Untether AI

What is Untether AI?

Untether AI designs hardware for running trained neural networks quickly and with low power. Its at-memory architecture places compute elements next to memory cells. The first-generation runAI200 chip has 511 custom RISC processors and 204 MB of SRAM, rated up to 502 TOPS. The second-generation speedAI240 uses 1,400 RISC-V processors and is rated up to 2 PetaFlops. These chips ship on PCIe accelerator cards: the speedAI240 Slim and tsunAImi tsn200, both with 75-watt TDP, and the 300-watt tsunAImi tsn800 with four runAI200 devices. The imAIgine SDK imports models from TensorFlow, PyTorch and ONNX, then compiles, quantizes, profiles and runs them through a C-based runtime API. Target markets include vision AI, finance, government, automotive and AgTech. Prices are not published; the site points to a purchase request and product briefs.

Use cases

• Run vision AI inference at the edge
• Accelerate transformer and LLM inference
• Deploy models in on-premise financial systems
• Process video analytics in real time
• Compile TensorFlow and PyTorch models for hardware
• Profile and simulate model performance

Who it's for

• Hardware and ML engineers
• Edge AI developers
• Data center teams
• Financial services technologists

Why use Untether AI

• Targets low power use in edge and data center inference
• A push-button flow takes models from standard frameworks to the hardware
• Low-profile 75-watt cards fit edge deployments
• Chip-to-chip interconnect lets finance teams scale from one chip to a rack on premises
• Reported MLPerf Inference v4.1 results give a reference for ResNet-50 performance

Untether AI features

• runAI200 and speedAI240 inference chips
• speedAI240 Slim and tsunAImi PCIe accelerator cards
• At-memory compute architecture
• imAIgine SDK with compiler, toolkit and runtime
• Model import from TensorFlow, PyTorch and ONNX
• Quantizer to optimize models for inference
• Model Garden and custom kernel development flow
• Card health and temperature monitoring
• C-based API for application integration

Untether AI integrations

• TensorFlow
• PyTorch
• ONNX

How to get started with Untether AI

1. Train a model in TensorFlow or PyTorch, or export it to ONNX.
2. Import it into the imAIgine SDK and set performance and power targets for the compiler.
3. Quantize and profile the model with the toolkit and simulation tools.
4. Deploy it on a tsunAImi or speedAI card and integrate through the C API.
5. Use the runtime to monitor card health and temperature.

Untether AI limitations

• Prices are not published
• Hardware must be purchased or requested; there is no self-serve plan
• The imAIgine SDK has an Early Access Program and demo request

Untether AI pricing

Untether AI is sold on custom quotes, so its prices aren't published.

Prices can change and may vary by country or exclude tax. Check the official pricing page.

Untether AI alternatives

The main Untether AI alternatives listed on Debuly are Groq, Cerebras, SambaNova, NVIDIA Dynamo-Triton, ExecuTorch and Alibaba Cloud Model Studio. The table compares pricing and platforms; order is A–Z.

1. 

   Alibaba Cloud Model Studio

   Alibaba Cloud platform giving developers API access to Qwen and partner models for text, image, video and speech, billed per token or by subscription.

   Developer platformsAudio and voiceVideo
   PaidVisit Alibaba Cloud Model Studio (official site)
2. 

   Cerebras

   Inference cloud and API running open models on wafer-scale chips, with a free tier, pay-per-token Developer plan and coding plans.

   Developer platformsCoding
   FreemiumVisit Cerebras (official site)
3. 

   ExecuTorch

   PyTorch's open-source framework for exporting models and running them on phones, microcontrollers and other edge devices.

   Developer platformsCoding
   Open sourceVisit ExecuTorch (official site)
4. 

   Groq

   GroqCloud inference API that serves open models such as GPT-OSS and Whisper on Groq's LPU hardware, billed per token.

   Developer platformsAudio and voiceCoding
   FreemiumVisit Groq (official site)
5. 

   NVIDIA Dynamo-Triton

   Open-source inference server, formerly Triton Inference Server, for serving models from TensorRT, PyTorch, ONNX and more in production.

   Developer platforms
   Open sourceVisit NVIDIA Dynamo-Triton (official site)
6. 

   SambaNova

   Fast inference API for open-source models like DeepSeek, Llama and gpt-oss on SambaNova's RDU chips, billed per million tokens.

   Developer platformsCoding
   FreemiumVisit SambaNova (official site)

Untether AI FAQ

Sources & verification

Facts on this page come from the official sources below, each with the date we read it. Written by Debuly; the vendor did not review this page.

Image credits

• Logo: none collected yet, so a monogram is shown.
• Cover: none collected yet, so a generated tile is shown.

Spot an error? Report an issue with this listing. We also remove a logo or image on the owner's request.