Huawei TaiShan 200 2280 Server Block Diagram
Here is the block diagram for the server:
Most of these components we have covered in our hardware overview. There were some challenges with the system. As an example, we could find the HNS adapter on the PCIe bus, but it seems very picky with optics.
The other one is that the Realtek OCP NIC 3.0 did not show up in the system. Here is a quick lscpu output from the Kunpeng 920, and the above is the topology output from when we initially powered on the system.
Here is the lshw output for the CPU that we found to be a Kunpeng 920-4826 or a 48 core 2.6GHz part:
product: ARM (To be filled by O.E.M.)
physical id: 2b
bus info: cpu@0
version: Kunpeng 920-4826
configuration: cores=48 enabledcores=48 threads=48
For those wondering, here is the HiSilicon PCIe Gen4 bridge in this system:
description: PCI bridge
product: HiSilicon PCIe Root Port with Gen4
vendor: Huawei Technologies Co., Ltd.
physical id: 100
bus info: pci@0000:00:00.0
width: 32 bits
capabilities: pci pciexpress msi pm normal_decode bus_master cap_list
resources: irq:27 ioport:1000(size=4096) memory:e6500000-e66fffff ioport:80000200000(size=2097152)
We can post more output if folks are interested in the hardware devices.
This turned into a huge piece without even going in-depth on the Kunpeng 920 Arm CPUs. We are going to have the Kunpeng 920 piece in the next few days. While we are still working on getting the system’s performance up, the firmware side is fairly clear. The Huawei server feels more like what we experienced with the Cavium/ Marvell ThunderX2 systems rather than the more modern Ampere Arm servers. This is a server that is fighting us more than normal. It took hours of work just to get what we would consider proper PCIe Gen4 NVMe SSD performance. Our general advice is to get Ampere Altra (Max) over this. Many companies in China are choosing Ampere over these Huawei Kunpeng 920 chips, and it seems like that is for good reason. Ampere Altra Max is twice the core count per socket and is socketed.
In terms of surprises beyond the CPU, the Broadcom SAS controller is one that perhaps we should have expected, but without a Broadcom NIC onboard nor a standard ASPEED BMC, we thought we would get a different solution.
Aside from the HNS 25GbE RDMA adapter, these two cards were clearly the most interesting in the server. We have a quad 1GbE OCP NIC 3.0 card using less than $4 of NICs and a PCIe Gen4 x16 connector. There is also a BMC based on a chip we could not find information on via Google.
In terms of overall quality, the Huawei TaiShan 200 (2280) was also surprising. Huawei is set up to have margins more akin to Cisco. We were expecting the quality to be similar to Cisco. Instead, from the Chinese domestic server vendors we review, the Inspur servers have higher quality. Lenovo servers are probably a step above Huawei. Then this Huawei server would be third. There are a lot of very good mechanical design elements in the TaiShan 200, but working on the server, there are fairly clear differences.
Overall, the hope is that this look at a Huawei server is one that folks find useful. We will have a look at the Kunpeng 920 very soon, then more content as parts arrive and more in this server gets working. Still, there was so much here, and this was also a very ambitious project to undertake, so we had to split it up.
We realize the comments section may get lively on this piece, but please keep it civil. We wanted to simply share this because these servers exist, yet there is not much information on them.