It’s not often a gaggle of tech journalists aims a disproportionate share of their questions at just one speaker, but AMD’s newest golden boy, at least in the eyes of the press gang at the EPYC event in Las Vegas, was Gary Smerdon.

He joined following AMD’s acquisition of storage software startup Mext in June. The formal announcement of that deal was fairly unexciting. No financial details and a handful of paragraphs penned by GM Dan McNamara skimpily outlined how the startup’s memory optimization technology could improve AI infrastructure performance.

It wasn’t until AMD’s In-Depth showcase, an event intended to hold aloft the importance of its processor offerings, that the gravity of what Mext represents sank in.

For those unlucky enough not to be in the bowels of the Cosmopolitan to learn more about Mext, the simple explanation as to what it entails is that it almost tricks operating systems into making flash storage function as high-speed DRAM. Through Mext's software, memory demand is predicted, with less-used memory pages swapped between DRAM and NVMe flash to expand usable memory. Even more simply, it could significantly slash costs by using a memory option that is 50 times cheaper per bit and requires 30 times less power per bit.

The trade-off with flash is that it’s considerably slower than DRAM – 500 times slower in Smerdon’s own words – but Mext’s special software sauce is designed to successfully mask that speed penalty.

The now AMD-owned software uses a proprietary AI to monitor deployments, proactively pushing those pages back into DRAM. An “ensemble” of AI models, as the exec said, runs in user space to make the predictions, sitting alongside a Linux driver that handles the memory-management integration. The workhorse powering that ensemble is what Smerdon described as a “transformer-derived” system like ChatGPT or Claude, but instead of the next line of text, it taps an application’s memory-access patterns to predict which pages are likely to be needed next. It predicts which pages on the slower NVMe flash tier will soon be needed and moves them back into DRAM ahead of time, but before the application requests them.

Without getting too in the weeds, the Mext software hides flash’s much higher latency all while extending memory capacity for applications with little or no performance impact.

And if that doesn’t tantalize you as it did the press corps, all that happens on just a single EPYC CPU core … no GPU. For those virtual machine-minded, it runs at the guest level with one core per guest.

“We train in under one second, and we then begin making very good predictions. It'll get better a little bit as time goes on, but as the application goes through a phase change, it'll retrain and adjust along with it,” Smerdon said.

Think about that. A means to cut memory costs and aid existing capacity without requiring a separate accelerator; suddenly, amid encroaching competition from the likes of Nvidia and Arm, it could well become a genuine differentiator.

That logic wasn’t lost on Smerdon, who touted Mext as having emerged during “one of the biggest gale force tailwinds of history.” Now as part of AMD, he was even more bullish, saying the software unit was on its “first step” and “what is going to be achieved over the next decade is going to be amazing.”

Doing the DRAM math

With memory woes set to extend as far as 2030 if SK Hynix is to be believed, a concept like Mext for IT buyers is a potential financial life raft, a means to boost server capacity without having to sell a kidney to afford the physical hardware.

That was the central point to Smerdon and AMD’s pitch, saving both dosh and DRAM.

“Instead of saying buying two terabytes of DRAM for your EPYC-based server, you could buy one terabyte of DRAM, and then the next terabyte come with Mext,” he highlighted, though in this instance he referred to the software paired with a dedicated NVMe flash drive.

“If you're familiar with the cost of memory, going from a 256 gigabyte DIMM [dual in-line memory module], the biggest, most expensive ones you could get, and going to a 128GB DIMM is far less than half the cost. So I reduced the cost because I had less memory, but I also had a less expensive DIMM. It's a double benefit.”

Naturally, with claims like these, reporters and analysts alike in the room were abuzz with questions in what this editor quickly dubbed “the Gary Show.” But to dish some numbers on this plate of perspective, Smerdon and the AMD team thought where better to put Mext to the test than with the processor giant’s own IT team.

A case study published quietly in the days leading up to the Vegas showcase details AMD IT deployed Mext on 86 hosts, with the software supplying approximately 58 terabutes (TB) of additional capacity. Of course, these are AMD’s own numbers, but that would equate to a purchasing equivalent of DRAM capacity at around $3.3 million – while software deployment costs were just $601,000.

If Smerdon’s presentation to the press pool is anything to go by, that deployment seems to have risen somewhat to 145 hosts and an estimated saving in avoided DRAM costs stood at almost $5 million, “and that's growing pretty much on a daily basis,” the exec said, claiming he had updated related slides twice in the lead up to the event “because the number of installations keeps going up.”

Whatever the final tally, the moving snapshot conveyed the scale of the opportunity that not only captured the imagination of its EPYC value proposition for both gathered press and AMD’s IT team, but partners too. John Carter, VP, for server product management at Hewlett Packard Enterprise (HPE) admitted the following day that it was testing Mext.

“[Mext] removes one of my biggest constraints right now, which is the memory situation. So that means that customers with the same budget can afford more. So not only can they invest in their AI, they can modernize the rest of their general purpose infrastructure at the same time and get those savings across the board,” he said during an OEM panel.

While Carter was the only executive on hand who freely acknowledged that his employers were evaluating Mext, an AMD source told SDxCentral that several other partners were quietly trialing the software behind the scenes.

And why wouldn’t they be? In a market where physical DRAM pricing is skyrocketing, a simple piece of software running on a single node capable of slashing expansion costs is an instant best friend.

Of course, the hope is that memory prices won’t stay painful forever. Asked what happens to Mext’s market once its tailwind of soaring DRAM costs return to normal, Smerdon didn’t dodge. Prices might well fall back, he allowed, but “physics isn’t going to change.” He argued that DRAM cells aren’t going to keep shrinking, and with core counts climbing while memory scaling lags, “other solutions have to come in,” with Mext one of them.

To bring the Gary Show to a close, there’s also a neatly poetic side to this story. Fresh out of college, Smerdon found himself at AMD for some 15 years working as a sales manager and later as director of marketing for its networking products. After that came a run through Marvell, LSI, where he led the SandForce flash acquisition, Fusion-io, and a string of startups, including TidalScale, which tackled in-memory performance long before memory got this expensive.

Now, having sold Mext to AMD, he’s back to where it all started but on a much, much bigger stage. His official AMD title is CVP of business development. Probed by SDxCentral as to what such a remit actually entails, the man of the hour simply said he was reporting in to McNamara on the EPYC team, and that he’ll be primarily running the engineering organization while also being “very involved” across the organization and the go-to-market.

“I think many people get that this is such a big deal and a big opportunity. I'm really looking forward to focus on and making this a huge success,” the exec added.

Not a bad way to come home.