HyperAccel

What HyperAccel does

HyperAccel is a fabless company developing AI chips, specifically LLM Processing Units (LPUs), to enable widespread deployment of Generative AI (GenAI). They focus on inference over training and LLMs over traditional AI models to overcome cost barriers by designing efficient hardware solutions. Their products include the Bertha LPU, Bertha 500, Forte 55X, and HX-F55X, along with a user-friendly software platform that supports leading AI frameworks and transformer-based LLMs.

Products

• BerthaA LPU (LLM Processing Unit) purpose-built for AI inference, designed to scale without compromise, redefining AI performance at the silicon core.
• Bertha 500An AI chip designed for efficient and sustainable AI inference, built to do more with less while minimizing energy consumption and resource waste.
• Forte 55XOptimized for low-latency, power-efficient AI inference, combining LPU technology with HBM-equipped AMD Alveo U55C high-performance compute card.
• HX-F55X (formerly Orion)HyperAccel Xceleration system comprised of Forte 55X and an ASUS server for high-speed datacenter inference, enabling on-premise AI with affordability.
• HyperAccel SoftwareA user-friendly software platform that provides a standardized ecosystem for Generative AI inference, supports LLM inference and model frameworks (vLLM, HuggingFace), accommodates all transformer-based LLMs and multi-modal models, offers Pytorch support and Python-embedded domain-specific language (eDSL), facilitates a developer page and model zoo, and implements device runtime and driver for LPU execution.

Key capabilities

• AI-Specialized Core (Coarse-grained, programmable processor for LLM inference)
• Streamlined Dataflow (Aligned bandwidth, model parameter reuse)
• Multi-Chip Scalability (Overlap of communication and computation)
• LPU technology (LLM Processing Unit)
• Optimized for low-latency, power-efficient AI inference
• Supports LLM inference and model frameworks (vLLM, HuggingFace)
• Accommodates all transformer-based LLMs (e.g., GPT, Llama, Qwen, Mistral, Grok, DeepSeek, Falcon, Gemma) and multi-modal models
• Pytorch support and Python-embedded domain-specific language (eDSL)
• Developer page and model zoo for easy compilation
• Device runtime and driver for LPU execution
• Seamless LLM inference experience for developers familiar with GPUs
• Streamlined Memory Access (minimal data buffering and reshaping)
• Internal network controller, Expandable Synchronization Link for multi-chip scalability
• Computation-communication overlapping to minimize communication overhead

Use cases

• AI inference for Generative AI
• LLM inference
• On-premise AI
• Edge appliances
• Robots
• Datacenter inference

AI approach

HyperAccel designs and develops AI semiconductors (LPU - LLM Processing Unit) specifically for Generative AI inference. Their hardware, like Bertha and Forte 55X, is optimized for performance, efficiency, and scalability in LLM inference. They also provide a software platform that supports leading AI frameworks and various transformer-based LLMs.

Tech named: LPU (LLM Processing Unit), AI-Specialized Core, Streamlined Dataflow, Multi-Chip Scalability, Bertha 500, Forte 55X, HX-F55X (HyperAccel Xceleration system), vLLM, HuggingFace, GPT, Llama, Qwen, Mistral, Grok, DeepSeek, Falcon, Gemma, Pytorch, Python-embedded domain-specific language (eDSL), Expandable Synchronization Link, DRAM, Samsung 4nm, FPGA (Field-Programmable Gate Array)

Industries served

• Semiconductor Manufacturing
• Cloud datacenters
• AI services
• Consumer electronics (e.g., LG on-device AI)
• Robotics

What it says sets it apart

• Focus on inference over training for GenAI
• Focus on LLMs instead of traditional AI models
• Overcomes cost barriers with efficient hardware solutions
• Unparalleled performance, efficiency, and scalability
• LPU technology for redefining AI performance
• Bertha 500 designed for efficiency and sustainability, minimizing energy consumption and resource waste
• Forte 55X optimized for low-latency, power-efficient AI inference
• HX-F55X enables on-premise AI with unmatched affordability
• Fully compatible with leading AI frameworks
• Supports a wide range of transformer-based LLMs and multi-modal models
• Offers a standardized ecosystem for Generative AI inference
• Aims for the most sustainable and efficient AI inference
• Optimized for power-efficient and cost-effective inference without compromising on performance
• LPU technology offers 2.4 times better performance than conventional GPUs and 50% faster processing speed (CEO claim)
• Utilizes DRAM instead of expensive HBM for cost-effectiveness (CEO claim)
• Samsung 4nm backing for 'Bertha' prototype production

Funding rounds we track

From the AI funding tracker — rounds as reported by the linked publications.