我手头的P106-100和P104-100，AI绘图做个简单对比_p106吧

先说我的配置

分别对比了三张图
第一张

branch, plum bossom, pink|blue _flower, petals, 1girl, pink|blue theme, long_hair, solo, petals_on_liquid,portrait, lace ,lace trim,Large black frame
Negative prompt: (worst quality, low quality:1.4), monochrome, zombie,EasyNegative,nsfw, bad-picture-chill-75v, badhandv4, easynegative, verybadimagenegative_v1.2-6400, verybadimagenegative_v1.3
Steps: 20, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 1340814394, Size: 512x512, Model hash: fa98706258, Model: 2.5D：koreanstyle25D_koreanstyle25DBaked, Clip skip: 2, ENSD: 31337
p106：花费时间: 32.80秒 Torch active/reserved: 1724/2312 MiB, Sys VRAM: 3814/6144 MiB (62.08%)
P104：花费时间: 26.42秒 Torch active/reserved: 2586/2956 MiB, Sys VRAM: 4491/8192 MiB (54.82%)
第二张

branch, plum bossom, pink|blue _flower, petals, 1girl, pink|blue theme, long_hair, solo, petals_on_liquid,portrait, lace ,lace trim,Large black frame
Negative prompt: (worst quality, low quality:1.4), monochrome, zombie,EasyNegative,nsfw, bad-picture-chill-75v, badhandv4, easynegative, verybadimagenegative_v1.2-6400, verybadimagenegative_v1.3
Steps: 40, Sampler: DPM++ 2S a Karras, CFG scale: 7, Seed: 1340814394, Size: 600x800, Model hash: fc2511737a, Model: 真实系：chilloutmix_NiPrunedFp32Fix, Clip skip: 2, ENSD: 31337
p106：花费时间: 2分 7.04秒 Torch active/reserved: 1781/2952 MiB, Sys VRAM: 4472/6144 MiB (72.79%)
P104：花费时间: 1分 20.06秒Torch active/reserved: 2780/3470 MiB, Sys VRAM: 5015/8192 MiB (61.22%)
第三张

branch, plum bossom, pink|blue _flower, petals, 1girl, pink|blue theme, long_hair, solo, petals_on_liquid,portrait, lace ,lace trim,Large black frame
Negative prompt: (worst quality, low quality:1.4), monochrome, zombie,EasyNegative,nsfw, bad-picture-chill-75v, badhandv4, easynegative, verybadimagenegative_v1.2-6400, verybadimagenegative_v1.3
Steps: 40, Sampler: DPM++ 2S a Karras, CFG scale: 7, Seed: 1340814394, Size: 1024x1024, Model hash: 17277fbe68, Model: 二次元：CounterfeitV30_v30, Clip skip: 2, ENSD: 31337
p106：花费时间: 6分 12.27秒 Torch active/reserved: 1929/4434 MiB, Sys VRAM: 5954/6144 MiB (96.91%)
P104：花费时间: 4分 4.04秒 Torch active/reserved: 3619/5092 MiB, Sys VRAM: 6637/8192 MiB (81.02%)
三张图使用了不同的模型，分辨率不一样，第一个图steps20，其他两图steps40。
总的来说，p104效率是p106的1.5倍，价格是1.2倍，单说AI出图确实性价比更高。
另外，p102效率应该是p104的1.5倍，价格是它的近3倍，优点是内存大。
具体情况，有卡的可以用我上边参数出图试试。
下边是今天查的价格和性能参考表

之前完整的性能图，价格有变化，已无参考价值

送TA礼物

IP属地:河北

1楼2023-05-01 16:38回复

价格图里那个1060 6G应该是p106-100，两者性能相近

IP属地:河北

2楼2023-05-01 16:43

收起回复

我的p106是华硕冰骑士，p104是微星的，均未超频，实际使用的都是无悔大大的531版驱动

IP属地:河北

3楼2023-05-01 16:48

收起回复

P102 10g 跟 tesla M40 24G 价格差不多，不知道跑图速度差多少？
炼丹当然M40强。

IP属地:广东

4楼2023-05-01 16:56

收起回复

这张图里40Hx有两个，一个是没半精度的，那矿卡是哪个呢？作图真实速度是多少？

IP属地:福建

6楼2023-05-01 17:17

收起回复

您好大神们，做手机通过USBHUB3.0连接主机投屏手机不开USB，那个投屏很吃显卡我的598GME只能开9个窗口就卡，请问用P102 104 106 K20 K40 P4 M40哪个好我花了100买了张K620也能支持10个投屏窗口同时直播。想要能支持更多更稳定的卡麻烦推荐

IP属地:浙江

7楼2023-05-01 18:05

你好 104 你用的什么驱动我的试了各种驱动装不上装上的也用不了

IP属地:广西

来自iPhone客户端8楼2023-05-01 18:17

收起回复

我是想拿p104当计算卡，我主卡588，只能安魔改驱动么，不能用英伟达官方上1070的驱动么，如果上魔改驱动方法和106一眼么，会影响主卡么

IP属地:天津

来自Android客户端9楼2023-05-01 20:59

收起回复

水一下

IP属地:江苏

来自Android客户端10楼2023-05-01 21:17

关于pascal与半精度效率，特地去试验了一下：
显卡：P104-100 8G / 驱动：531.41 / CUDA Tookit:12.1
WebUI参数：COMMANDLINE_ARGS=--medvram --xformers --opt-split-attention --opt-channelslast
WebUI环境：python: 3.10.11 / torch: 2.0.0+cu118 / xformers: 0.0.18 / gradio: 3.23.0
测试（默认值）： 20采样 / Euler a / 512*512
Half: avg 1.6it/s
16.19sTorch active/reserved: 2119/3596 MiB, Sys VRAM: 5281/8192 MiB (64.47%)
16.11sTorch active/reserved: 2119/3598 MiB, Sys VRAM: 5283/8192 MiB (64.49%)
16.08sTorch active/reserved: 2127/3602 MiB, Sys VRAM: 5303/8192 MiB (64.73%)
16.25sTorch active/reserved: 2111/3582 MiB, Sys VRAM: 5249/8192 MiB (64.07%)
16.14sTorch active/reserved: 2110/3580 MiB, Sys VRAM: 5247/8192 MiB (64.05%)
16.11sTorch active/reserved: 2111/3582 MiB, Sys VRAM: 5249/8192 MiB (64.07%)
对比：--precision full --no-half --no-half-vae 或 --no-half --no-half-vae
No Half: avg 1.3it/s
22.03sTorch active/reserved: 3768/5252 MiB, Sys VRAM: 6919/8192 MiB (84.46%)
21.87sTorch active/reserved: 3778/5262 MiB, Sys VRAM: 6947/8192 MiB (84.8%)
21.89sTorch active/reserved: 3778/5264 MiB, Sys VRAM: 6949/8192 MiB (84.83%)
21.88sTorch active/reserved: 3768/5252 MiB, Sys VRAM: 6919/8192 MiB (84.46%)
21.83sTorch active/reserved: 3778/5262 MiB, Sys VRAM: 6947/8192 MiB (84.8%)
21.84sTorch active/reserved: 3778/5264 MiB, Sys VRAM: 6949/8192 MiB (84.83%)

IP属地:河北

11楼2023-05-02 00:27

收起回复

顺便贴张网上找的显卡参数，自己P了下对比图，以及p104还是挺方便超频的。用MSI Afterburner的话，核心超250+MHz，显存直接拉1000MHz没问题的，6、7分钟的图可以减少30秒左右。
个人是在降压超频用。

// P104-100 GPU : VEN_10DE&DEV_1B87&SUBSYS_123710DE&REV_A1&BUS_1&DEV_0&FN_0 //
// 1.0310V average core overclock is 158MHz, memory overclock is 200MHz //
0.6250 V == 待机
0.8430 V == 1607 MHz 默频
0.9620 V == 频率区间电压……
0.9750 V
0.9930 V
1.0000 V
1.0250 V
1.0310 V
1.0430 V == 频率区间电压……
1.0500 V == 1885 MHz 睿频
1.0620 V
1.0750 V(MAX)
=============================
1607 MHz ~ 18xx MHz
0.8430 V ~ 1.0500 V
=============================
降压超频：
AVG:1885(MAX:1911)MHz == 1.050V
AVG:2025(MAX:2037)MHz == 0.962V
+250 MHz -0.088V
+126 MHz -88mV
=============================
GPU : 2025M (+250 MHz)
MEM : 5508 (+500 MHz)
VOLT : 0.9620V (-088 mV)
-----------------------------
GPU : 2012M (+230 MHz)
MEM : 5200M (+200 MHz)
VOLT : 0.9620V (-088 mV)
=============================
测试稳定跑图时长：347m 49.12sTorch
AI绘图：默频、超显存、超核心、超核心显存，时长对比。
(AVG == 7m 22s) Benchmark
NC:1885M_Vram:5005M = 7m 21.67sTorch
NC:1885M_Vram:5005M = 7m 21.77sTorch
NC:1885M_Vram:5005M = 7m 21.33sTorch
NC:1885M_Vram:5005M = 7m 24.90sTorch
NC:1885M_Vram:5005M = 7m 26.36sTorch
(AVG == 7m 17s) -5s
OC:1885M_Vram:6003M = 7m 17.41sTorch
OC:1885M_Vram:6003M = 7m 19.43sTorch
OC:1885M_Vram:6003M = 7m 16.45sTorch
(AVG == 7m 01s) -16s
OC:2025M_Vram:5005M = 7m 02.27sTorch
OC:2025M_Vram:5005M = 7m 01.25sTorch
OC:2025M_Vram:5005M = 7m 00.37sTorch
OC:2025M_Vram:5005M = 7m 00.72sTorch
OC:2025M_Vram:5005M = 7m 01.55sTorch
(AVG == 6m 56s) -26s
OC:2025M_Vram:5508M = 6m 57.75sTorch
OC:2025M_Vram:5508M = 6m 56.67sTorch
OC:2025M_Vram:5508M = 6m 58.22sTorch
OC:2025M_Vram:5508M = 6m 56.00sTorch
OC:2025M_Vram:5508M = 6m 55.20sTorch
(AVG == 6m 50s) -32s
OC:2025M_Vram:6003M = 6m 53.46sTorch
OC:2025M_Vram:6003M = 6m 51.29sTorch
OC:2025M_Vram:6003M = 6m 48.47sTorch
OC:2025M_Vram:6003M = 6m 48.22sTorch
OC:2025M_Vram:6003M = 6m 50.82sTorch

IP属地:河北

12楼2023-05-02 00:38

收起回复

画这些要干嘛，网上全是这种ai生成人

IP属地:福建

13楼2023-05-02 11:40

收起回复

p106-90和-100什么区别

IP属地:湖北

来自Android客户端14楼2023-05-02 23:23

收起回复

第一个图重新测了一下，上边数据有些问题，第一次出图时间长，这次比第四次出图的时间
branch, plum bossom, pink|blue _flower, petals, 1girl, pink|blue theme, long_hair, solo, petals_on_liquid,portrait, lace ,lace trim,Large black frame
Negative prompt: (worst quality, low quality:1.4), monochrome, zombie,EasyNegative,nsfw, bad-picture-chill-75v, badhandv4, easynegative, verybadimagenegative_v1.2-6400, verybadimagenegative_v1.3
Steps: 20, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 1340814394, Size: 512x512, Model hash: fa98706258, Model: 2.5D：koreanstyle25D_koreanstyle25DBaked, Clip skip: 2, ENSD: 31337
原始第四次：Time taken: 22.03sTorch active/reserved: 2593/2962 MiB, Sys VRAM: 4497/8192 MiB (54.9%)
原始关半精度第四次：Time taken: 17.81sTorch active/reserved: 4864/5742 MiB, Sys VRAM: 7295/8192 MiB (89.05%)
改动源文件第四次：Time taken: 18.83sTorch active/reserved: 2464/3022 MiB, Sys VRAM: 4557/8192 MiB (55.63%)

IP属地:河北

15楼2023-05-02 23:42

收起回复

日	一	二	三	四	五	六

我手头的P106-100和P104-100，AI绘图做个简单对比

登录百度账号

扫二维码下载贴吧客户端