Spaces:
Running
Running
Update README.md
Browse files
README.md
CHANGED
|
@@ -1,10 +1,86 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
-
|
| 3 |
-
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
|
|
|
|
| 8 |
---
|
| 9 |
|
| 10 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
title: README
|
| 3 |
+
emoji: π»
|
| 4 |
+
colorFrom: purple
|
| 5 |
+
colorTo: pink
|
| 6 |
+
sdk: static
|
| 7 |
+
pinned: false
|
| 8 |
+
license: mit
|
| 9 |
+
short_description: High-efficiency LLM acceleration engine
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
<h1 align="center">Anima Core Inc.</h1>
|
| 13 |
+
<h3 align="center">High-efficiency LLM acceleration for real-world AI inference</h3>
|
| 14 |
+
|
| 15 |
+
<p align="center">
|
| 16 |
+
<b>7β11Γ faster attention on NVIDIA H100 NVL</b><br>
|
| 17 |
+
<b>~90% lower energy per token</b><br>
|
| 18 |
+
<b>Software-only, no custom hardware required</b>
|
| 19 |
+
</p>
|
| 20 |
+
|
| 21 |
+
---
|
| 22 |
+
|
| 23 |
+
### π AN1 Acceleration Engine
|
| 24 |
+
The AN1 Engine provides drop-in accelerated attention and matrix operations for PyTorch LLM inference.
|
| 25 |
+
It achieves competitive speedups to dedicated AI hardware but runs on standard NVIDIA GPUs.
|
| 26 |
+
|
| 27 |
+
| Feature | Value |
|
| 28 |
+
|--------:|:------|
|
| 29 |
+
| Acceleration | 7.21Γ to 11.3Γ faster |
|
| 30 |
+
| Energy Savings | ~90% lower joules/token |
|
| 31 |
+
| Hardware | H100 NVL, A100, L40S, more |
|
| 32 |
+
| Integration | PyTorch (vLLM & TensorRT adapters coming) |
|
| 33 |
+
| Availability | Production pilots by request |
|
| 34 |
+
|
| 35 |
+
---
|
| 36 |
+
|
| 37 |
+
### π Benchmarked on NVIDIA H100 NVL
|
| 38 |
+
|
| 39 |
+
```text
|
| 40 |
+
Baseline PyTorch (fp16, 2048 seq):
|
| 41 |
+
Latency: 11.63 ms
|
| 42 |
+
Tokens/sec: 1.41M
|
| 43 |
+
TFLOPs: 47.27
|
| 44 |
+
|
| 45 |
+
AN1 Accelerated (fp16, 2048 seq):
|
| 46 |
+
Latency: 1.36 ms
|
| 47 |
+
Tokens/sec: 12.04M
|
| 48 |
+
TFLOPs: 404.01
|
| 49 |
+
|
| 50 |
+
Speedup: 7.65Γ
|
| 51 |
+
Energy Savings: ~90%
|
| 52 |
+
```
|
| 53 |
+
|
| 54 |
+
---
|
| 55 |
+
|
| 56 |
+
### π Public Repository
|
| 57 |
+
π https://github.com/Anima-Core/an1-engine
|
| 58 |
+
|
| 59 |
+
---
|
| 60 |
+
|
| 61 |
+
### π§ͺ Pilot Access (Private GPU Backend)
|
| 62 |
+
The CUDA backend (`an1_core_gpu`) remains proprietary.
|
| 63 |
+
To request access for benchmarking or integration:
|
| 64 |
+
|
| 65 |
+
π© **pilot@animacore.ai**
|
| 66 |
+
|
| 67 |
+
Please include:
|
| 68 |
+
- Name
|
| 69 |
+
- Organization
|
| 70 |
+
- GPU hardware available
|
| 71 |
+
|
| 72 |
---
|
| 73 |
+
|
| 74 |
+
### π§ Research Interest
|
| 75 |
+
- Efficient LLM inference
|
| 76 |
+
- Software-based structured reuse for GPU compute
|
| 77 |
+
- Symbolic and neuro-symbolic acceleration
|
| 78 |
+
- Meaning-based computational models
|
| 79 |
+
|
| 80 |
---
|
| 81 |
|
| 82 |
+
### π± About Anima Core
|
| 83 |
+
Anima Core builds AI systems focused on secure, ethical, and computationally efficient machine intelligence.
|
| 84 |
+
We believe performance and responsibility belong together.
|
| 85 |
+
|
| 86 |
+
π https://www.animacore.ai
|