Design Choices have Consequences

I spent 2012-2018 at Plexxi Networks, a software defined network company.??Before Plexxi, I had worked at four other network startups and two public companies (Nortel and Ciena).??Networking is usually viewed by non-networking people to be the complicated world of the dark arts.??I had never sold storage or servers until I arrived at HPE.??

A few days ago, I was talking to a SI as we were looking to get a quote on some servers for the Drut lab.??When we got on a Zoom, the questions started.??How much memory, what kind of GPUs, AMD or Intel, what kind of storage, any SmartNics, etc.???It trigged a thought from my first year at HPE.

We did some amazing and complex network designs at Plexxi and some of these designs took a day or so to design and optimize.??When HPE acquired Plexxi in 2018, I knew it was going to be a struggle to turn compute salespeople into networking salespeople.??What I was unprepared for was the complexity of the compute design cycle.??Over on the network team we could easily design networking for racks of gear in short order.??Just tell us the NIC speeds, uplink speeds, fabric OS, routing information; it was kind of straight forward.??Then our network teams started working with the compute sales teams and the whole design process became complicated.??

The first lesson I learned was the server was configured delicately to meet the needs of the customer and it was not really the CPU generation that was the tricky part.??Some servers had many GPUs, some had more storage.??The tricky part was trying to fit in all the customer demands around memory, GPUs, storage, IPUs, NPUs, TSUs into the server chassis.??If you want a lot of one resource, say GPUs, that will preclude other resources.??Design choices have consequences.?

If you take a step back and look at all the innovative cycles going on in server space, it is not really in the server, but rather in the resources that go into the server.??The semiconductor industry calls this “disaggregation.”??I was speaking to a CEO of optical component company, and he told me that he was reluctant to do a full system startup again.??We both worked for a time at Nortel in 2000.??As CEO of an optical component start-up, he sees the interesting innovation focused on various components across the infrastructure spectrum: optics, GPUs, memory, IPUs, TPUs, Storage, FPGAs.??All the things you can put into a server or a system level product to perform discrete functions.??These discrete resources are the future as they are the processing building blocks designed for the specific needs of the customer.??This was the starting point for me to think about how to remove these barriers, the consequences around design choices.???

Bill Koss