Strata 0.1.31 + 3090Ti+ 3080 + 64GB RAM

[Thanaots]

LLM 本地推理 Benchmark

Hardware

Item	Configuration
GPU	NVIDIA GeForce RTX 3090 Ti, 23028 MiB NVIDIA GeForce RTX 3080, 10240 MiB
CPU	Intel(R) Core(TM) Ultra 7 270K Plus
RAM	60Gi
OS	Ubuntu 26.04 LTS
Kernel	7.0.0-38-generic
Driver	595.91.07
Backend	llama.cpp

Benchmark Settings

Parameter	Value
Tool	llama-benchy
PP	2048
TG	128
Depth	16K / 32K / 64K / 120K
Concurrency	1
Runs	3
Cache	Disabled
TG Length	Exact
Latency Mode	generation

Results

model	test	t/s	peak t/s	ttfr (ms)	est_ppt (ms)	e2e_ttft (ms)
Qwen/Qwen3.8-Flash-Next	pp2048 @ d16384	487990.14 ± 4576.01		39.82 ± 0.36	37.78 ± 0.36	7600.50 ± 239.32
Qwen/Qwen3.8-Flash-Next	tg128 @ d16384	93.71 ± 4.94	94.33 ± 4.92
Qwen/Qwen3.8-Flash-Next	pp2048 @ d32768	549207.46 ± 18264.55		65.51 ± 2.16	63.47 ± 2.16	13641.16 ± 93.11
Qwen/Qwen3.8-Flash-Next	tg128 @ d32768	104.92 ± 6.12	105.67 ± 6.18
Qwen/Qwen3.8-Flash-Next	pp2048 @ d65536	602404.29 ± 2694.60		114.24 ± 0.50	112.19 ± 0.50	25405.68 ± 138.26
Qwen/Qwen3.8-Flash-Next	tg128 @ d65536	95.69 ± 5.91	96.33 ± 5.73
Qwen/Qwen3.8-Flash-Next	pp2048 @ d120000	620944.40 ± 4961.17		198.61 ± 1.57	196.57 ± 1.57	46767.65 ± 10.61
Qwen/Qwen3.8-Flash-Next	tg128 @ d120000	99.18 ± 4.63	99.33 ± 4.50

Notes

• PP = Prompt Processing
• TG = Token Generation
• 120K depth is used for the ~128K context test so PP/TG still fit inside a 128K context window.
• Results are mean ± standard deviation across 3 runs.
• Prefix cache avoidance enabled with --no-cache.
• Fixed generation length enabled with --exact-tg.

[terry]

我弟，这特么不能翻译成汉语吗？

[Thanaots]

@terry 说:

我弟，这特么不能翻译成汉语吗？

我没在ubuntu上装中文软件。然后跑的内容全在上面。只能勉强先达成英文得了。Windows版本的数据我就可以上中文了。
主要是，我现在打字用的是小鹤双拼+讯飞输入法。不喜欢搜狗的流氓广告。讯飞好像不支持linux。未来ubuntu用的不多。