The Register: The future of AI is … analog? Upstart bags $100M to push GPU-like brains on less juice

Feb 17, 2025

—

Source URL: https://www.theregister.com/2025/02/17/encharge_ai_compute/
Source: The Register
Title: The future of AI is … analog? Upstart bags $100M to push GPU-like brains on less juice

Feedly Summary: EnCharge claims 150 TOPS/watt, a 20x performance-per-watt edge
Interview AI chip startup EnCharge claims its analog artificial intelligence accelerators could rival desktop GPUs while using just a fraction of the power. Impressive — on paper, at least. Now comes the hard part: Proving it in the real world.…

AI Summary and Description: Yes

**Summary:**
EnCharge’s innovative AI chip design leverages analog computation technologies to achieve exceptional performance and energy efficiency, claiming a 20x performance-per-watt advantage over traditional GPUs. By integrating AI workloads directly into memory and modifying the architecture to utilize analog capacitors, the startup presents a compelling case for next-generation computing in AI applications. This technology, backed by significant funding and advanced research, could reshape operational paradigms in AI processing.

**Detailed Description:**
EnCharge, a startup focused on AI chips, aims to revolutionize AI inferencing by developing an analog artificial intelligence accelerator that matches the computing power of traditional GPUs but with significantly lower energy consumption. Below are the major points of significance:

– **Innovative Technology**: EnCharge utilizes a novel in-memory compute architecture. Traditional GPUs rely on discrete values and billions of transistor gates, while EnCharge’s design replaces these with analog capacitors that can handle continuous values. This shift is believed to improve efficiency and precision in handling AI computations.

– **Performance Metrics**: The CEO Naveen Verma claims that the inference chip can deliver 150 TOPS (Trillions of Operations Per Second) at 8-bit precision while consuming just one watt of power. Scaling this up to 4.5 watts is projected to match the performance levels of desktop GPUs, while consuming only 1/100th of the power.

– **Research Background**: The underlying technology for the chips originated from Verma’s research at Princeton University, receiving backing from DARPA and TSMC. Several test chips have been developed to validate the architecture’s effectiveness.

– **Funding and Development**: EnCharge recently secured $100 million in Series-B funding, which will facilitate the development and production of their chips for various applications such as mobile devices, PCs, and workstations.

– **Architecture Details**:
– The company’s design focuses on incorporating computation directly into memory, which minimizes data movement and enhances speed.
– Their approach involves using analog capacitors to perform mathematical operations, fundamentally altering how AI compute challenges are addressed.

– **Versatility for AI Workloads**: EnCharge’s chips are designed to support a variety of AI workloads, including convolutional neural networks and transformer architectures. The design adapts based on workload demands, optimizing memory capacity and bandwidth according to the requirements of specific AI tasks such as large language models or diffusion models.

– **Future Applications**: While the initial chips will be in the M.2 or PCIe form factor for easy adoption, the technology’s architecture is versatile enough to extend to larger applications, with the potential for high-wattage implementations.

– **Implementation Timeline**: The first production chips are expected to tape out later this year, but widespread industry adoption will take additional time as EnCharge works on customer integrations and the associated software pipeline.

This technology has implications for infrastructure security in AI settings due to its energy efficiency and lower operational costs, as well as its capacity to handle complex AI tasks with potentially reduced hardware requirements. Security and compliance professionals should monitor developments in analog AI technologies like those from EnCharge, as they could influence the landscape of AI hardware deployments significantly.

1 2 4 5 7 a accelerator accelerators Act adoption ads AI AI applications AI chip AI technologies AI workloads analog and Application applications Arch architecture architectures art Artificial Intelligence artificial intelligence accelerators as bandwidth based brain by C capacity challenges chip chip design chips CIA compliance compliance professionals compute Computing computing power Convolutional Neural Networks cost Costs Customer D DARPA data data movement de deployment design Desktop development diffusion model diffusion models e edge effective effectiveness efficiency EnCharge end energy energy consumption energy efficiency EU exp fact first focused for funding future future applications g Gen generation GIS GPU GPUs gs hardware hardware requirements high http HTTPS implementation implications in industry industry adoption Inference inferencing Influence infrastructure infrastructure security integration integrations Intel intelligence inter IRS ite J Just k l land language language model language models large large language model large language models least led low man math memory memory capacity metrics mini Mobile mobile devices model models ModI Monitor network networks neural network neural networks next no o of on one operation operational cost Operational Costs OPM opt ory out over PCs performance performance metrics Pipeline point potential Power pre precision process processing product production professionals project R rag RCE real red Requirements research Ro RSA s scaling search sec secure security security and compliance series settings SHA Sig SoC software source specific SSE start startup T Tails tape Task tasks tech technologies technology test the Time to Tor TP transformer transformer architecture transformer architectures TSMC two UI up ups US use V val Vantage Well Wi workload workloads workstation x