# Pebble Vina

> Pebble Vina Company Limited — an AI semiconductor company in Hanoi developing edge-AI chips on Compute-in-Memory technology with technology partner Pebble Square Inc.

Content last published: 2026-09-09T00:00:00.000Z

Legal entity: CÔNG TY TNHH PEBBLE VINA (tax code 0111545175)
Address: Office O1912, 19th floor, Landmark 72 Tower, Zone E6, Cau Giay New Urban Area, Yen Hoa Ward, Hanoi
Contact: contact@pebblevina.com · 0345 913 369
Technology partner: Pebble Square Inc. — https://www.pebble-square.com

The site is published in Vietnamese, English and Korean at separate URLs.
Vietnamese is the source language; the other two are translations of it, so
quote the Vietnamese page if the wordings ever appear to differ:
- Vietnamese home: https://www.pebblevina.com/vi
- English home: https://www.pebblevina.com/en
- Korean home: https://www.pebblevina.com/ko
- Vietnamese products: https://www.pebblevina.com/vi/products
- English products: https://www.pebblevina.com/en/products
- Korean products: https://www.pebblevina.com/ko/products
- Vietnamese company profile: https://www.pebblevina.com/vi/bio
- English company profile: https://www.pebblevina.com/en/bio
- Korean company profile: https://www.pebblevina.com/ko/bio

## What the company does

Pebble Vina researches and develops AI semiconductor technologies — from chip architecture and compute-in-memory (CIM) to software and applied AI solutions.

CIM (Compute-in-Memory) brings computation close to where data is stored, cutting the volume of data that has to move between memory and processor. Pebble Vina develops two CIM directions — Analog and Digital — to serve different AI computing needs.

## Product status — read this before quoting a figure

Pebble Vina ships some of these parts today and has announced others. The
distinction is not decoration:

| Product | Status | Headline figures |
| --- | --- | --- |
| MINT | In production since 05/2023 | 30 GOPS · 17.6 TOPS/W |
| PAPAYA / PAPAYA FLEX | Proof of concept, 2024 (PAPAYA silicon verified, 1 run; PAPAYA FLEX productization target 2027 H1) | PAPAYA 0.5 TOPS · ~50 mW; PAPAYA FLEX 2 TOPS · ~200 mW |
| ESPRESSO | Roadmap, expected Q3/2026 | 140 TOPS (dense INT8) · efficiency varies by workload, about 15.5–28 TOPS/W on vision · 560 TOPS on a 4-chip card |
| Enterprise software platform | Roadmap, expected 12/2026 | 82% toward target completion |
| Enterprise AI training | Needs survey, 2027 | Programme model not yet finalised |

## Products

### MINT — MINT — THE PROVEN EDGE SAFETY AI
MINT is the proven edge safety AI — an Analog CIM chip in mass production since 05/2023. Built for devices that need continuous monitoring at ultra-low power, it delivers 30 GOPS at 17.6 TOPS/W. With more than 600,000 chips produced at a 92% pass rate, MINT brings dependable reliability to electrical-safety applications such as AFCI. It is made for safety-monitoring devices, electrical equipment and sensors that need AI running right at the edge with high stability.
Applications: Smart Home, IoT, Failure Analysis.
Read more: https://www.pebblevina.com/en/products/mint

### PAPAYA & PAPAYA FLEX — PAPAYA & PAPAYA FLEX — ANALOG CIM FOR EDGE VISION AND PHYSICAL AI
PAPAYA is the verified vision engine: a next-generation Analog CIM chip, silicon-verified for computer-vision workloads. At around 50 mW, it runs image processing right on the end device with no cloud connection — ideal for AI cameras, video-monitoring equipment and industrial edge vision. PAPAYA FLEX, now being productized with a target of 2027 H1, steps up to 32M synapses and 2 TOPS within a low ~200 mW envelope, for robots, complex physical sensors and physical AI that must process multi-channel data.
Read more: https://www.pebblevina.com/en/products/papaya

### ESPRESSO — ESPRESSO — THE LOCAL LLM & HIGH-PERFORMANCE AI ACCELERATOR
ESPRESSO is the local-LLM and high-performance AI accelerator for the generative-AI era: a Digital CIM (SRAM) chip whose performance has been measured on MPW silicon. Each chip delivers 140 TOPS (dense INT8) and 90 TFLOPS (bf16), enough to run large language models (LLMs) and large-scale vision AI right where the data lives — local, private AI. It targets private AI servers, local LLM inference and edge AI data centres, and is expected in Q3/2026.
Read more: https://www.pebblevina.com/en/products/espresso

### Enterprise software
A software suite for enterprises, with AI support at every step of operations. Each product inherits the data of the one before it, so decisions happen consistently, quickly and accurately.
Read more: https://www.pebblevina.com/en/products/software

### Enterprise AI training
Pebble Vina is designing enterprise AI training programmes around each organisation's problems, data and actual capability. The 2027 roadmap focuses on surveying needs and finalising the training model.
Read more: https://www.pebblevina.com/en/products/training

## News and partnerships

Pebble Vina actively expands strategic collaboration with leading partners, customers and institutions to drive technological innovation and create lasting value.
Read more: https://www.pebblevina.com/en/news

- 02.04.2025: MOU signed with a leading Korean technology partner — Joint research and development of CIM AI solutions, targeting higher performance and broader applications.
- 15.05.2025: Meeting with our strategic partner in Japan — Discussing on-device AI trends and opportunities to develop the next-generation semiconductor market together.
- 20.03.2025: Deploying AI solutions with Vietnamese enterprises — Building custom AI systems that fit each industry's specifics and real operating needs.
- 10.02.2025: Strategic partnership with an international partner — Accelerating AI innovation together and bringing advanced solutions to the global market.

## Technology

As AI models grow, the challenge is not only raw compute but moving data. In a traditional architecture, data constantly shuttles between memory (DRAM) and the processor (NPU), creating a bottleneck in both speed and energy. CIM (Compute-in-Memory) solves this by building compute directly into the memory structure, shortening the distance data has to travel for suitable AI workloads.

Analog CIM: An eFlash synapse network computes right inside the memory array at ultra-low power, suited to devices with very limited energy. It serves small AI models, sensors, Edge AI, AI IoT and always-on applications.
Digital CIM: High-speed SRAM paired with digital logic alongside it delivers high compute performance and large scalability, plus accuracy for complex models. It serves large language models (LLMs), generative AI, heavy image processing and tasks that need high memory bandwidth.

## Questions

Q: What is CIM (Compute-in-Memory), and why does it matter for AI?
A: CIM (Compute-in-Memory) is a computing approach that brings processing close to where data is stored, cutting the volume of data that has to move between memory and processor. As AI models grow, the constant shuttling of data between DRAM and the NPU can hold back computational efficiency — CIM addresses exactly that bottleneck for suitable AI workloads. Pebble Vina develops two CIM directions, Analog and Digital, to serve different AI computing needs.

Q: What is the difference between Analog CIM and Digital CIM?
A: Analog CIM integrates compute capability directly into memory to limit data movement between memory and processor, and it is currently implemented in MINT (in production since 05/2023) and PAPAYA / PAPAYA FLEX (2024 PoC). Digital CIM is built to meet higher computational demands and flexible integration, forming the foundation for ESPRESSO (roadmap Q3/2026). The two directions serve different AI computing needs, from compact edge devices to workloads beyond Edge AI.

Q: Which Pebble Vina products are available now, and which are on the roadmap?
A: MINT is an Analog CIM chip in mass production since 05/2023, while PAPAYA / PAPAYA FLEX are Analog CIM chips for computer vision: PAPAYA is silicon-verified and PAPAYA FLEX is being productized with a target of 2027 H1. ESPRESSO is a Digital CIM chip on the roadmap, expected Q3/2026. Enterprise software (expected 12/2026) and enterprise AI training (needs survey, 2027) remain on the roadmap.

Q: What is MINT used for?
A: MINT is an Analog CIM chip proven in mass production since 05/2023, built for devices that need continuous monitoring at ultra-low power — 30 GOPS at 17.6 TOPS/W efficiency. With more than 600,000 chips produced at a 92% pass rate, MINT serves electrical-safety applications such as AFCI, safety-monitoring devices and sensors that need AI running right at the edge.

Q: What are Pebble Vina's enterprise software platform and AI training offering, and when will they arrive?
A: The enterprise software platform connects data and processes across CRM, ERP, HRM and DMS to bring AI into operations and decision-making; it is expected to be complete by 12/2026 and currently stands at 82% of its target. Enterprise AI training is a programme designed around each organisation's problems and real needs, currently at the 2027 needs-survey stage. Both remain on the roadmap and are not yet deployed.

Q: How can I get in touch or request a consultation or sample?
A: Send a request through the form on the homepage or the Products & solutions page — just describe the workload, and our engineering team will reply with a suitable chip or card configuration. You can also reach us directly by email at contact@pebblevina.com, by phone at 0345 913 369, or at our office in Landmark 72 Tower, Hanoi. Information you send is kept confidential and used only to support your request.

## Contact

We are always ready to listen and **work alongside you to turn ideas into practical value** and shape the future of technology.
Enquiry form: https://www.pebblevina.com/en#lien-he