AMD Wants To Help Partners Cut High Memory Costs With Paid Software
The push by AMD to lure customers with the promise that it can meaningfully reduce DRAM costs is coming more than a year after the AI data center boom created a global memory shortage that is causing prices of such components to rapidly increase.
AMD plans to tap into channel partners to help data center customers cut down on exorbitant memory costs with paid proprietary software that can make cheap flash storage drives appear as DRAM to a system—without a major performance hit.
The plan, laid out during an AMD press event earlier this month, stems from the chip designer’s June acquisition of a Silicon Valley startup called Mext that has developed a software solution it calls “AI-powered predictive memory.”
[Related: Nvidia Partners Tell Customers ‘Not To Fear’ In Wake Of Jensen Huang’s Exit]
The push by the chip designer to lure customers with the promise that it can meaningfully reduce DRAM costs is coming more than a year after the AI data center boom created a global memory shortage that is causing prices of such components to rapidly increase.
Alexey Stolyar, CTO of Northbrook, Ill.-based AMD systems integration partner International Computer Concepts, told CRN last Thursday that the appeal of Mext to the channel is obvious, even when not factoring in the current memory crisis.
That’s why, according to Stolyar, his company was working with Mext to sell its software well before it was acquired by AMD to serve as its primary hardware partner. International Computer Concepts even has a few customers using it now, he added.
“This was one of our more innovative ways [to help customers] besides just going out and buying RAM,” the executive said.
However, Stolyar said, the acquisition “disrupted the flow” of how Mext worked with his company, and he’s now looking to get the relationship back on track.
“We’re figuring out how do we go re-engage again, so we’re having those discussions right now,” the executive said.
AMD Claims Significant DRAM Cost Reductions With Mext
Santa Clara, Calif.-based AMD is claiming that Mext can significantly reduce the amount of DRAM required for data center applications—anywhere from 20 percent to 75 percent, or more in some cases—in exchange for a relatively small degradation to application performance.
For example, the company found that the solution can enable an electronic design automation workload to run with half of the system’s original DRAM capacity at 95 percent of the original performance. It also found similar tradeoffs for agentic AI workloads, with a memory-hungry agent running at 90 percent performance with half the original DRAM kit.
“So instead of saying buying 2 TB of DRAM for your EPYC-based server, you could buy 1 TB of DRAM,” said Gary Smerdon, Mext’s CEO and co-founder who is now corporate vice president of business development at AMD.
This translates into significant cost savings for data centers, he said, pointing to how AMD has already managed to save an estimated $4.9 million in avoided DRAM costs on an initial deployment of 145 servers that will eventually expand to 2,000 next year.
“The overall cost saving per system is 83 percent. Huge cost savings,” he said.
While AMD is pitching Mext as a way for data center operators to cut down on how much DRAM they need to reach a certain performance level, Smerdon said the software can also help customers improve an application’s performance.
“In this case, we may take, for example, a TB of DRAM, and make it look like 1.2 or 1.5 TB, or 2 or 3 TB, or even 4 TB of memory in that system, all through the expansion with our software plus the NVMe drives,” he said.
How Mext Reduces DRAM Requirements: AI And Flash Storage
Mext enables data centers to cut down on DRAM costs by offloading cold memory to flash storage and using an “ensemble” of AI models to predict when those memory pages will likely be needed in DRAM, according to Smerdon.
This effectively makes all relevant memory pages for an application appear as available in DRAM, allowing systems to make up for less memory capacity with flash storage drives— namely NVMe drives—that are orders of magnitude cheaper than DRAM kit, he added.
“That’s where we realized we could use AI to monitor the system and proactively push those pages back into DRAM, so that as far as the application or the operating system knew, they were there the whole time,” Smerdon said. “Another way of putting it is that we transparently make flash appear as DRAM.”
AMD Plans To Sell Mext Through Channel Partners, OEMs
Smerdon, who previously worked at AMD for 15 years earlier in his career, told CRN that Mext will have a “paid license component” and can only run on the chip designer’s EPYC server CPUs, though he hinted it could extend to other AMD CPUs over time.
“We’re working with our channel partners to bring it out,” said the executive, who is no stranger to the channel through previous leadership roles at companies like TidalScale, LSI and Fusion-io, which are now part of HPE, Broadcom and SanDisk, respectively.
Smerdon said AMD is working with HPE as part of a multi-phase engagement that will initially focus on highlighting “Mext-ready” systems that come equipped with high-core count EPYC CPUs and dedicated NVMe storage drives or bays.
The engagement will eventually expand to allow the OEM to bundle Mext software licenses with these systems, which will come pre-installed with Mext’s RAM View observability tool.
Smerdon did not elaborate on when AMD will make Mext commercially available or Mext’s licensing costs, and AMD declined to make him available for follow-up questions.
John Carter, vice president of compute at HPE, told CRN that HPE is a lead go-to-market partner for Mext and said the company is currently qualifying Mext to run on its “lead” Gen13 server products that use AMD’s EPYC “Venice” SP7 CPUs.
“We’ve done some of the basic testing to understand what works, what doesn’t work, and then I think it’s how you optimize the performance there and really pick the right workloads,” he said. “They’ve done some great work in targeting a few key lead workloads, which align really well with a couple of the systems we just announced.”
Asked about HPE’s plan to sell Mext and Mext-ready systems through channel partners, Carter said his company is still working out those details. He added that HPE may have more information to share at the HPE Discover event in Barcelona, Spain, next month.
“It’s going to be a perfect synergy, bringing together that Mext software and those AMD platforms to get some cool new things to happen,” he said.
Whether On-Prem Or In The Cloud, Mext Needs Lots Of CPU Cores
Mext, which runs on Linux, will give partners a variety of deployment options: on premises, in the cloud, in virtual machines or containers or on bare metal, according to Smerdon. The only requirement is that the underlying servers use AMD’s EPYC CPUs.
Because it isn’t practical to pre-train Mext’s AI models on memory access patterns, Smerdon said the software is set up to train these models on a customer’s memory usage continuously in the background. But instead of requiring a GPU to do the training, the software accomplishes this by using a single core on an EPYC CPU.
“We train in under one second, and we begin making very good predictions,” he said. “It’ll get better a little bit as time goes on, but as the application maybe goes through a phase change, it’ll retrain and adjust along with it.”
Smerdon said customers will need “plenty of cores” because “as you expand your memory over time, you want to be able to have the additional cores and the needed cores to take advantage of that memory, keeping the ratios intact with where they should be.”
Partner Sees One Limitation With AMD’s Mext Push
Stolyar, the executive at International Computer Concepts, said he has seen one limitation when pitching Mext to customers who want to use it for existing data centers: the lack of space for adding new NVMe drives.
“We presented it to a few customers that are interested in it, essentially [for] the right use case, but they didn’t have any more room in their systems for NVMe drives, so that kind of slowed it down,” the executive said.
When it comes to new data center deployments, however, Stolyar said, “there’s more flexibility, and it’s definitely interesting.”