HP, IBM and Dell are the names that come to mind when talking about servers, HPC and blade systems. Outside of the big three are a number of other vendors targeting this space. One of those, Supermicro, is currently winning awards for its power efficient power supplies. It is also a major supplier to the white label OEM market where it's beginning to make a name for itself in HPC.
Tau Leng, General Manager, Program Management, Marketing & HPC, Supermicro recently sat down with DCT to talk about what they see happening in the market.
Given the increased interest being shown by companies in the use of HPC inside their business how far away are we from commodity pricing?
I see commodity hardware and software from the last 10 years playing an important role in HPC. Today we cannot accommodate the future needs of software on generalised hardware. We are already seeing the transition to application optimised hardware for most systems.
Even with HPC you will have certain customisation. These are not standard servers
People want to have sharing between the cooling and power supply. The memory of each machine is just 4GB in a 2U. This is about efficiency not about commodity and we see that trend coming. We call that trend application optimised.
This does not necessarily mean it will be expensive. When I say application optimised we will still use commodity parts such as the motherboard, chipset, CPU and everything else. What we do have is a different design for HPC and the datacentre.
We've already NEC turn the entire closure into the backplane.
Its a matter of scalability, a smaller backplane serving several motherboards. The concept is the same as NEC and Hitachi but we have to find a balance between the cost and design. We cannot easily get multiple motherboards in a chassis and then add cooling and power sharing.
This sounds a little like vendor lock-in
We don't see it as being locked in. They will see commodity components including things like the chipsets. The different form factors is the only thing that makes the difference. You can mix and match with other generations of hardware and general purpose servers. It might just not be fully optimised at that point in time
Procurement will be incremental. People look at the CPU roadmap and they know it will be different over time. The same thing with the form factor and the special design, people accept those difference over time. If you look at HPC today you see this today. Each generation of HPC is different from the last.
Why not look at acquiring companies such as ScaleMP who already have a solution for building HPC out of different generations of commodity hardware?
Using existing hardware is an important challenge for ScaleMP. They are good at mismatching different generations of hardware and still work fine. But for a small company they cannot accommodate all the needs of those who want HPC. What they have to do is select the most important solution for their market.
Those kinds of solutions won't be big and it will be a limited set of customers. It will also be driven by the size of the company. Will be it be mainstream? I doubt it. They have had some successful customer wins but it is about the applications.
How exactly are you going to make this work without ?
UIO is one of our technologies. It allows people to cut out the component they don't want today or they can put in the component they do want in the future.
It is a very innovative thought and design for HPC people. They can have different types of nodes this way. They could have a bunch of compute nodes that only does CPU or we could design for memory.
If using a regular server and lots of components are wasting energy, we can cut that off allowing people to assemble what they want. They can put new components in such as CPU and memory through an add-on card. This will allow people to do some flexible thing and will prolong the HPC life for the hardware.
How does this work in a channel model? You don't sell direct but a lot of dealers buying your computers may not have the skills to assemble HPCSelecting HPC is based on the customer not the dealer. In many universities all the graduate students can build a cluster with no help so why bother to have someone to help them.
The other extreme is a company with no Linux or HPC solution who needs a turnkey solution. For that we selecting specific partners who do have the experience to do this.
Professional services would allow you to engage directly with those customers who do need help but can you do that with your existing dealer channel?
Still purely on the hardware services side, we have a service group to help on this. What we don't have is an HPC or vertical market channel.
Our major business is through the channel and you don't want to upset your partners. We will always be in the assistant type of role to dealers for HPC deals. It does raise some challenging questions for us. We do have some direct customers but they have come directly to us and not through the channel.
The main reasons we would consider doing this is that the customer has a specific design and if we don't step in, they would end up being tier 1 business. These direct cases are where the customer comes to us. We do not approach them. They are not buying from the channel so there is no conflict.
There are some cases of conflict such as when the customer is not satisfied with dealer or distributor and they want somebody in between. Sometimes we have to say no and it's hard to get a perfect balance.
Can virtualisation be embedded onto the motherboard or in the BIOS?
Embedded virtualisation is best done by the USB drive on the motherboard. Putting the feature into the BIOS is a huge commitment. Maintaining a bios version is not an easy task and will be a long term commitment. The USB key is perfect.
You are not going to change things all the time and some servers will be rebooted only once per year. You don't need high speed data transfer. Putting the NAND chip on the motherboard gives a small increase in reliability but it's costly. It is about finding a balance point and you don't have to put it on the board. Flexibility is more important. There is also very little difference in cost between commodity and industrial grade USB.
Leng Tau, thank you.