Re: [情報] GMP 建議不要用 Zen5
看板PC_Shopping (個人電腦購買)作者zseineo (Zany)時間4小時前 (2025/08/28 09:34)推噓38(39推 1噓 281→)留言321則, 32人參與討論串2/2 (看更多)
https://gmplib.org/gmp-zen5
連結內有一些CPU的圖跟兩台的配備
全部貼來PTT好像有點長,我截下面比較重要的部分+AI翻
We really don’t know what is wrong here. Let us list some observations and
thoughts:
The ambient temperature was in both cases quite low.
These are supposedly top-quality motherboard (We’ve had about 50 ASUS
motherboard over the years, and only had one failure).
The offset mounting does not look right to us. But this is how things are
supposed to be done.
It is possible that we caused the 2nd CPU failure by inadequate mounting of
the heat sink. It surely seems bizarre to have the thermal paste squeezed like
shown on the pictures above!
But note that the 1st failure happened with a more centred heat sink. We only
made the off-centre mounting for the 2nd system as to minimise the risk of a
repeated system failure.
The so-called TDP of the Ryzen 9950X is 170W. The used heat sinks are
specified to dissipate 165W, so that seems tight. We have extra fans in the
cases and a low ambient temperature, making these 5W seem negligible. But of
course, this is not completely kosher.
We have a sister system for the 9950X, a 7950X which is similarly configured
and built. That system actually seems to run slightly hotter, but it has been
stable for a long time (under the same crazy load!).
The systems were under maximal load at the time their CPUs died, running very
tight handcrafted asm loops sustaining one MULX per cycle. That might be a “
hot” instruction.
Did the CPUs die of heat stroke? Modern CPUs measure their temperature and
clock down if they get too hot, don’t they?
We don’t overclock or overvolt or play other teen games with our hardware.
The extremely thin layer of thermal paste which resulted from the off-centre
heat sink mount, might seem fine. We however suspect that there might be a
problem, as when the system heats and cools, things bend a little. With a
remaining layer of thermal paste, such bending will not create any void as the
paste’s elasticity will maintain some contact. When the thermal paste is
squeezed away as with the suggested Noctua mounting, an ever-so-slight void
between the CPU and the heat sink might be created.
Neither of the 9950X CPUs died immediately, instead they died the exact same
way after a couple of months at high load. This seems to suggest a gradual but
predictable degradation.
我們真的不清楚問題出在哪裡。以下是一些我們的觀察與想法:
兩次故障發生時,環境溫度都相當低。
我們使用的是公認品質頂尖的主機板
(多年來我們用過約 50 張 ASUS 主機板,只壞過一張)。
偏置安裝看起來很奇怪,但這卻是官方建議的安裝方式。
第二次 CPU 故障有可能是我們安裝散熱器不當所致。散熱膏被擠壓成上圖那樣,確實有
點匪夷所思。
但值得注意的是,第一次故障時散熱器是置中安裝的。我們在第二套系統上改用偏置安裝
,正是為了避免重蹈覆轍。
Ryzen 9950X 的所謂熱設計功耗(TDP)是 170W,而我們使用的散熱器解熱能力規格為
165W,規格上確實有點吃緊。不過考量到我們在機殼內加裝了風扇,且環境溫度低,這
5W 的差距應該可以忽略。但嚴格來說,這確實不是最標準的作法。
我們有一套與 9950X 系統配置相近的 7950X 系統。那套系統的運作溫度似乎還稍微高一
些,但在同樣極端的負載下,它已經穩定運行了很長時間。
CPU 燒毀的當下,系統都處於滿載狀態,持續運行著我們手動編寫、極度精簡的 asm 迴
圈,每個時脈週期執行一次 MULX 指令。這或許是個發熱量極大的指令。
CPU 是因為過熱燒毀的嗎?現代 CPU 不是都有溫度監控,會在過熱時自動降頻保護嗎?
我們從不對硬體進行超頻、加壓或任何類似的胡亂設定。
偏置安裝導致散熱膏層極薄,表面上可能沒問題。但我們懷疑,當系統反覆升溫和降溫時
,金屬會產生微小的形變。如果還有一層散熱膏,膏體的彈性可以填補縫隙,維持接觸。
但若像 Noctua 建議的安裝方式那樣將散熱膏完全擠開,CPU 和散熱器之間就可能因形變
產生一道極其微小的空隙。
兩顆 9950X CPU 都不是立即損壞,而是在高強度負載幾個月後,以完全相同的方式故障
。這似乎指向一種緩慢但可預測的漸進式損耗。
關於文中提到的off-centre mounting
We use a Noctua cooling solution for both systems. For the 1st system, we
mounted the heat sink centred. For the 2nd system, we followed Noctua’s
advice of mounting things offset towards what they claim to be the hotter side
of the CPU. Below is a picture of the 2nd system without the heat sink which
shows that offset. Note the brackets and their pins, those pins are where the
heat sink’s pressure gets centred. Also note how the thermal paste has been
squeezed away from that part, but is quite thick towards the left.
我們的兩套系統都採用 Noctua 散熱器。第一套系統,我們將散熱器置中安裝。到了第二
套系統,我們則遵循 Noctua 的建議,將散熱器偏向他們所謂 CPU 較熱的一側進行安裝
。下方是第二套系統未安裝散熱器的照片,可以清楚看到偏置安裝的狀況。請注意支架和
卡榫,散熱器的壓力就集中在這些卡榫上。同時也請注意,散熱膏如何從壓力集中處被完
全擠開,但在左側卻保留了相當的厚度。
----
因為天氣夠冷所以散熱用稍微不夠是可以的嗎?尤其是在大量數學計算上?
積熱應該還是會積在裡面吧?熱交換不夠快外面再涼也沒意義?
(編輯)
補充:看推文討論溫度可能不是主要原因
--
※ 發信站: 批踢踢實業坊(ptt.cc), 來自: 1.160.16.3 (臺灣)
※ 文章網址: https://www.ptt.cc/bbs/PC_Shopping/M.1756344840.A.5CD.html
※ 編輯: zseineo (1.160.16.3 臺灣), 08/28/2025 09:39:47
→
08/28 09:42,
4小時前
, 1F
08/28 09:42, 1F
→
08/28 09:42,
4小時前
, 2F
08/28 09:42, 2F
推
08/28 09:45,
4小時前
, 3F
08/28 09:45, 3F
→
08/28 09:46,
4小時前
, 4F
08/28 09:46, 4F
→
08/28 09:47,
4小時前
, 5F
08/28 09:47, 5F
推
08/28 09:47,
4小時前
, 6F
08/28 09:47, 6F
→
08/28 09:47,
4小時前
, 7F
08/28 09:47, 7F
→
08/28 09:47,
4小時前
, 8F
08/28 09:47, 8F
→
08/28 09:48,
4小時前
, 9F
08/28 09:48, 9F
→
08/28 09:48,
4小時前
, 10F
08/28 09:48, 10F
推
08/28 09:50,
4小時前
, 11F
08/28 09:50, 11F
→
08/28 09:50,
4小時前
, 12F
08/28 09:50, 12F
→
08/28 09:51,
4小時前
, 13F
08/28 09:51, 13F
→
08/28 09:52,
4小時前
, 14F
08/28 09:52, 14F
→
08/28 09:53,
4小時前
, 15F
08/28 09:53, 15F
→
08/28 09:53,
4小時前
, 16F
08/28 09:53, 16F
→
08/28 09:55,
4小時前
, 17F
08/28 09:55, 17F
→
08/28 09:55,
4小時前
, 18F
08/28 09:55, 18F
→
08/28 09:57,
4小時前
, 19F
08/28 09:57, 19F
→
08/28 09:57,
4小時前
, 20F
08/28 09:57, 20F
→
08/28 09:58,
4小時前
, 21F
08/28 09:58, 21F
→
08/28 09:58,
4小時前
, 22F
08/28 09:58, 22F
推
08/28 10:02,
4小時前
, 23F
08/28 10:02, 23F
→
08/28 10:02,
4小時前
, 24F
08/28 10:02, 24F
→
08/28 10:02,
4小時前
, 25F
08/28 10:02, 25F
→
08/28 10:03,
4小時前
, 26F
08/28 10:03, 26F
→
08/28 10:03,
4小時前
, 27F
08/28 10:03, 27F
→
08/28 10:03,
4小時前
, 28F
08/28 10:03, 28F
→
08/28 10:06,
4小時前
, 29F
08/28 10:06, 29F
→
08/28 10:06,
4小時前
, 30F
08/28 10:06, 30F
推
08/28 10:06,
4小時前
, 31F
08/28 10:06, 31F
→
08/28 10:06,
4小時前
, 32F
08/28 10:06, 32F
→
08/28 10:07,
4小時前
, 33F
08/28 10:07, 33F
→
08/28 10:07,
4小時前
, 34F
08/28 10:07, 34F
→
08/28 10:07,
4小時前
, 35F
08/28 10:07, 35F
推
08/28 10:07,
4小時前
, 36F
08/28 10:07, 36F
→
08/28 10:07,
4小時前
, 37F
08/28 10:07, 37F
→
08/28 10:08,
4小時前
, 38F
08/28 10:08, 38F
→
08/28 10:08,
4小時前
, 39F
08/28 10:08, 39F
還有 243 則推文
→
08/28 12:12,
2小時前
, 283F
08/28 12:12, 283F
→
08/28 12:13,
2小時前
, 284F
08/28 12:13, 284F
→
08/28 12:13,
2小時前
, 285F
08/28 12:13, 285F
推
08/28 12:13,
2小時前
, 286F
08/28 12:13, 286F
→
08/28 12:13,
2小時前
, 287F
08/28 12:13, 287F
→
08/28 12:13,
2小時前
, 288F
08/28 12:13, 288F
→
08/28 12:13,
2小時前
, 289F
08/28 12:13, 289F
→
08/28 12:14,
2小時前
, 290F
08/28 12:14, 290F
→
08/28 12:14,
2小時前
, 291F
08/28 12:14, 291F
→
08/28 12:14,
2小時前
, 292F
08/28 12:14, 292F
→
08/28 12:15,
2小時前
, 293F
08/28 12:15, 293F
→
08/28 12:15,
2小時前
, 294F
08/28 12:15, 294F
※ 編輯: zseineo (1.160.16.3 臺灣), 08/28/2025 12:16:07
→
08/28 12:15,
2小時前
, 295F
08/28 12:15, 295F
→
08/28 12:15,
2小時前
, 296F
08/28 12:15, 296F
→
08/28 12:17,
2小時前
, 297F
08/28 12:17, 297F
→
08/28 12:17,
2小時前
, 298F
08/28 12:17, 298F
→
08/28 12:17,
2小時前
, 299F
08/28 12:17, 299F
→
08/28 12:18,
2小時前
, 300F
08/28 12:18, 300F
→
08/28 12:18,
2小時前
, 301F
08/28 12:18, 301F
推
08/28 12:18,
2小時前
, 302F
08/28 12:18, 302F
推
08/28 12:18,
2小時前
, 303F
08/28 12:18, 303F
→
08/28 12:18,
2小時前
, 304F
08/28 12:18, 304F
→
08/28 12:18,
2小時前
, 305F
08/28 12:18, 305F
推
08/28 12:19,
2小時前
, 306F
08/28 12:19, 306F
→
08/28 12:19,
2小時前
, 307F
08/28 12:19, 307F
→
08/28 12:19,
2小時前
, 308F
08/28 12:19, 308F
→
08/28 12:19,
2小時前
, 309F
08/28 12:19, 309F
→
08/28 12:20,
2小時前
, 310F
08/28 12:20, 310F
→
08/28 12:21,
2小時前
, 311F
08/28 12:21, 311F
→
08/28 12:21,
2小時前
, 312F
08/28 12:21, 312F
→
08/28 12:22,
2小時前
, 313F
08/28 12:22, 313F
→
08/28 12:22,
2小時前
, 314F
08/28 12:22, 314F
→
08/28 12:23,
2小時前
, 315F
08/28 12:23, 315F
推
08/28 12:23,
2小時前
, 316F
08/28 12:23, 316F
→
08/28 12:23,
2小時前
, 317F
08/28 12:23, 317F
→
08/28 12:24,
2小時前
, 318F
08/28 12:24, 318F
→
08/28 12:24,
2小時前
, 319F
08/28 12:24, 319F
推
08/28 12:25,
2小時前
, 320F
08/28 12:25, 320F
→
08/28 12:25,
2小時前
, 321F
08/28 12:25, 321F
討論串 (同標題文章)
PC_Shopping 近期熱門文章
PTT數位生活區 即時熱門文章