0xNeoAi/llama-cpp-rx7800xt-benchmarks

llama.cpp ROCm Benchmarks on AMD Radeon RX 7800 XT

This dataset contains llama.cpp HIP/ROCm benchmark results collected on a local AMD Radeon RX 7800 XT Linux desktop. It includes both parsed result tables and the original raw benchmark logs so the numbers can be audited.

Dataset URL:

https://huggingface.co/datasets/0xNeoAi/llama-cpp-rx7800xt-benchmarks

Test system

Component	Value
OS	Ubuntu 24.04.4 LTS
Kernel	Linux 6.17.0-29-generic
CPU	AMD Ryzen 7 5700X3D, 8 cores / 16 threads
System memory	32 GiB RAM, 8 GiB swap
GPU	AMD Radeon RX 7800 XT
GPU architecture	gfx1101
VRAM reported by llama.cpp	16,368 MiB
ROCm / HIP	HIP 7.2.53211, ROCm 7.2.3 toolchain
llama.cpp commit	f12cc6d0fa96d6a3c33952f06b7439ac43a3c3fe
Backend	ROCm / HIP build of llama.cpp

Benchmark method

The normal throughput results were produced with llama-bench from the HIP build:

/home/neo/bench/src/llama.cpp/build-hip/bin/llama-bench \
  -m <model.gguf> \
  -ngl 999 \
  -p 512 \
  -n 128 \
  -b 512 \
  -ub 512 \
  -r 5

Meaning of the reported tests:

• pp512: prompt processing throughput for a 512-token prompt.
• tg128: token generation throughput for 128 generated tokens.
• Values are reported as tokens/sec ± standard deviation from the raw llama-bench output.

Files

Path	Description
llama_bench_results.csv	Parsed benchmark table rows from completed llama-bench outputs.
llama_bench_results.jsonl	Same parsed results as JSONL.
metadata.json	Host, GPU, llama.cpp commit, and notes.
raw/	Original benchmark logs, summaries, and failed/incomplete artifacts.

Results summary

Highest parsed values in this dataset:

Metric	Best result
Prompt processing	6002.21 ± 85.75 on llama-bench-LFM2.5-8B-A1B-Q8_0.gguf-20260529-100822.txt
Token generation	199.70 ± 1.92 on llama-bench-LFM2.5-8B-A1B-Q5_K_M.gguf-20260529-100802.txt

LFM2.5-8B-A1B results

Model / quant	Size	Params	Backend	pp512 tok/s	tg128 tok/s	Raw file
lfm2moe 8B.A1B Q4_K - Medium	4.79 GiB	8.47 B	ROCm	4844.05 ± 20.61	167.88 ± 1.81	llama-bench-LFM2.5-8B-A1B-Q4_K_M.gguf-20260529-100753.txt
lfm2moe 8B.A1B Q5_K - Medium	5.61 GiB	8.47 B	ROCm	4625.49 ± 33.50	199.70 ± 1.92	llama-bench-LFM2.5-8B-A1B-Q5_K_M.gguf-20260529-100802.txt
lfm2moe 8B.A1B Q6_K	6.47 GiB	8.47 B	ROCm	3336.65 ± 23.56	195.71 ± 1.27	llama-bench-LFM2.5-8B-A1B-Q6_K.gguf-20260529-100812.txt
lfm2moe 8B.A1B Q8_0	8.38 GiB	8.47 B	ROCm	6002.21 ± 85.75	171.18 ± 1.99	llama-bench-LFM2.5-8B-A1B-Q8_0.gguf-20260529-100822.txt

Qwen3-14B results

Model / quant	Size	Params	Backend	pp512 tok/s	tg128 tok/s	Raw file
qwen3 14B Q3_K - Medium	6.81 GiB	14.77 B	ROCm	1004.75 ± 31.82	42.93 ± 0.05	llama-bench-Qwen3-14B-Q3_K_M.gguf-20260528-194019.txt
qwen3 14B Q4_K - Medium	8.38 GiB	14.77 B	ROCm	1091.35 ± 23.42	36.75 ± 0.19	llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-112104.txt
qwen3 14B Q4_K - Medium	8.38 GiB	14.77 B	ROCm	1095.16 ± 17.51	36.83 ± 0.02	llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-112415.txt
qwen3 14B Q4_K - Medium	8.38 GiB	14.77 B	ROCm	1091.85 ± 17.52	36.76 ± 0.02	llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-114637.txt
qwen3 14B Q4_K - Medium	8.38 GiB	14.77 B	ROCm	1086.99 ± 21.08	36.53 ± 0.07	llama-bench-Qwen3-14B-Q4_K_M.gguf-20260528-194128.txt
qwen3 14B Q5_K - Medium	9.79 GiB	14.77 B	ROCm	1036.95 ± 17.77	41.57 ± 0.01	llama-bench-Qwen3-14B-Q5_K_M.gguf-20260528-194247.txt
qwen3 14B Q2_K - Medium	5.65 GiB	14.77 B	ROCm	1064.51 ± 16.34	49.66 ± 0.08	llama-bench-Qwen3-14B-UD-Q2_K_XL.gguf-20260528-193947.txt
qwen3 14B Q3_K - Medium	7.07 GiB	14.77 B	ROCm	1038.61 ± 18.82	43.11 ± 0.04	llama-bench-Qwen3-14B-UD-Q3_K_XL.gguf-20260528-194054.txt
qwen3 14B Q4_K - Medium	8.53 GiB	14.77 B	ROCm	1087.87 ± 31.11	38.14 ± 0.02	llama-bench-Qwen3-14B-UD-Q4_K_XL.gguf-20260528-194210.txt
qwen3 14B Q5_K - Medium	9.82 GiB	14.77 B	ROCm	1047.31 ± 15.11	40.31 ± 0.05	llama-bench-Qwen3-14B-UD-Q5_K_XL.gguf-20260528-194327.txt
qwen3 14B Q6_K	12.37 GiB	14.77 B	ROCm	1101.35 ± 30.22	36.60 ± 0.05	llama-bench-Qwen3-14B-UD-Q6_K_XL.gguf-20260528-194409.txt

Notes:

• Several Qwen3-14B Q4_K_M runs are included because the benchmark was repeated during setup and validation.
• UD-* and non-UD quants are kept with their original source labels as reported by llama-bench.

Qwen3.6 27B result

Model / quant	Size	Params	Backend	pp512 tok/s	tg128 tok/s	Raw file
qwen35 27B Q3_K - Medium	12.86 GiB	27.32 B	ROCm	549.09 ± 5.50	23.39 ± 0.03	llama-bench-Qwen3.6-27B-Q3_K_M.gguf-20260528-124239.txt

Incomplete and invalid artifacts

The following raw files are included for transparency, but are not included in llama_bench_results.csv / .jsonl as completed benchmark rows:

Qwen3.6 35B A3B MTP speculative attempt

File:

raw/qwen36-35b-a3b-q3km-mtp-speculative-20260529-103518.txt

This is not a valid throughput benchmark. All attempted llama-speculative runs failed immediately with:

--model-draft is required

Do not compare this file against the normal llama-bench results.

Qwen3-14B UD-Q8_K_XL incomplete run

File:

raw/llama-bench-Qwen3-14B-UD-Q8_K_XL.gguf-20260528-194455.txt

This raw file contains benchmark metadata and the Markdown table header only; it has no completed pp512 or tg128 result rows. It is preserved as raw evidence, but intentionally excluded from the parsed result files.

Reproducibility notes

• The raw logs include model paths, hashes where captured, exact command lines, and raw llama-bench tables.
• The parsed CSV/JSONL files are derived only from completed Markdown table rows in raw/llama-bench-*.txt.
• Failed or incomplete runs are documented above rather than mixed into the parsed result tables.

License

The benchmark metadata and result tables in this dataset are released under the MIT license. Original model licenses remain with their respective model authors and repositories.