提高内存频率性能会提高，但是发热太厉害，我又给降回来了
先改善散热吧，这个肯定是最大的瓶颈，然后把内存频率尽可能加高 deviljz 发表于 2019-12-9 11:04
有帖子地址吗，我去仔细看看
https://www.extremetech.com/computing/302650-how-to-bypass-matlab-cripple-amd-ryzen-threadripper-cpus
https://bbs.saraba1st.com/2b/thread-1866511-1-1.html
https://www.pugetsystems.com/labs/hpc/How-To-Use-MKL-with-AMD-Ryzen-and-Threadripper-CPU-s-Effectively-for-Python-Numpy-And-Other-Applications-1637/
https://www.reddit.com/r/matlab/comments/dxn38s/howto_force_matlab_to_use_a_fast_codepath_on_amd/ albertfu 发表于 2019-12-8 22:28
不知道什么专门的软件可以看
你可以试试hwinfo记录一下，跑同样测试，只有内存频率不同，CPU的SoC、CPU...
2400
2934
我仔细观察了下，这个87度不是上来就87度，是一点点升上去的，一开始70多
deviljz 发表于 2019-12-9 12:18
2400
2934
看了下你发的图，2933时SoC电压可以适当调低些
默认给了1.1V以上，实际1V~1.05V比较合适，高了并不有利于内存上高频。SoC电压调低也能减少10W左右功耗，210W降到200W也是好的
albertfu 发表于 2019-12-9 03:07
2950X已经是双路了，2990WX其实是四路，而且两个NUMA节点没法直接访问内存...
嗯 4ccx 两组内存控制器
所以算起来的话其实是两个 双ccx算一个节点
说起来amd至今没做出多路系统
—— 来自 vivo NEX S, Android 9上的 S1Next-鹅版 v2.1.0-play albertfu 发表于 2019-12-9 13:30
看了下你发的图，2933时SoC电压可以适当调低些
默认给了1.1V以上，实际1V~1.05V比较合适，高了并不有利 ...
选了xmp1档，自动就这样，我去找找在哪改 albertfu 发表于 2019-12-9 13:30
看了下你发的图，2933时SoC电压可以适当调低些
默认给了1.1V以上，实际1V~1.05V比较合适，高了并不有利 ...
soc电压改成1.075v，降了一点点
看了下这机箱能装9个机箱风扇，我打算装满 deviljz 发表于 2019-12-9 15:25
soc电压改成1.075v，降了一点点
SoC Power降了12W，和预估的差不多
package power不变估计是core power上去了，吃掉了SoC部分节约的功耗，而package power受制于散热规模上不去
搜了一下hardocp做过2990WX配合银箭TR4的评测，可惜这人跳槽outel之后把hardocp网站关掉了，现在只能看google cache，看不到图：
The delta between the Thermalright and the Cooler Master on this run is a minimal ~50MHz. Even with our 2990WX it can still edge out a tiny lead, which I find somewhat impressive. Of course the standout here is our XSPC cooler which allows our 2990WX to run a full ~666MHz faster than the Silver Arrow TR4...devilish. If you look however, both our air cooler systems are running clocks lower than they were with Precision Boost 2. Certainly that is not good. PBO is penalizing us on air compared to PB2.
Keep in mind, that our 2990WX has a specified TDP of 250 watts. The Silver Arrow TR4 run shows to be using 283 watts package power. The Wraith Ripper system package power is 42 watts lower. I think this spread between the two air coolers does not seem quite "correct." Surely as the 2990WX get hotter, its power usage becomes exponential and certainly does not ramp linearly, but it still seemed like a big bump to me. If you look, the Cooler Master cooler is actually below TDP. Certainly the water cooled 2990WX is rolling through the power. While 442 watts might seem excessive, we have seen package power above 650 watts doing hand overclocking here recently. No matter how you look it, PBO is an issue for our air coolers in terms of clocks and power.
水冷能比银箭快666MHz，package power能压到442W，银箭能压制283W package power
你这里只有210W的话，可能是机箱温度比hardocp的测试环境温度要高
先试试银箭上挂满3风扇吧，高马甲条的话竖直机箱没法固定风扇，我的做法是主板平放，用胶带把TY-143扇子固定住 你的全家 发表于 2019-12-9 14:01
嗯 4ccx 两组内存控制器
所以算起来的话其实是两个 双ccx算一个节点
互联大多用在socket内了，导致最多就2 socket（不过也128C了还要啥自行车，牙膏28C*4还没这个多 albertfu 发表于 2019-12-9 16:13
SoC Power降了12W，和预估的差不多
package power不变估计是core power上去了，吃掉了SoC部分节约的功耗 ...
https://item.jd.com/100003803069.html
这机箱能装9个机箱风扇，我打算装满试试 我用hwinfo看了下我i7-7700的温度……99度，快能烧开水了。而且我在这个状态下断断续续跑了几个月 宏 发表于 2019-12-9 19:07
一体水也就比风冷高40w，300出头，到不了400w的
https://www.tomshardware.com/reviews/amd-ryzen-threa ...
32C讲真现成卖的散热方案都是渣渣，除了个别420冷排的一体水，没什么方案压得住440W的
一体水就自由多了，上手先怼个1080排
lz这个又是工作用机不好上水，只能尽量挖掘风冷潜力了 deviljz 发表于 2019-12-9 19:03
我用hwinfo看了下我i7-7700的温度……99度，快能烧开水了。而且我在这个状态下断断续续跑了几个月 ...
牙膏干了吧，换换散热，不嫌吵，加加暴力风扇
—— 来自 Sony H8296, Android 9上的 S1Next-鹅版 v2.1.2 今天装了3个机箱风扇，温度基本没变化，峰值87.5，功率峰值234
我自己的i7换了个风扇倒是降了30度
页:
1
[2]