97 lines
4.1 KiB
Markdown
97 lines
4.1 KiB
Markdown
# sayuri
|
||
|
||
## Hardware
|
||
|
||
HP Z440 workstation.
|
||
|
||
* [Intel Xeon E5-2683 v4](https://ark.intel.com/content/www/us/en/ark/products/91766/intel-xeon-processor-e52683-v4-40m-cache-2-10-ghz.html)
|
||
* 4×4 GiB DDR4 2400 MHz ECC memory
|
||
* 250GB Samsung 970 Evo Pro NVMe SSD
|
||
* 256GB micron MTFDDAK256TBN-1AR15ABHA SATA SSD
|
||
* 2TB Toshiba HDWA120 HDD
|
||
* Sapphire Nitro+ Radeon RX 480 (8 GiB VRAM)
|
||
|
||
### Mods
|
||
|
||
#### Fan
|
||
|
||
The original fans are really loud if you run them at a higher speed.
|
||
What fans are used depend on the exact model of the Z440,
|
||
mine had a Delta QUR0912VH as rear case fan,
|
||
a Delta AFB0912VH as front fan
|
||
and a Foxconn PVA092G12S as CPU fan.
|
||
Since the firmware only allows dynamic fan control via Intel QST,
|
||
which is not supported in the kernel,
|
||
the “minimum fan speed” set in the firmware configuration is always used.
|
||
I replaced all three fans (rear case fan, front “PCIe” fan, CPU fan) with aftermarket products.
|
||
For the rear case fan and CPU fan I used Noctua NF-A9 PWM fans,
|
||
for the front fan an Arctic F9 PWM (for the sole reason that it is cheaper).
|
||
Since HP decided to use different connectors for all of the fans
|
||
(with the CPU fan connector having a proprietary 6-pin connector),
|
||
I had to get creative with plugging them in.
|
||
One alternative suggested by Michael Stapelberg in [his article on fan replacement in a HP Z440](https://michael.stapelberg.ch/posts/2021-08-28-silent-hp-z440-workstation/)
|
||
is to remove the guard rails of the fans.
|
||
Since this is destructive and I might want to sell the fans again if I don’t use them anymore,
|
||
I decided against this
|
||
and instead opted for connecting them manually with cheap jumper cables.
|
||
This certainly is not the prettiest solution, but it works.
|
||
As for the CPU fan,
|
||
the fifth and sixth wire are actually not needed
|
||
and a 4-pin PWM fan can be plugged into the connector (with jumpers).
|
||
|
||
#### CPU
|
||
|
||
The original CPU that came in my model was an Intel Xeon E5-1620 v4 CPU
|
||
with 4 cores (8 threads),
|
||
a base clock speed of 3.5 GHz
|
||
and a boost clock speed of 3.8 GHz.
|
||
To achieve better multicore performance,
|
||
I upgraded it to an Intel Xeon E5-2683 v4, which can be found used for semi-cheap on AliExpress.
|
||
It has 16 cores (32 threads),
|
||
a base clock speed of 2.1 GHz
|
||
and a boost clock speed of 3 GHz.
|
||
While the CPU worked out-of-the box with the 2020 firmware revision I had on it,
|
||
it did not offer frequency control in linux (and therefore stayed at its base clock).
|
||
Upgrading to the newest firmware did not fix this issue.
|
||
A workaround is to enable CPU HWPM in the firmware,
|
||
which strips the kernel from frequency control and instead hands it to the firmware,
|
||
which at least allows the CPU to reach 2.7 GHz
|
||
Since this setting makes it impossible to determine the CPU clock via `cpupower frequency-info`,
|
||
they have to be obtained by running `grep -E '^cpu MHz' /proc/cpuinfo` (one line for every thread).
|
||
|
||
However, all CPUs compatible with the socket of this system (2011-3) share the same problem:
|
||
They suffer from many security vulnerabilities,
|
||
the mitigation of which slows them down massively.
|
||
A possible workaround is to disable mitigations (https://make-linux-fast-again.com/),
|
||
which, however, makes the CPU vulnerable to all those attacks again!
|
||
This is implemented by the `yolo` specialisation,
|
||
which can be selected at boot.
|
||
|
||
The result of this is that,
|
||
while it has double the cores and is a desktop/server CPU,
|
||
it still is around 15 % slower (!) than my laptop’s Ryzen 7 5850U in multithreaded synthetic workloads.
|
||
It looks even worse in single-threaded workloads,
|
||
in which my laptop is five times as fast.
|
||
|
||
#### Memory
|
||
|
||
I have not yet upgraded the memory for cost reasons.
|
||
It still is the original 4 sticks of Hynix HMA451R7AFR8N-UH (4 GiB DDR4 2400 MHz ECC).
|
||
An upgrade is necessary due to many processes’ memory usage scaling linearily with CPU cores,
|
||
the prime example being compiling with `make -j32` et al.
|
||
|
||
#### Conclusion
|
||
|
||
Is it worth it?
|
||
Probably not, especially as a desktop machine.
|
||
It might come in handy at a later time as a server.
|
||
|
||
## Purpose
|
||
|
||
Tasks that require a decent amount of GPU power
|
||
and/or have to run while I do other things (on my laptop).
|
||
|
||
## Name
|
||
|
||
Sayuri Kurata is a student from *Kanon*
|