Inspur NE3412M5 Edge AI and Storage Platform Review
26 May 2021
In our Inspur NE3412M5 review, we are going to see what this unique server offers. Unlike many rack servers, this server is designed to be portable and shipped to various edge locations whether those are in buildings or in vehicles. It was specifically designed to be transportable to various edge locations so hyper-scalers and other organizations can offer local and low latency CPU and GPU compute along with storage. If you want to see a truly portable server designed to be transported in the back of a truck and deployed quickly upon arrival, this is that system from Inspur. Those same properties make for an interesting review.
Inspur NE3412M5 Hardware Overview
As we have been doing in our recent reviews, we are going to split this into external and internal versions. There is a lot to see here so it makes sense to split this up.
Inspur NE3412M5 External Hardware Overview
The system itself looks like a tower chassis, except different in a key feature. It looks like a Pelican case and a tower server were mixed together. In a way, that is exactly what this NE3412M5 is.
The chassis itself has rubber feet on the bottom and the bottom side panel to help with shock and vibration when it is deployed. While most of Inspur’s servers we review are rack-mounted, this is a machine designed for edge deployment.
The system itself when it is completely closed does not have any ports exposed. What it does have are a few panels that expose functionality below. Perhaps the most exposed functionality aside from the rubber feet is the top handle which allows easy transportation of the system.
Around that handle is a panel that can be used to store the cabling. Since this unit came directly from deployment in China, it had an appropriate regional power cable but one can use this area for other power cable options as well as to store gear such as network cables.
On each end of the chassis, there are two doors that open. We called this the “port” side as traditional front and rear labels were probably less descriptive. Here we can see an out-of-band management port, a VGA port, two 1Gbase-T ports, two SFP+ 10GbE ports, and an array of six USB 3.0 ports. There are also a large number of LEDs and power/ reset buttons on this side.
One can see two fan vents for cooling and there is also what looks like another expansion slot that is not populated. We will show the other side of that when we get to the internal overview.
The opposite side has additional fans and the power input. The power supply in our test unit is a single non-redundant PSU which is more common on devices like this since in the field there is often a single power source. Perhaps the big feature though is that one can see a double-width PCIe expansion slot with a GPU installed. This system is designed to provide AI inferencing to the edge and this is a great example of how that is accomplished. In our internal overview, we are going to get more into how this is deployed since it is a novel solution.
One item that is really nice here is that covering both port sides of the chassis the doors can be retracted into the chassis. This allows the system to be protected during transport but then have the ports exposed when it arrives at its destination.
What you may have noticed from the first few images is that the side panels are big, and have an extensive system to secure them to the chassis. There is a bit more than just that going on. This solution is designed to securely hold the doors in place while distributing the pressure on the chassis to maintain proper seals.
Inspur provides a tool to open the sides of the chassis, but one will likely want to use a drill bit instead to remove them. There are a total of twelve screws on each side, and each side may need to be accessed in the system depending on the components required. Using the hand tool, this took a while to unscrew 24 screws. One item we wish is that Inspur adds in a future generation is the ability to secure this tool in the compartment that houses the power cable around the handle. It would be great to see a dedicated spot there for the tool.
Now, that we have discussed the panels, let us get inside the system.
Inspur NE3412M5 Internal Hardware Overview
Pulling a panel off, one can see that this external plastic is not the only place where there is protection. We also see that inside the motherboard is in its own metal compartment. Here we also get a nice service guide. Since the unit we are reviewing is a configuration that is used by major Chinese hyper-scalers as an edge device, similar to how Amazon AWS has its Snowball and other similar boxes, we see the service guide is in Chinese. From what we understand, this can be swapped to a different language if necessary. Again, we are not discussing which major hyper-scaler is using this machine, but this is a popular edge deployment system.
When we showed the screws holding the side panel on, one item we wanted to note here is that the external plastic is only one level of protection. There is also an internal gasket that runs around the internal compartment to keep moisture out of the system. This gasket is a bit inside of the exterior shell so it is protected by a hard plastic buffer around the system.
Removing the primary motherboard area cover, we can see the system. Here we see how this side of the system is laid out. The CPU and memory area is covered with an airflow guide. Behind that is the PCH and M.2 storage as well as the ports. Below this area, we have the BMC and both power and PCIe connectors.
The fans are on their own partition secured by two screws. Something that is a bit different is that these fans pull air from inside the chassis and expel air outside of the chassis.
One can see that the fans have more robust power connectors than typical server 4-pin PWM fan headers we see in a system.
The system itself is a single Intel Xeon Scalable (first and second generation) design with six-channel memory and eight DIMM slots. As one would expect on an edge server like this, the heatsink is passive so the chassis fans can provide redundant cooling.
In front of the CPU and memory portion, we have a M.2 storage slot that will usually be occupied by a boot SSD. The heatsink below this is the PCH. One can see ports are on the left side of this system here.
Below this area, we have the ASPEED AST2500 BMC as well as high-density SATA as well as a traditional 7-pin SATA port.
Below the CPU area, we have a section dedicated to cabled connections. Specifically for power and PCIe connectivity. There are no traditional PCIe x16 slots on the motherboard here, so instead this is designed for cables to attach to I/O on the other side of the system. This area gets less airflow, so it makes sense that it is used for cable connectors.
On the other side of the system, we get something that is very different. The top portion is a secured metal casing for expansion. There is also cooling and power being delivered to this section that also has the power supply.
First, the big feature. The peripheral section unscrews/ latches and swivels out locking in either the up or down positions (shown above and below.) In this peripheral section, we get the ability to put 3.5″ storage, and three drives for capacity or redundancy. We also get a PCIe section for up to two cards. In our test system, we utilized a single double-width GPU. One can see the fans directing air into this peripheral section.
Underneath we can see the PCIe cables and SATA cables along with power entering the peripheral section. We can also see the FSP PSU below.
Something we mentioned earlier is that in this area one can see the empty I/O expansion slot below the fans. Our system did not have this populated but it seems as though there may be a PCI mezzanine option that could be placed here. Another interesting feature is that the bottom has two round holes and some mounting points. The faceplate we have on this system had blanks over these two holes, but we could see them being used for a variety of applications.
Overall, there is a lot of engineering that went into this solution. Our test unit has been to multiple continents and one can see has certainly seen a lot of use. Still, it worked directly upon startup.
Next, let us get to how the system performs and the management.
Inspur NE3412M5 Management
We have covered Inspur’s management interface several times at this point. Since this is a first and second-generation Intel Xeon Scalable part, as we would expect this is is a fairly standard solution that we have looked at a number of times with an out-of-band IPMI/ Redfish management interface that includes a web GUI.
Inspur generally has a solid number of advanced features in its WebGUI and these are often a step beyond the basic level we see in some machines.
There are also standard features such as an included HTML5 iKVM solution with remote media and power capabilities.
Inspur’s servers also have customized POST and splash screens. They offer a bit more information than some of the basic BIOS solutions we have seen, but then also have modes, where they look relatively clean as can be seen below that, was taken via the HTML5 iKVM.
BIOS in the system is a fairly standard American Megatrends (AMI) solution as one can see here:
Overall, Inspur’s management solution is more aligned to what we would consider an industry-standard approach rather than those proprietary solutions from Dell, Lenovo, and HPE. Still, within that industry-standard framework, Inspur has added quite a bit of functionality to differentiate it from vendors who are providing base-level vendor images that are simply rebranded with logo swaps. There is a lot more going on here.
Inspur NE3412M5 Performance
In terms of performance, we have looked at the performance of the Intel Xeon Silver 4208 and the NVIDIA GeForce RTX 2070 GPU previously. We wanted to see what this solution would provide in terms of performance compared to our baselines. Specifically, we wanted to see if this more ruggedized and mobile edge platform performed similarly to standard tower offerings or if performance was being sacrificed in order to achieve a wider range of deployment scenarios.
First, looking at performance, we are using the HPE ProLiant ML110 Gen10 as the baseline. That HPE server is not designed for portable edge deployments, but it is HPE’s mainstream tower server that we also tested with the Xeon Silver 4208. We also swapped the memory configuration from what was seen in the hardware overview to 6x 16GB DIMMs in order to get a similar baseline to the HPE system which we did the same for.
The performance here was similar to what we saw in the HPE system that is not as ruggedized. Our conclusion is that we are not losing performance. Still, this is a lower-end CPU configuration in both machines, but the Xeon Silver is popular in this segment.
Looking at GPU performance, we used our baseline results for blower-style cooler GPUs to compare here.
We will note that we tried fitting a NVIDIA GeForce RTX 3090 Founders Edition in this system but it did not fit. If you are looking to add a larger GPU to this system, we suggest a standard blower-style cooler configuration.
Overall, the key takeaway was that the performance was very similar to our control setups which means that we are not losing performance with this edge format.
Next, let us get to the STH Server Spider and our final words.
Inspur NE3412M5 Power Consumption
Our Inspur test server, we hooked this system up to our lab power meters and took a few power readings at 120V using our Extech 380803 True RMS power analyzers.
STH 70% Load: 91W
100% CPU Load: 132W
Maximum Recorded: 451W
As one would imagine, there is a lot of configurability here with CPUs, GPUs, memory, and storage. We see these more as directional guidelines in our configuration.
STH Server Spider: Inspur NE3412M5
In the second half of 2018, we introduced the STH Server Spider as a quick reference to where a server system’s aptitude lies. Our goal is to start giving a quick visual depiction of the types of parameters that a server is targeted at.
Our STH Server Spider generally measures density, but density is not the point of the NE3412M5. Instead, this system is about being mobile and being able to handle the stresses of being shipped across distances.
Overall, this is a very interesting system. Sometimes we see systems on the market that are marketed as portable but that are more of systems that simply add a handle to them. The Inspur NE3412M5 is much better designed than those systems. There are features such as the handle compartment to store cables that are differentiating. Other features such as the multiple layers of plastic, gaskets, and metal areas for protection during shipping are extremely useful as well.
The key point for our readers is that this is a system that was designed for hyper-scalers to provide an edge compute (CPU and GPU) offering with local storage to cloud customers. Inspur often engages in co-design with its large customers and this is an example of where it appears as though those design cycles have led to some major design directions.
The above photo is a very good example of this. One can see the gaskets, as well as just the thickness of the hard plastic covering that, encases the metal structures of the server and the valuable components. We did not try this, but this feels like a server that can survive a 1-2m drop onto a concrete floor unscathed.
One item we wish that the server had was easily replaceable external air filters. Some edge environments are very dusty so those would have been nice to see. This system is designed more for rugged shipping but indoor and in-vehicle operation so that may be the reason behind that design direction.
Overall the NE3412M5 is an impressive edge comptue platform that is designed to survive transportation to remote locations and provide a local outpost for compute capabilities.