Etched
Transformer-specific ASIC chip company
Etched is a semiconductor startup designing application-specific integrated circuits, or ASICs, built to run only transformer-based neural networks, the architecture underlying most modern large language models. By hard-wiring the chip's…
Definition
Etched is a semiconductor startup designing application-specific integrated circuits, or ASICs, built to run only transformer-based neural networks, the architecture underlying most modern large language models. By hard-wiring the chip's circuitry around the transformer's core operations, such as attention and matrix multiplication, rather than building a general-purpose programmable processor, Etched aims to deliver substantially higher inference speed and efficiency for transformer workloads than general-purpose GPUs can achieve, at the cost of being unable to run other kinds of computation at all.
Overview
Etched's central bet is that the transformer architecture, first introduced in the 2017 paper "Attention Is All You Need" and now underlying the vast majority of large language and multimodal models, has become stable and dominant enough that it makes sense to build silicon that does nothing else. Where GPUs are designed as flexible, massively parallel processors capable of running almost any kind of numerical workload, an ASIC trades away that generality entirely, etching the transformer's specific computational pattern directly into the chip's circuitry so there is no wasted die area or instruction overhead for operations the chip will never perform. Mechanically, this means Etched's chips are built around fixed, highly optimized data paths for the matrix multiplications and attention computations that make up the bulk of a transformer's forward pass, rather than a general instruction set that a compiler translates arbitrary code into. Because the hardware doesn't need to support conditional branching, varied instruction types, or the broad range of operations a GPU must handle for graphics and general compute, more of the chip's transistor budget can go directly toward transformer-relevant arithmetic and on-chip memory bandwidth, the two resources that most directly bound inference speed for large models. This specialization can yield large throughput and efficiency gains specifically on transformer inference, though it offers no benefit, and indeed cannot function at all, for training workloads or non-transformer architectures. Etched sits at the most narrowly specialized end of the AI chip spectrum, distinct from broadly programmable AI accelerators and from companies like Untether AI or d-Matrix that build memory-centric architectures still capable of running a range of network types. Its closest conceptual relatives are other workload-specific ASIC efforts, though few competitors have committed as explicitly to a single architecture family as Etched has to transformers. This is a higher-risk, higher-reward position: if a successor architecture displaces transformers as the dominant paradigm for large models, a transformer-only ASIC would need a fundamentally new chip design to remain relevant, whereas GPU-based infrastructure can simply run new model code. In practice, Etched's chips are aimed at high-volume inference serving for large language models, where the same transformer computation is repeated for every user request at massive scale, making the fixed cost of specialized silicon worthwhile because it is amortized across enormous inference volume. Organizations running consumer-facing chat products, coding assistants, or other latency-sensitive LLM applications are the natural customers, since throughput and cost per token served scale directly with how efficiently the underlying hardware executes each transformer layer. The trade-off is inflexibility: an ASIC that only runs transformers cannot be repurposed if a workload shifts to a different architecture, cannot be used for model training in most designs, and requires a software and deployment ecosystem that can reliably compile transformer models onto its fixed pipeline. Organizations that need general-purpose compute, that train models frequently, or that experiment with novel architectures outside the transformer family will still rely on GPUs; Etched's proposition only pays off for buyers confident that transformer-based inference at scale will remain their dominant, long-running workload.
Key Features
- Application-specific integrated circuit built only for transformer models
- Fixed, hard-wired data paths for attention and matrix multiplication
- No support for conditional branching or general-purpose instructions
- More transistor budget devoted to transformer arithmetic and memory bandwidth
- Cannot run training workloads or non-transformer architectures
- Designed to amortize fixed silicon cost across massive inference volume