Etched Sohu ASIC
Vendor: Etched
Architecture: Transformer ASIC
Category: ASIC Dedicated Accelerator
Introduction
An ASIC chip hard-coded specifically for the Transformer architecture. It removes redundant logic used for graphics rendering and general-purpose computing in GPUs, claiming to be several orders of magnitude faster than the NVIDIA H100 on large models such as Llama.
Specifications
| Model | Compute | Memory | Interface | TDP | Process |
|---|---|---|---|---|---|
| Sohu | Transformer-dedicated (multiple times H100) | External HBM | PCIe 5.0 | TBA | TBA |
Official Website
OS Support
| Windows | Linux | macOS | Android |
|---|---|---|---|
| ❌ | ✅ (Etched Cloud API) | ❌ | ❌ |
Version History
| Version | Release Date | Description |
|---|---|---|
| Sohu Pre-release | 2025 | Claims Llama inference speed far exceeds H100 |
Performance Benchmarks
| Model | Task | Performance Metric |
|---|---|---|
| Sohu | Llama 2 70B Inference | Ultra-high throughput (official data) |
| Sohu | Transformer Inference | Dedicated Transformer inference acceleration |
Pricing
| Model | Reference Price | Notes |
|---|---|---|
| Sohu | Cloud API | Etched Cloud API |
| Sohu | Contact vendor | Enterprise deployment |
Quick Installation
Etched Cloud (API)
pip install etched-sdk
Sohu is a dedicated Transformer ASIC, supporting only Transformer architecture model inference.
Code Examples
Python (Etched API)
from etched import EtchedClient
client = EtchedClient(api_key="your-key")
response = client.generate(
model="llama-3-70b",
prompt="Hello",
max_tokens=100
)
Architecture Highlights
- Transformer ASIC: The world's first ASIC designed specifically for the Transformer architecture, implementing attention mechanisms directly in hardware
- Extreme Inference: Bypasses the flexibility overhead of general-purpose computing, pushing Transformer inference efficiency to the limit
- Limitations: Only supports Transformer architecture models; does not support CNNs, graph neural networks, etc.
Model Compatibility
| Model/Framework | Support | Notes |
|---|---|---|
| Transformer LLM | ✅ Native | Llama/GPT/Qwen etc. |
| CNN Models | ❌ | Not supported |
| Non-Transformer | ❌ | Transformer architecture only |
Related Products
If you're evaluating alternatives, the following products may also fit your scenario:
- Groq LPU v1 — Groq (LPU Language Processor)
- Cerebras WSE-3 — Cerebras (ASIC Dedicated Accelerator)
- Tenstorrent Wormhole — Tenstorrent (ASIC Dedicated Accelerator)
- NVIDIA GPU / CUDA — NVIDIA (GPU Graphics Processor)
- AWS Trainium / Inferentia — Amazon AWS (ASIC Dedicated Accelerator)
- Intel Gaudi (Habana) — Intel (ASIC Dedicated Accelerator)
- Cambricon Siyuan MLU — Cambricon (ASIC Dedicated Accelerator)