San Francisco Intel Corp. will make a play for market leadership over Advanced Micro Devices Inc. in the emerging world of mainstream multicore processors at the Intel Developer Forum here this week. But the deeper, systems-level implications of a rising tide of multicore CPUs are the unfinished story of tomorrow's computers.
With the advent of multicore processors, servers are increasingly becoming nodes that support multiple operating systems in network clusters. That is driving two new requirements: the need for virtualization sharing resources across networks and operating systems and the need to ease I/O bottlenecks.
Attacking the latter problem, Intel will announce an approach to accelerating Gigabit Ethernet traffic within a server that could marginalize the efforts of a broad group of chip makers that have been working on enhanced Ethernet products. Separately, the PCI Special Interest Group has quietly kicked off a working group to set standards for virtualization of hardware resources.
Intel will roll out its I/O Acceleration Technology starting in 2006 with modified Xeon processors, chip sets and Gbit Ethernet motherboard chips. The company promises a 30 percent reduction in the time it takes to deliver data that comes into a networked server.
The technology essentially pipelines the processing of network data packets on a CPU while waiting for system I/O memory copies without inserting any new data headers in the TCP/IP traffic. That's an alternative to the remote direct-memory-access technique being developed by a broad group of server and Ethernet chip makers under the RDMA Consortium.
"RDMA requires modifying the software. We don't use RDMA, so we don't have to change applications," said Stephen Chenoweth, a marketing manager in Intel's digital enterprise group. The approach targets all Web- and midtier-server computing work, leaving only a 10 percent niche, in database and backup applications, for the enhanced Ethernet techniques in the works from RDMA Consortium members and others developing chips using TCP offload engines.
Uri Elzur, director of advanced technology for Broadcom Corp.'s enterprise computing group, said the Intel approach is expensive and wasteful. Intel will dedicate a core in a multicore Xeon to network processing, he said, but the approach will not be able to keep up with the volume of TCP acknowledgements.
Broadcom has demonstrated Gigabit Ethernet with TOE and RDMA that delivers more than 10 times the 30 percent performance increase Intel is claiming, while using less than 10 percent of the capacity of the host CPU. But Elzur would not say when Broadcom plans to deliver 10-Gbit Ethernet products.
Many network efforts aim to enable virtualization, or the sharing of a computing function across a network of servers and operating systems. "Easing the I/O bottleneck will help people stay on the path to virtualization," said Intel's Chenoweth.
Virtualization is considered so key that 19 companies helped kick off an effort to set standards for resource sharing over PCI in January.
The rise of multicore processors is providing the impetus for the virtualization work, said Renato Recio, a senior I/O technologist at IBM who co-chairs the new PCI SIG working group.
Intel claims to have more than 10 multicore designs now in the works. Intel has said it will ship two multicore desktop processors by June, including a dual-core version of its Pentium Extreme Edition for PC gamers.
Intel is playing catch-up with AMD, which announced in 1999 that it had designed its Opteron and Athlon CPUs with dual-core designs in mind. AMD plans a midyear launch for three versions of dual-core Opteron chips.