Esperanto Technologies is folding its chip business, according to EE Times, leaving it to seek a technology buyer or licensor. The company had developed a data-center AI accelerator (NPU) that incorporated technologies such as low-voltage SRAM cells and RISC-V cores.
Founded in 2014, the company sampled its ET-SoC-1 chip in 2021 and began production in 2023. When the company started, the important AI models were convolutional neural networks (CNNs) employed for object detection and similar computer-vision applications. A company like Facebook (Meta) could, for example, use a CNN to identify people in users’ photos. Subsequently, recommendation models were used to choose content and ads to show users. Today, transformers, the neural-network type underpinning large language models (LLMs) dominate NPU workloads.
Lashing together 1,088 RISC-V cores with tensor and vector extensions, Esperanto created a flexible architecture capable of running these disparate models. Moreover, the ET-SoC-1’s low-voltage design kept typical power at 25 W. However, a problem is that a CNN-focused design assumes a greater locality of reference than one optimized for transformers. Whereas a CNN operates on pixel clusters, a transformer compares tokens over a large window.
The ET-SoC-1’s meager DRAM bandwidth reflects this. Its 132 GB/s is 1/25th that of the Nvidia H100. Customers deploying 13-billion-parameter models will find the Esperanto chip power efficient. That’s about the size of Meta’s Llama 4 Scout LLM, but models with 70 billion or more parameters are common. DeepSeek-R1, for example, has 671 billion. Moreover, recent LLMs have context windows spanning 100,000 or more tokens, further stressing memory resources. The ET-SoC-1 is stale in other ways. For example, it supports FP32, FP16, and INT8 types, but FP8 and FP4 are becoming important.
A planned second-gen chip was to add FP64 capabilities to address HPC customers and employ HBM. Meanwhile, however, Nvidia and AMD are trimming FP64 in upcoming designs to make way for additional low-precision math units, recognizing it’s not practical for one chip to serve two markets. Having landed few customers after spending years developing the ET-SoC-1, Esperanto no longer has the resources to keep going, despite its low-power differentiation.