Validates Performance with AMD and Launches Mango Inference to Accelerate Global AI Platform Expansion

- Up to 1.25× higher token throughput than industry-leading GPU platforms
- Up to 5.1× higher throughput and over 2× lower latency than conventional GPU clusters

SAN FRANCISCO: #AIInference--MangoBoost announced the commercial launch of Mango Inference, its serverless AI inference platform, while unveiling breakthrough inference performance on AMD's latest Instinct™ MI355X accelerators at AMD Advancing AI 2026 in San Francisco.

At the event, the company demonstrated inference performance surpassing industry-leading GPU platforms, highlighting a shift in AI competition from model size to efficient inference delivery and reinforcing its position as a global AI inference platform provider.

Using AMD Instinct™ MI355X GPUs combined with MangoBoost's inference optimization software LLMBoost™, the system achieved up to 1.25× higher token throughput than industry-leading GPU architectures under the same latency constraints in GLM-5/5.1 (FP8) benchmarks. End-to-end latency was reduced by up to 5.1×, cutting AI response time from 142 seconds to 28 seconds. The system also demonstrated outstanding performance on DeepSeek-R1, a reasoning-focused AI model. These results were further validated through MLPerf, the industry's leading AI performance benchmark. With these performance results, MangoBoost's product capabilities have led to a strategic collaboration with AMD.

As AI moves from training to inference, real-world competitiveness depends on how effectively compute, networking, storage, memory, and system software are integrated and optimized. MangoBoost's achievements demonstrate that AI leadership is defined by system-level optimization rather than hardware specifications alone. Through its collaboration with AMD, MangoBoost validated and improved AI inference performance via MLPerf submissions and unofficial InferenceX projects on AMD Instinct™ GPUs.

Additionally, the event highlighted Korea's growing leadership in AI infrastructure. Deputy Prime Minister and Ministry of Science and ICT Bae Kyung-hoon and AMD Chair and CEO Lisa Su visited the MangoBoost booth to review its AI inference technologies and future vision. Their visit underscored MangoBoost's growing global recognition as a key contributor to South Korea's flagship AI initiatives.

Following successful Seed and Series A funding, MangoBoost is preparing for its Series B financing, supported by validated technology, commercial launch of Mango Inference, and growing global customer traction. The investment will accelerate expansion across North America, the Middle East, Asia, and Europe while advancing its AI inference platform and enterprise offerings.

MangoBoost aims to become a global full-stack AI infrastructure leader, optimizing computing, networking, memory, and storage across next-generation AI data centers. Its long-term vision is to establish the global standard for AI infrastructure optimization in the Agentic AI era.

Jangwoo Kim, CEO of MangoBoost, said, "Our industry is no longer competing over which GPU is used. The real competition is about how efficiently hardware and software are integrated into a complete AI infrastructure. Together with AMD, we will continue setting new benchmarks for AI inference and expand Mango Inference as a global AI platform."

Overall, this marks a turning point demonstrating that the future of AI inference will be defined by system optimization. With world-class performance and a commercially available AI inference platform, MangoBoost is reshaping the competitive landscape of AI infrastructure worldwide.

About MangoBoost
MangoBoost builds full-stack AI infrastructure software and hardware that raises the efficiency of AI workloads. Founded out of Seoul National University and headquartered in Bellevue, WA, the company develops LLMBoost™ inference optimization software, DPU acceleration and Agent OS, and operates Mango Inference, a managed serverless inference platform. More at https://www.mangoboost.io/.

Fonte: Business Wire