely be whether Lenzo can build a software development environment comparable to CUDA. No matter how capable the hardware, it won’t matter unless developers can make use of it. A key question will be how effectively the company can reinvest its mining revenue into software development. This is where Lenzo’s real capabilities will be tested.
Still, the value of the challenge itself is significant. As noted by renowned VC firm Andreessen Horowitz (a16z), the convergence of AI and crypto represents a major force driving the next wave of computing. Lenzo appears to have identified this intersection earlier — and perhaps more clearly — than most.
Whether they can meaningfully challenge an industry giant like NVIDIA and spark renewed momentum in Japan’s semiconductor sector remains to be seen. But their efforts are certainly worth watching.
