Sanmina execs tout Opteron, muted rise in tech spending

The successful original device manufacturing (ODM) outsourcing model that has reduced prices for PCs and low-end servers is poised to make its way up into the data center, and Advanced Micro Devices Inc.'s (AMD's) Opteron chip will come along for the ride, said executives from Sanmina-SCI Corp. in an interview Thursday.

PC vendors have been outsourcing manufacturing and some design functions to Taiwanese companies for years, and that production model has enormous cost-cutting potential for vendors of large enterprise servers, said Don Johnson, executive vice president for enterprise computing and storage systems at Sanmina-SCI, based in San Jose, California.

Sanmina-SCI recently acquired Newisys Inc., a server design company, to add expertise it needs to strengthen its position as a manufacturer and designer of various types of hardware. Newisys rose to prominence as one of the primary server design companies to work with AMD's new Opteron server processor, but its team of designers have experience with RISC (reduced instruction set computing) and Intel Corp. servers, and will be used on all types of enterprise servers that Sanmina-SCI can offer its customers, Johnson said.

Sanmina-SCI purchased the company in July for an undisclosed amount, and Newisys will operate as a wholly-owned subsidiary of Sanmina-SCI.

For Newisys, the deal allows them to place a billion dollar company behind their Opteron servers, alleviating the concerns of potential buyers who might have been scared off by the company's small size, said Phil Hester, chief executive officer of Newisys. And Sanmina-SCI will be able to offer high-quality Opteron servers to large vendors that want to support the chip, Johnson said.

Servers based on Opteron will eventually make their way into the heart of the data center, although that will probably take two to three years, Hester said. Companies will adopt one-way and two-way Opteron servers for applications like Web hosting at first, and move the processor into larger servers with heavier workloads as independent software vendors develop more applications for Opteron's 64-bit capabilities, he said.

In order for this to happen, AMD will need to improve the scalability of the Opteron beyond eight processors, Hester said. Sanmina could build an eight-way Opteron server, but it doesn't scale very well beyond that size due to the current cache design of the Opteron, he said.

Opteron's 1M-byte Level 2 cache size is low compared to that of chips used in large enterprise servers, such as Intel's Xeon MP or Itanium. Those chips feature up to 2M bytes of Level 2 cache on the Xeon MP, and up to 6M bytes of Level 2 cache on the newest Madison Itanium 2 processors. Cache allows the processor to store frequently accessed data close to the processor, reducing the amount of time needed to process an instruction compared to the time it would take to retrieve that data from the main memory.

One-way, two-way and four-way servers don't require as much cache, said Nathan Brookwood, principal analyst for Insight 64 in Saratoga, California. Because AMD used a high-speed memory interface based on the Hypertransport standard, the chip's overall performance is competitive with Intel's Xeon DP processor, he said.

Eventually, AMD will have to address the cache issue by increasing the size of the cache, as well as the method by which the processor manages cache coherency, Brookwood said.