EDA
System
IP
SolvNetPlus
Training & Education
To go up against the biggest players in the AI hardware industry, startup chip designers need to take a radically different approach — and innovate rapidly.
That’s exactly what AI chip unicorn Rebellions has accomplished with its new REBEL-Quad semiconductor. Unveiled at Hot Chips 2025, REBEL-Quad is the world’s first UCIe-Advanced AI accelerator, purpose-built for frontier LLMs with hundreds of billions of parameters. The 4-homogeneous-chiplet system-on-chip (SoC) design delivers up to 2,048 teraflops in 8-bit Floating Point (FP8), putting it in the same performance class as leading GPUs.
The key difference: superior energy efficiency.
REBEL-Quad delivers 1.6× higher throughput and 50% lower power consumption, resulting in 3.2× more transactions per second per watt (TPS/W) than top-tier GPUs (tested on Llama 3.3 70B in FP8).
For data centers grappling with skyrocketing energy usage due to AI workloads, Rebellions is betting that REBEL-Quad offers significantly better total cost of ownership (TCO) with similarly high performance.
“This is a powerful solution, and we can deliver much higher energy efficiency compared to GPUs,” says Jinwook Oh, Rebellions co-founder and chief technology officer. “We believe REBEL-Quad delivers the highest performance per TCO ever and will make a huge impact in the AI inference market.”
Enhance Your Chip Design with AI
Explore the enhanced Synopsys.ai brochure, featuring cutting-edge advancements in Advanced Optimization, Generative AI, and Agentic AI to transform your chip design process.
Targeting new markets
REBEL-Quad marks Rebellions’ first big swing in the U.S. and Europe. Founded in just 2020, the chip startup is backed by Arm as well as South Korean leaders SK Hynix, SK Telecom, Korea Telecom, and Samsung Ventures & Securities. Its previous ATOM-Max semiconductor, now in mass production, primarily targets Asian markets.
But with REBEL-Quad, Rebellions is eyeing bigger opportunities.
“U.S. customers are very sensitive to TCO,” says Oh. “Even though it’s a pretty crowded market, if we can provide a strong TCO benefit, I think there will be strong demand.”
Outside the U.S., Oh expects REBEL-Quad to provide an alternative for many countries that want to host their own AI data centers but struggle to source GPUs.
“They are craving GPUs,” he says. “Our solution will be a good substitute. I believe this solution will replace many GPU markets.”
Rebellions REBEL-Quad AI Accelerator
Collaborating to compete
A key part of Rebellions’ go-to-market strategy is positioning itself as a systems company, collaborating with the likes of Marvell Technology and Credo Technology for SoC design and chiplet components, as well as AI server rack and infrastructure partners like Pegatron and Penguin Solutions. The company’s goal is to design and manufacture proprietary accelerator boards and full server rack systems specifically tailored to REBEL-Quad.
“We want to provide a performant and energy-efficient solution in a rack scale,” says Oh. “By understanding the requirements and limitations of AI-dedicated data centers, we can provide the best, optimized solution for those environments.”
Expanding its circle of collaborators, Rebellions recently joined the Arm Total Design ecosystem, a group of industry leaders developing custom systems based on Arm Neoverse Compute Subsystems (CSS).
“Arm is becoming a keystone of this ecosystem,” says Oh. “Choosing the right partners and building the right solution on top of an ecosystem are the most important factors for increasing our performance so that we can compete against the big companies.”
Jinwook Oh, co-founder and chief technology officer of Rebellions
Accelerating chip validation and software development
Time to market in the race for AI supremacy is critical, and Rebellions partnered with Synopsys to accelerate the validation of its breakthrough semiconductor using ZeBu and Virtualizer tools.
At the core of REBEL-Quad is a modular SoC with four identical compute chiplets integrated using high-bandwidth UCIe-Advanced links. Alongside the chiplets on an I-Cube S package, four HBM3E memory stacks provide 144 GB total capacity and approximately 4.8 TB/s aggregate bandwidth.
“Validation is challenging, but ZeBu is a perfect platform for that,” says Oh. “Its speed, capacity, and support from the Synopsys team made the emulation of this project successful.”
Working closely with Synopsys, Rebellions implemented REBEL-Quad using a hybrid combination of ZeBu and Virtualizer, leveraging the latter for virtual prototyping and accelerated software development.
“Our software programmers picked a design based on their needs,” says Oh. “If they needed to validate all the functionality thoroughly, they went with the whole mapped solution. But if they wanted to develop their software stack quickly, they would use Virtualizer.”
Rebellions designers were able to run an LLM, enter prompts on a host, and observe token acceleration within the full system context. Integrating the user’s end application into the design process allowed developers to “shift left” TPS/W testing to earlier stages of development. While end users of AI accelerators focus on maximizing tokens processed per second, fast validation using platforms like ZeBu is essential during development to minimize the time required to execute each token. With emulation, this execution can be reduced to just minutes, depending on the complexity of the setup.
“We prepared most of the software stack during the pre-silicon phase, which helped us go from first silicon to live demo in just five weeks. That’s a major achievement,” says Oh. “Even more impressive was how closely ZeBu’s performance predictions matched the actual silicon — 98% accuracy.”
The race is on
In the race to deliver powerful, energy-efficient AI semiconductors to the world, startups need to innovate faster than ever. With the development of REBEL-Quad, Rebellions has engineered an AI accelerator card for efficient, scalable inference — and demonstrated that it’s one to watch in a highly competitive market.
Continue Reading
Industry 1st CXL 4.0 Verification IP: Transforming AI and HPC Systems
Synopsys Bold Prediction: New Cars Will Contain 600 Million Lines of Code by 2027
