High fan idle on Maxsun dual B60's

[cdscustoms-coder]

Hello,

1. Seems most of xpu-smi is not wired up yet. No switches for fan mode. I have 4 Maxsun B60 cards and with the docker running and model loaded the fans are at least half speed despite excellent case cooling. I spoke to Maxsun about this, they mentioned an Intel plugin might be causing the issue. Far too loud for an AI workstation.

Oddly when I run small inference the fans spool down to a reasonable level.

Can we expect updated drivers and xpu-smi soon?

2. I'm seeing 16 t/s on llama 3.3-70b-instruct with tp=8. Dual Xeon Gold 6430 CPUs so it's cross NUMA. I did ask on Intel's discrete GPUs forum about running Ubuntu server vs Desktop, I'm on desktop, could that be having an effect.

3. Ubuntu 25.04 is nagging me for the 25.10 update, doubt I should risk it at this point.

Lastly sorry to post here, but this is where the Intel mods sent me with these questions. #1 is a serious issue though, lower fans during inference and high idle seems like inverted fan curves and is very annoying. Won't matter when it's in a rack but its a pain for building Gradio and model testing.

[adamhaosuser]

Adding to this for #1 - I'm using a single SPARKLE B60 and the fan speed idles at about ~1200 rpm. If I increase the load on the card gradually the temperature seems to settle around 54c and the fan at 2200rpm.

If I suddenly introduce a large workload it'll spike to ~5300rpm and stay there - it's outrageously loud. My working theory is the card's default curve is trying to get the spike in temperature under control and overcompensates.

This is no issue on Windows with the Intel Arc Control app because you can set a fan curve. Any sort of fan control for Ubuntu would be super helpful.

[cdscustoms-coder]

Adding to this for #1 - I'm using a single SPARKLE B60 and the fan speed idles at about ~1200 rpm. If I increase the load on the card gradually the temperature seems to settle around 54c and the fan at 2200rpm.

If I suddenly introduce a large workload it'll spike to ~5300rpm and stay there - it's outrageously loud. My working theory is the card's default curve is trying to get the spike in temperature under control and overcompensates.

This is no issue on Windows with the Intel Arc Control app because you can set a fan curve. Any sort of fan control for Ubuntu would be super helpful.

Mine often idle at 2-3k but with just the docker running and model loaded, no inference, I can see 6k rpm. Ask a question, they almost go silent unless it's a long response, then ramp right back up again. In other words it seems like mine have default high idle and dynamic load curve that will DROP fan speed at light inference. Seems backwards.

[hzdzkjdxyjs]

以下建议希望对你有用
1.检查主板及bios配置是否符合要求，主板是否适配B60显卡，bios是否支持x8x8
2.我想知道您说的风扇是B60风扇还是服务器主板的风扇？
3.如果是显卡问题，我也很好奇这个问题，我下周将会尝试在ubuntu系统下复现一下，如果确实存在，我会和铭瑄团队check一下，但是我很好奇您是怎么查看风扇频率的，xpu-smi似乎没有提供查询接口

[cdscustoms-coder]

BIOS is already set to x8/x8 bifurcation and I’ve tweaked all the NUMA-related options. The board is a T3DGQ from ttyinfo.com (Gen4/Gen5 Xeon scalable support). It took me a while to figure out the exact PCIe slot mapping in BIOS, but all eight cards show up correctly in both xpu-smi and lspci -v | grep "VGA".

The noisy fans are definitely the B60 cards themselves (not the server chassis fans). You can see the card fan speeds with s-tui or Mission Center — interestingly both tools report three fans per card, with the two non-existent ones showing 0 RPM (probably because the driver/firmware is still based on desktop Arc parts). Motherboard/system fans are only visible through IPMI/BMC or ipmitool sensors.

Exactly — even the latest xpu-smi still has most fields not wired up yet (no fan speed readout/control, no memory usage, etc.).
The hardware has huge potential, but the software stack is still very half-baked right now. Hoping the upcoming big driver/firmware release fixes these monitoring and control gaps.
For reference: I’m getting ~16 t/s on the full-precision FP16 Llama 3.3 70B Instruct with TP=8. Maxsun thought it should be higher, but I’m on a dual-CPU (cross-NUMA 4+4) setup, so the UPI bottleneck is probably the main limiter right now.

Still no clear answer on upgrading to Ubuntu 25.10 — might just try mainline or a newer Intel-specific kernel on top of 25.04 first.

BIOS 已经设置成 x8/x8 分叉，也把所有跟 NUMA 相关的选项都调过了。主板是 ttyinfo.com 的 T3DGQ（支持 4/5 代 Xeon 可扩展处理器），PCIe 5.0。之前花了不少时间才把 BIOS 里的槽位对应关系搞清楚，目前 8 张卡在 xpu-smi 和 lspci -v | grep "VGA" 里全部正常显示。
噪音大的风扇肯定是 B60 显卡本身的风扇（不是机箱或主板风扇）。可以用 s-tui 或者 Mission Center 看到显卡风扇转速，有意思的是这两个工具都显示每张卡有三个风扇，另外两个不存在的风扇显示 0 RPM（估计是驱动/固件沿用了桌面 Arc 的代码）。主板系统风扇只能通过 BMC 或 ipmitool sensors 看到。
对，目前最新的 xpu-smi 还有很多功能没接上（风扇转速无法显示和调节、显存占用也看不到等等）。
硬件潜力很大，但现在的软件栈确实还很半成品，希望接下来那波大的驱动/固件更新能把监控和调控功能补齐。
顺便汇报一下性能：全精度 FP16 的 Llama 3.3 70B Instruct，TP=8，目前大概 16 t/s。铭瑄原本预期更高，但我这边是双路 CPU（跨 NUMA 4+4），UPI 带宽瓶颈应该就是主要限制。
关于升不升 Ubuntu 25.10 还是没明确答案，我可能先在 25.04 上试试 mainline 或者更新的 Intel 专用内核。
Thanks again for willing to reproduce the fan issue next week — really appreciate it!

[cdscustoms-coder]

Btw attached

T3DGQ_PCIe_CPU_Map_CPU1_1-2-5_CONFIRMED.pdf

is my slot mapping report for this sort of unicorn motherboard. Available on eBay from a few sellers. Fits in a Phanteks Enthoo Pro 2 Server case. I'm running a Superflower Leadex Platinum 2800w 240vac PSU and all Noctua Industrial PPC fans.

I'm only adding this for people looking for a dual CPU skip slot board that can natively hold four double wide cards. It's a solid board but the bios is confusing to get bifurcation right so I documented it. No affiliation at all.

I do know Maxsun now has a single CPU motherboard that will support four of these cards. I did not know that when I started my build.

[hzdzkjdxyjs]

xpu-smi最新的驱动是可以看到显存占用的，但是GPU利用率是有bug的只能显示11.1或22.2，前两天和intel的团队沟通，据说后面xpu-smi及其驱动会支持24.04版本的，我想大概你只要固定内核，就可以升级系统了，和ubuntu版本无关和内核有关吧

[hzdzkjdxyjs]

顺带我想请教您，如何在linux系统下获得GPU风扇转速信息，这对于复现问题很重要，因为xpu-smi本质上是一个将传感器数据转为人类能阅读数据的软件栈，如果xpu-smi都无法获取数据的话，那别的软件可以获得风速信息吗，或许我明天要到机房里，听一下转速了，虽然机房本身就很吵

[cdscustoms-coder]

English:
Interesting that you need root/sudo to see all the stats in xpu-smi.
By the way, if you want to monitor fan speed right now:

Terminal: try s-tui. Sometimes it crashes on the first Python run with Intel GPUs, just restart it a few times and it usually loads successfully. Once it’s up, it can show fan speeds on many Intel Arc/Data Center GPUs.
If you have a desktop/GUI: Mission Center version 1.1.0 (or newer) also displays fan speeds correctly for Intel dGPUs.

Chinese（中文）：
有意思，xpu-smi 要用 root 或 sudo 才能看到完整的统计数据。
顺便说一下，如果你现在就想监控风扇转速，可以试试这两个工具：

命令行：s-tui（https://github.com/amanusk/s-tui），第一次跑有时候会 Python crash，多重启几次基本就能正常打开，很多 Intel Arc/数据中心 GPU 上都能显示风扇转速。
如果有图形界面：Mission Center 1.1.0（以及更新的版本）也能正确显示 Intel 独显的风扇转速。

希望能帮到你～不用明天特地跑机房听风扇了哈哈

[hzdzkjdxyjs]

不需要用root哈，可以设置权限的

[hzdzkjdxyjs]

即使我复现了问题，我更要check一下GPU实际风扇风速了，因为官方提供的xpu-smi或者GPU驱动类问题是intel提供的，铭瑄主要是设计了GPU主板，如果仅仅只是s-tui显示不准确，但实际转速正常，那我无法针对这个问题去向铭瑄或是intel反馈，（因为这不是他们的技术栈）除非是他们的技术栈出现了问题，比如实际转速变快了变慢和s-tui一致。
而且一开始我也有疑问，这里仅仅只是设计AI模型推理的项目，并不涉及B60驱动相关的问题，即使您反馈了相关问题，可能也得不到解答。
所以我明天大概的思路是，也欢迎您提出建议
1.我会找到一台铭瑄的双芯服务器，安装s-tui，检查转速情况
2.启动vllm运行模型，进行压测，使用s-tui检查转速情况
3.若显示转速确有下降，我需要进入吵闹的机房，检查GPU转速的情况
4.若的确存在该情况，我会拍摄测试GPU转速变化的视频，直接向铭瑄和intel技术团队反馈情况
5.我会记录我操作的每一步骤，主要是驱动，docker等情况
非常凑巧，刚好我是个非常热爱工作的海绵宝宝类型的实习生，我很愿意探索一下～^_^
也想和您确认一下，您有没有确认过实际转速和s-tui是否一致

[cdscustoms-coder]

English version:
I’ve done some verification by comparing the readings between s-tui and Mission Center — both seem to be using lm-sensors as the backend.
There’s something I first noticed on the A-series cards and now I’m seeing the exact same behavior in xpu-smi on these B-series cards: Intel really likes keeping them pinned at exactly 50 °C all the time. My old Arc A770 behaved the same way.
I loaded the model and ran watch -n2 xpu-smi stats -d7. You can see the results in the attached screenshot (the model was refactoring ~800 lines of Python code at the time). Power draw stayed around 100 W and temperatures were in the mid-60 °C range throughout a long inference run, yet s-tui was still reporting 5000–6000 RPM on the fans.
Chinese version（中文版）：
我已经通过把 s-tui 和 Mission Center 的读数做对比进行了一定程度的验证，两者看起来底层都是调的 lm-sensors。
我之前在 A 系列显卡上就发现的现象，现在在 B 系列的 xpu-smi 里又看到了：Intel 特别喜欢把温度牢牢控制在 50 °C。我以前那张 Arc A770 也是完全一样。
我加载了模型，然后跑了 watch -n2 xpu-smi stats -d7，结果见附件截图（当时模型正在重构大概 800 行 Python 代码）。整个长时间推理过程中功耗只有 100 W 左右，温度在中 60 °C 区间，但 s-tui 依然显示风扇转速 5000–6000 RPM。

[hzdzkjdxyjs]

海绵宝宝忍不住晚上尝试了一下复现该情况，这是一款我设计的支持B60的应用，我在0和1卡上部署了模型，并且我也看了后端的情况，似乎风扇并没有转动异常，我做了这样的测试

1.未部署模型时显卡的情况，0卡仅有miner-u小模型，1卡无模型，您可以看到的显存情况

似乎没有风扇是正常的

2.当我放入模型，并进行对话，注意我本地的工作流是调用0卡的miner-u和1卡的llm，但是风扇也没有变化

3.那就是我明天即将进行的测试，长压测试，我会进行不同并发数量的测试，以验证这个情况，但是我想表达一点哈，服务器和服务器是不一样的，我使用的服务器是我们公司定制的服务器，他的风扇非常猛，所以可以严格的控制温度，B60是双芯卡，他的显存发热是非常严重的，比起蓝戟的卡，而且温度会直接导向gpu外壳，所以您的服务器如果不是专用服务器，那可能压不住这个温度，因为我的服务器运行这三个模型快七天了，但是温度一直控制的很好，这是我的软件显示的gpu监控情况

[cdscustoms-coder]

It looks like you're running vllm on windows and not intel's linux docker?

I noticed something. Someone is listening! I ran updates yesterday, still loud fans at idle. I ran updates tonight, much different fan performance.

Honestly the only reason I care is that currently this is a workstation as I build the front and backend. Previously it was very annoying, much better now. I'm going to let it idle for several hours and see if I see a speedup. I have noticed that was happening in the past, hours of loaded but not used and it would be very loud.

看起来你是在 Windows 上直接跑 vLLM，而不是用 Intel 官方的 Linux Docker 容器？
我发现了一件事——有人在听啊！
昨天我更新完系统，闲置时风扇还是很吵；今天晚上又更新了一次，风扇表现完全不一样了。
老实说，我之所以在意这个，只是因为现在这台机器还在当工作站用，我得一边写前端一边写后端。以前闲置时噪音真的很烦人，现在明显好太多了。
我准备让它再闲置几个小时观察一下，看看会不会恢复高速转速。以前就遇到过这种情况：负载挂在那里好几个小时但没实际使用，风扇就会一直狂转很吵。

[hzdzkjdxyjs]

并不是windows跑的vllm哦，我们也是ubuntu25.04跑的intel的docker，如果我能直接在windows上跑，那我可以当intel院士了，只是我给他做了后端和前端界面，所以用户可以在网页使用B60显卡部署模型，并进行硬件和模型的运维，虽然我们是服务器厂家，但我也算isv厂商吧，我们自己对B60做了前后端，方便客户使用，老实说我前两天参加了intel的展会，但是好像没人对我的软件感兴趣QAQ。

据我所知，intel的docker的团队仅仅只是负责优化vllm的，并不会直接优化底层驱动，所以可能是别的原因导致的，我现在有点怀疑您的服务器风扇问题，您如果是工作站的话，说实话，那散热并不会很优秀，而我们是服务器，所以会有很优秀的散热性能