Putting terabytes of memory into servers, the cheap way

Disk drives have only recently begun to be measured in terabytes. MetaRam CEO Fred Weber is talking about putting terabytes of memory into servers.

The San Jose, Calif.-based company, which is just coming out of stealth mode and has received money from three of the four founders of Sun Microsystems, has devised a chipset that effectively lets computer makers or owners double or quadruple the amount of memory in their systems fairly cheaply. A standard two-processor server on the market right now might have a ceiling of 64GB (gigabytes) of memory. By inserting DIMMs (dual inline memory modules) equipped with MetaRam's chips into the same server, the capacity can go to 128GB or 256GB.

Eight-processor servers normally capped at 256GB of memory thus can hit 512GB or 1 terabyte of memory. As memory density increases, so does the capacity allowed by MetaRam. Thus, if the limit for an eight-processor server climbs to 512MB because memory chips can hold more data, a MetaRam-enabled computer could conceivable hold 2 terabytes.

The company's chipsets are largely aimed at ameliorating the growing gap between processor capabilities and memory capacity, said Weber, who used to be the CTO at Advanced Micro Devices. Chipmakers are now coming out with chips with four and more cores. (Sun's Niagara has eight cores with eight processing threads per core.)

Performance continues to double about every 18 months or so. Memory capacity on servers, however, doubles only about every three years, Weber said. As a result, the performance of certain computers (depending on the application) is limited by memory more than processor power.

To make up the difference, server makers have to add more memory to their boxes. But in existing situations that means adding more slots for DIMMs on a motherboard, which is expensive and requires a redesign. MetaRam's chips let server makers do the same thing without adding slots. The chipset--which is shipping now--costs US$200 in 1,000-unit quantities.

A conventional four-processor 256GB server can run into the hundreds of thousands; the same server with MetaRam-enabled DIMMs might cost US$50,000, Weber said.

"I like it. It solves the basic problem of how to get a lot of memory into a finite number of slots," said Nathan Brookwood, principal analyst at Insight 64.

Potential customers include financial institutions, which run rapid, data-intense simulations, scientific researchers, and the people rendering 3D and animated movies--about 10 percent to 20 percent of the server market.

Smart Modular Technologies and South Korea's Hynix have already agreed to produce DIMMs with MetaRam's chips. Servers and workstations from Rackable Systems, Colfax International, and other manufacturers focused on high-end computing have also said they will put these DIMMs into products in the first quarter of 2008.

MetaRam's chipset serves as a conduit between the memory chips located on the DIMM and the computer's memory controller, known as the Northbridge. Typically, memory chips speak straight to the Northbridge. In MetaRam's system, signals to and from memory go through its chips first.

"You might call it (MetaRam's chipset) an intermediate memory controller," Weber said. "Instead of talking to four 1-gigabit DRAMs, we make it think it is talking to one 4-gigabit DRAM." (Chips are measured in gigabits, while DIMMS which hold chips get measured in gigabytes.)

The company also has tailored its business plans to avoid antagonizing the incumbents, he added. Memory makers like it because they can design a new DIMM to accommodate the chip, and voila, start selling a new, higher-margin product. In the past, Rambus came up with a way to improve memory performance, but it required designing a whole new memory chip. Big mistake, Weber said.

Server makers like it because it lets them add more memory to their boxes without tinkering with a new motherboard. (Because the chip comes on a prepackaged DIMM from one of their usual suppliers, server makers also don't have to deal directly with a start-up, something they are rarely fond of.)

Intel and Advanced Micro Devices likely won't try to develop this type of chip, Weber argued. The type of functionality that comes from MetaRam's chips isn't easy to integrate into conventional chipsets. As a result, the interest of the processor and chipset giants is likely greatly diminished.

"This is something that has to go onto the memory module to be useful," Insight 64's Brookwood concurred.

MetaRam comes with a fairly deep pedigree in servers. While at AMD, Weber oversaw the design and development of Opteron, the server chip that pulled AMD out of the doldrums and gave the company a rare, temporary lead over Intel in performance.

Investors in MetaRam include Andreas Bechtolsheim, chief architect at Sun Microsystems; Khosla Ventures; Intel Capital; and Kleiner Perkins Caufield & Byers. Weber pointed out that Bill Joy is the Kleiner partner participating in the deal. Thus, three of Sun's four founders--Vinod Khosla, Bechtolsheim, and Joy--are investors.

MetaRam's basic technology was devised by Surest Rajan. The company received its initial funding in January 2006, which is a fast turnaround in the semiconductor world.

"We took on a fairly simple job and we wanted to do it in a year," Weber said. "I picked a problem we could do with 20 guys."