How fast is fast?

How fast is fast?

By Richard Buckle
Chief Storyteller, Pyalla Technologies, LLC.

Whenever IT folks get together, one topic of discussion seems to center on performance. What were once called the speeds and feeds – what the physical computing system could deliver in terms of turn-around. This was back in the days of batch processing. Could a batch process complete overnight? Perhaps as little as a couple of hours? When I first started my professional career in IT we would come into the office of a morning full of anticipation that the results of our endeavor had proved fruitful. Often this was measure in inches, indeed feet, of stacked printer output.

However, once support for online processing became economically viable – yes, the cost of those early CRT terminals was prohibitive leaving access in the hands of senior management – our problem of speed took on a completely different meaning. How fast were the comms lines? How big a pipe could we afford? 1200bps? 2400? It would be many years before such discussion moved to consider 4800 or even 9600 bps! Surprise? Along came Andrew Corporation (pioneered VSAT for data communications) with their low-profile satellite receivers with uplink capabilities (as I recall) and suddenly, every business had the potential for previously unimaginable bandwidth. Speeds? “Well, if there had been a speed limit, we have now paid the fine and charged on regardless,” was how one IT manager I knew reacted.

Internally, online programming proved to be one of the biggest boosts to productivity for applications development and maintenance. Even the older IBM 360 mainframes were not immune to having online terminals connected for their development teams and by the late 1970s, their presence began to proliferate. When I first saw a Tandem Computer (it was at the original Hannover Fair in the early 1980s) it wasn’t just the sight of multiple processor cabinets that surprised me, but rather, how online terminals were being used in support of operations as well as development. It would be several more years before online terminals were universally deployed on all manner of systems.

However, as much as I like to reminisce about the past, when today we talk of speeds and feeds, it is for an entirely different reason. Processors, particularly since the arrival of GPUs, perform with insane speed so fast that the mind simply boggles at the levels of performance that they provide. For the Nonstop community, living in a world of Hybrid IT where the presence of adjacent systems is becoming the norm, the ability to move data between systems has once again become a gating factor. How fast are the connections? But is the speed of the connection getting to a level where it outstrips the performance of the connected processors?

The arrival of AI is challenging many of our deeply-entrenched ideas of how to assemble a modern computer system. We thought InfiniBand (IB) was exciting and then we heard about RDMA over Converged Ethernet (RoCE) and were skeptical at first. Really? Ethernet? But then more recently the talk is about Ethernet links operating at ultra-high speed. As fabrics? Surely not, but then again, can you really stand aside from a solution anchored in open, industry-standard technologies? What of AI and how are the neural networks wired?

“The communications fabric in support of AI neural networks is the ultra-high-speed networking infrastructure that connects thousands of processors (GPUs/XPUs), memory pools, and storage. It is the critical “highway” that allows massive distributed AI models (like Large Language Models) to train and run without starving the processors for data,” as seen on the YouTube post Network Fabrics for AI Workloads. Now we are running speeds that essentially are at lightspeed and so we can say that the fine for exceeding former network speeds has been paid.

Talk of networks and workloads came to a head when at HPE Discover 2026 (Las Vegas), HPE CEO Antonio Neri declared that “architecting for AI starts with the network, positioning the company as an AI networking leader.” While this was in part a promotion of the potential positive outcomes following the integration of Juniper Networks it nevertheless highlights the importance of networking for HPE. As it now stands, according to Neri’s promotion in support of his keynote at HPE Discover, The Network as the AI Bottleneck, he stressed that “while GPUs get most of the attention, traditional networks lack the capacity for massive AI training and inference models.” It was then reported that Neri described networking as the critical defining constraint for the future of enterprise IT. Yes, speed matters and when it comes to AI, there should be no barriers limiting future speeds.

However, there’s one more critical factor to consider when thinking of speeds and feeds. It’s less so about feeds these days as it has more to do with latency. If indeed, as has been championed of late, we launch data centers into space, there will be the likelihood that latency will play a role and perhaps restrict the performance from space to applications that are less about learning and inference and more about car parts and financial ledgers. Or, maybe not. Fast will always be fast but, for the Nonstop community concerned about ensuring data created on Nonstop is ingested into Large Language Models (and avoid unnecessary hallucinations due to incomplete data), such improved speeds with lower latency will become important considerations.

Data that is meaningful arriving late is not really meaningful data but rather, just noise to be sidelined. Inference becomes incomplete and models miss out applying correcting weights to neural connections. Ouch … there goes my order of dental floss with Amazon! Seriously, the rate at which AI is advancing and impacting Nonstop, speed alone may not mean much until we begin to consider what AI brings with it, in terms of speed as applied to other parts of Nonstop that are beneficial for all who work with Nonstop.

As Keith Moore notes in his Op Ed included in this publication, Agility is Stability, the Nonstop community can leverage AI in several ways, including basic agility in terms of speed-to-test, speed-to-deploy, speed-to-secure, speed-to-change. And I suspect a lot more. Those days when we thought of line speeds of 4800 / 9600 bps came with limitations including cost, distance, and reliability / dependability, we had no idea where the technology would take us. With what we are seeing of late, driven by advances of AI, then perhaps when we come into the office or simply fire up our work laptop maybe that excitement we all shared in the past when something worked will return and fuel even more advances where Nonstop is an active participant.

Suddenly, where I feel the need for speed may not be as outrageous a request as Hollywood would suggest, but rather, be all the incentive we need to plug Nonstop into the ultra-high-speed networks forging a presence in our daily IT lives! Yes, quoting another from Hollywood, let’s “Make it so!” For certainly, as we all have come to understand from working with Nonstop, it should be just as easy to add, “This is the way!”