如何繞過matlab 弱化AMD CPU功能 - 3C
By Hamiltion
at 2019-11-29T14:08
at 2019-11-29T14:08
Table of Contents
: 推 a58524andy : 懶人包:if amd ignore avx2 sse3 sse4 11/28 11:35
: → windrain0317: 就matlab用了intel mkl啊 11/28 11:39
: → a58524andy : 裡面是說mkl開哪些指令是依據vendor而非指令列表 11/28 11:40
: → a58524andy : 所以用一些環境變數硬讓他開起來ryzen就扳回一城惹 11/28 11:41
: → a58524andy : 測試結果來說i+mkl還是有些優勢啦,像是矩陣乘法那 11/28 11:42
: → a58524andy : 項,18c36t的10980xe還是贏過24c48t的3960x 11/28 11:43
: → a58524andy : 3970x才贏過10980xe 11/28 11:43
: 推 ariadne : AMD也可以自己出MKL 這方面intel真的有下成本 11/28 11:47
: 推 doom3 : 對手CPU只用SSE1 也太針對了吧 11/28 13:34
: 推 twosheep0603: numpy系列好像也有MKL的問題 改參數會起飛XD 11/28 15:31
: → pig : 堆土機時代根本不會有人在乎這個啊 AVX1/2 有開跟沒 11/28 15:42
: → pig : 開差不了多少吧 XD 11/28 15:43
: → pig : 到了 zen 才開始有人注意,用的人多了自然就會改正 11/28 15:44
有玩機械學習寫 Python Code 的,如果換香香機應該也會想:
喵的怎麼有時候 3700x 跑不贏不知到多少年前的 E3-1230v2
對,我就是跑 SVM 跟 Naive Bayes 時越想越不對勁
然後我找到了 Reddit 討論串跟 GitHub 上的資源
http://bit.ly/2L3w5Fy , Reddit, Ryzen and Intel's Anti-competitive MKL
http://bit.ly/2R0sD2j , Ryzen 3900X vs Intel Xeon 2175W, Python numpy
重點就是這段,尤其開發環境是 Anaconda 的用戶應該連署去逼社群改寫這種垃圾奧步
Intel's MKL check the CPUID is GenuineIntel or not, if detected the non-intel cpu, MKL will choose the "maximum capability" code (i.e. SSE2 - slowset) Intel's "cripple AMD" function Anaconda's numpy use Intel TBB instead of OpenMP, Intel TBB use the intel's proprietary method to detect the CPU or NUMA topology, in this situation zen's SMT will be recognize to the real core, it hurt the ALU performance.
GitHub 上有改善方案,但我怕把 Anaconda 炸掉所以觀望
http://bit.ly/34vjvqr
但光是用 n_jobs 這個參數去強迫 sklearn 使用 Ryzen 的所有核心,都能把原來的運算時間壓縮到 20% 甚至更少
https://i.imgur.com/Dqfz9ta.png
https://i.imgur.com/XAua4cn.png
這招真的是 intel 效能輸人家的時候最愛在編譯器搞的爛招
我想應該最早可以追溯到 2000 前後的 3DMark,就是那個 Pentium4 被 Athlon 鎚爆的時代
對, AMD Ryzen 的 Python Performance 就是輸在 intel 把二十年前的爛招回鍋
--
while(user==alone){user=find(girlfriend);}
return user;
--
: → windrain0317: 就matlab用了intel mkl啊 11/28 11:39
: → a58524andy : 裡面是說mkl開哪些指令是依據vendor而非指令列表 11/28 11:40
: → a58524andy : 所以用一些環境變數硬讓他開起來ryzen就扳回一城惹 11/28 11:41
: → a58524andy : 測試結果來說i+mkl還是有些優勢啦,像是矩陣乘法那 11/28 11:42
: → a58524andy : 項,18c36t的10980xe還是贏過24c48t的3960x 11/28 11:43
: → a58524andy : 3970x才贏過10980xe 11/28 11:43
: 推 ariadne : AMD也可以自己出MKL 這方面intel真的有下成本 11/28 11:47
: 推 doom3 : 對手CPU只用SSE1 也太針對了吧 11/28 13:34
: 推 twosheep0603: numpy系列好像也有MKL的問題 改參數會起飛XD 11/28 15:31
: → pig : 堆土機時代根本不會有人在乎這個啊 AVX1/2 有開跟沒 11/28 15:42
: → pig : 開差不了多少吧 XD 11/28 15:43
: → pig : 到了 zen 才開始有人注意,用的人多了自然就會改正 11/28 15:44
有玩機械學習寫 Python Code 的,如果換香香機應該也會想:
喵的怎麼有時候 3700x 跑不贏不知到多少年前的 E3-1230v2
對,我就是跑 SVM 跟 Naive Bayes 時越想越不對勁
然後我找到了 Reddit 討論串跟 GitHub 上的資源
http://bit.ly/2L3w5Fy , Reddit, Ryzen and Intel's Anti-competitive MKL
http://bit.ly/2R0sD2j , Ryzen 3900X vs Intel Xeon 2175W, Python numpy
重點就是這段,尤其開發環境是 Anaconda 的用戶應該連署去逼社群改寫這種垃圾奧步
Intel's MKL check the CPUID is GenuineIntel or not, if detected the non-intel cpu, MKL will choose the "maximum capability" code (i.e. SSE2 - slowset) Intel's "cripple AMD" function Anaconda's numpy use Intel TBB instead of OpenMP, Intel TBB use the intel's proprietary method to detect the CPU or NUMA topology, in this situation zen's SMT will be recognize to the real core, it hurt the ALU performance.
GitHub 上有改善方案,但我怕把 Anaconda 炸掉所以觀望
http://bit.ly/34vjvqr
但光是用 n_jobs 這個參數去強迫 sklearn 使用 Ryzen 的所有核心,都能把原來的運算時間壓縮到 20% 甚至更少
https://i.imgur.com/Dqfz9ta.png
https://i.imgur.com/XAua4cn.png
這招真的是 intel 效能輸人家的時候最愛在編譯器搞的爛招
我想應該最早可以追溯到 2000 前後的 3DMark,就是那個 Pentium4 被 Athlon 鎚爆的時代
對, AMD Ryzen 的 Python Performance 就是輸在 intel 把二十年前的爛招回鍋
--
while(user==alone){user=find(girlfriend);}
return user;
--
Tags:
3C
All Comments
By Steve
at 2019-12-03T12:23
at 2019-12-03T12:23
By Robert
at 2019-12-03T21:53
at 2019-12-03T21:53
By Emma
at 2019-12-04T19:03
at 2019-12-04T19:03
By Rebecca
at 2019-12-07T18:11
at 2019-12-07T18:11
By Edwina
at 2019-12-09T06:45
at 2019-12-09T06:45
By Mia
at 2019-12-11T06:26
at 2019-12-11T06:26
By Lydia
at 2019-12-13T03:21
at 2019-12-13T03:21
By Damian
at 2019-12-17T06:55
at 2019-12-17T06:55
By Wallis
at 2019-12-19T13:12
at 2019-12-19T13:12
By Lydia
at 2019-12-21T19:15
at 2019-12-21T19:15
By Yuri
at 2019-12-21T22:18
at 2019-12-21T22:18
By Valerie
at 2019-12-25T06:04
at 2019-12-25T06:04
By Andrew
at 2019-12-28T03:03
at 2019-12-28T03:03
By Lydia
at 2019-12-29T18:25
at 2019-12-29T18:25
By Mia
at 2020-01-03T10:07
at 2020-01-03T10:07
By Rae
at 2020-01-05T16:10
at 2020-01-05T16:10
By Enid
at 2020-01-07T15:28
at 2020-01-07T15:28
By Isla
at 2020-01-08T01:00
at 2020-01-08T01:00
By Harry
at 2020-01-08T03:58
at 2020-01-08T03:58
By Rae
at 2020-01-12T22:51
at 2020-01-12T22:51
By Belly
at 2020-01-13T04:00
at 2020-01-13T04:00
By Franklin
at 2020-01-17T05:30
at 2020-01-17T05:30
By Emily
at 2020-01-20T21:48
at 2020-01-20T21:48
By Jacob
at 2020-01-21T02:51
at 2020-01-21T02:51
By Ivy
at 2020-01-25T22:56
at 2020-01-25T22:56
By Ina
at 2020-01-28T00:11
at 2020-01-28T00:11
By Skylar Davis
at 2020-01-30T14:47
at 2020-01-30T14:47
By Cara
at 2020-01-31T03:07
at 2020-01-31T03:07
By Kristin
at 2020-02-03T17:48
at 2020-02-03T17:48
By Michael
at 2020-02-07T14:14
at 2020-02-07T14:14
By Tom
at 2020-02-09T11:12
at 2020-02-09T11:12
By Genevieve
at 2020-02-13T23:29
at 2020-02-13T23:29
By Sierra Rose
at 2020-02-18T22:09
at 2020-02-18T22:09
By Kama
at 2020-02-22T20:36
at 2020-02-22T20:36
By Ula
at 2020-02-23T21:41
at 2020-02-23T21:41
By John
at 2020-02-27T12:29
at 2020-02-27T12:29
By Ula
at 2020-02-27T16:10
at 2020-02-27T16:10
By Liam
at 2020-03-03T01:52
at 2020-03-03T01:52
By Poppy
at 2020-03-04T07:51
at 2020-03-04T07:51
By Odelette
at 2020-03-08T18:45
at 2020-03-08T18:45
By Anthony
at 2020-03-08T23:04
at 2020-03-08T23:04
By Delia
at 2020-03-13T21:44
at 2020-03-13T21:44
By Carolina Franco
at 2020-03-14T15:03
at 2020-03-14T15:03
By George
at 2020-03-15T04:21
at 2020-03-15T04:21
By Eartha
at 2020-03-20T03:10
at 2020-03-20T03:10
By Lily
at 2020-03-23T16:48
at 2020-03-23T16:48
By Agnes
at 2020-03-24T21:12
at 2020-03-24T21:12
By Joseph
at 2020-03-27T14:38
at 2020-03-27T14:38
By Charlotte
at 2020-03-28T22:06
at 2020-03-28T22:06
By Connor
at 2020-03-29T19:34
at 2020-03-29T19:34
By Sarah
at 2020-04-01T13:10
at 2020-04-01T13:10
By Michael
at 2020-04-03T10:20
at 2020-04-03T10:20
By Ethan
at 2020-04-05T16:10
at 2020-04-05T16:10
By Oliver
at 2020-04-06T20:27
at 2020-04-06T20:27
By Lucy
at 2020-04-10T12:40
at 2020-04-10T12:40
By Vanessa
at 2020-04-10T20:45
at 2020-04-10T20:45
By Franklin
at 2020-04-14T08:41
at 2020-04-14T08:41
By Liam
at 2020-04-16T01:17
at 2020-04-16T01:17
By Joseph
at 2020-04-20T06:51
at 2020-04-20T06:51
By Oliver
at 2020-04-25T05:12
at 2020-04-25T05:12
By Liam
at 2020-04-27T15:10
at 2020-04-27T15:10
By Frederic
at 2020-04-28T07:35
at 2020-04-28T07:35
Related Posts
什麼地方可以幫忙切割機殼?
By Thomas
at 2019-11-29T13:17
at 2019-11-29T13:17
WD Black SN750 法亞黑五特價中
By Daph Bay
at 2019-11-29T12:23
at 2019-11-29T12:23
3700X和9700K怎麼選?
By Charlotte
at 2019-11-29T10:57
at 2019-11-29T10:57
技嘉發布世界首款 USB 3.2 Gen2x2 擴充卡
By Valerie
at 2019-11-29T08:57
at 2019-11-29T08:57
華碩 ROG STRIX 2080Ti 官方白化版!
By Tristan Cohan
at 2019-11-29T08:43
at 2019-11-29T08:43