File size: 1,768 Bytes
a30d21b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
---
license: mit
pretty_name: amd-experiments MI300X KV-cache compression results
tags:
  - kv-cache
  - quantization
  - rocm
  - amd-mi300x
  - triton
  - benchmark
---

# amd-experiments: MI300X KV-cache compression benchmark results

Benchmark result files for [amd-experiments](https://github.com/Bobliuuu/amd-experiments): a four-way comparison of **TurboQuant**, **IsoQuant**, **PlanarQuant** and **RotorQuant** KV-cache compression on an AMD Instinct MI300X (gfx942) with Triton/ROCm kernels and Mistral-7B-v0.1 shapes.

**No model weights are released.** The GitHub repo contains the kernels and benchmark scripts (and a README, video and report); this dataset repo holds the raw `bench_*.json` result files and the derived summary `readme_data.json` only.

## Headline (from `results/readme_data.json`)

| method | compression | compress GB/s | prefill K tok/s at 32K | cosine (3-bit) |
|---|---|---|---|---|
| TurboQuant (TQ3) | 4.92x | 2.9 | 42 | 0.983 |
| IsoQuant | 4.92x | 21.8 | 892 | 0.983 |
| PlanarQuant | 4.92x | 18.7 | 1,126 | 0.983 |
| RotorQuant | 4.92x | 17.3 | 855 | 0.983 |

- A fused Triton TQ3 attention kernel is 0.28x of fp16 SDPA speed at 1K tokens and about 1.9x at 32K (B=1).
- In vLLM serving of Mistral-7B (1024 in / 256 out, 32 prompts) output throughput is flat across fp16, TurboQuant decompress and TurboQuant fused (2,425 to 2,435 tok/s).
- Single GPU model, one LLM family; the needle-in-a-haystack test is synthetic.

## Files

`results/bench_*.json`: one file per benchmark (compress/decompress bandwidth, prefill, perplexity and cosine quality, NIAH, fused-attention timings, vLLM A/B). Load with `json.load`.
Scripts to regenerate them: https://github.com/Bobliuuu/amd-experiments (`benchmarks/`, `run_all_benchmarks.sh`).