Intel's Pat Gelsinger with the company's upcoming dual Core 2 Extreme "Skull Trail" system.
The CPUs aren’t even out yet but the war of words has already begun between Intel and AMD over whose CPU is bigger, err, better. While Intel showed off early samples of its next-gen Nehalem CPU at its annual developer conference this week, AMD was sniping at the company from the sidelines.
The big splash came from Intel which said that it would have its 45nm-based Penryn family of CPUs out this fall, the company also demonstrated its, next next-generation chip, code-named Nehalem, up and running a year before it's even scheduled to be launched.
Nehalem will be scalable from 1 to 8 cores and will bring back a form of Hyper-Threading, but the big change from previous Intel processors will be the inclusion of an integrated memory controller. The three-channel DDR3 controller will work along with a new chip-to-chip communication protocol dubbed QuickPath. Previous generations of quad-core chips, including the new Penryn processor, continue to use a shared front-side bus toplogy. For one core to communicate with another core on another die, it must use the much slower front side bus. Intel has addressed by increasing the bus to 1,600MHz for its server and workstation Penryns, but Nehalem will have a much faster connection between CPUs at it's beck and call.
That got AMD into the spin game by whispering to anyone that would listen, that they were right!
“What's amazing is that many of (Intel’s) ‘groundbreaking, innovative new technologies’ are close facsimiles of technologies AMD pioneered, is already shipping, and in some cases, has been shipping for years,” AMD crowed to the media. “For example, products that are more than a year away, like Nehalem (compare to native Quad-Core AMD Opteron), and QuickPath (compare to AMD Direct Connect Architecture and HyperTransport) are simply Intel's admission that AMD was right all along about an integrated memory controller being the key to a superior processor architecture.”
AMD first moved the memory controller from the chipset directly into the CPU die with the Opteron and Athlon 64. The chips also utilized a chip-to-chip communication protocol based on HyperTransport instead of a front-side bus topology. CPUs weren’t the only thing AMD pooh-poohed. The company also said Intel’s graphics project, code-named Larrabee (which Intel publicly confirmed this week) sounded pretty damned similar to its own Fusion project that will combine ATI graphics with AMD CPUs.
Obviously sensitive to AMD’s jibes, Intel exec Pat Gelsinger, acknowledged the similarities.
“When you look at that simple system architecture they look pretty similar,” Gelsinger responded rather testily to a media question. “On a gross level they’re very similar.”
But Intel's Gelsinger shot back, if Opteron’s on-die memory controller and chip-to-chip is so bad-ass, why does a Harptertown (Intel’s Penryn-based Xeon) using the outdated front-side bus already outperform Barcelona? Gelsinger attributed it to Intel’s superior cache design and the bus. “When we bring those big caches over to Nehalem, that superior cache architecture, wow.”
"(With Nehalem) we’re going to have a superior systems architecture, dramatically superior system bandwidth with superior caching technology,” Gelsinger booasted. “The instruction per clock of Merom is still ahead (of Opteron), and Nehalm jumps those even further. “So (Nehalem has) a superior microarchitecture, cache technology, and memory bandwidth -- This is a phenomenal platform.”
Translation: We’re going to kick AMD's ass. Gelsinger’s and AMD’s posturing aside, all bets are on Intel for this season. Penryn will make its debut on the desktop and server within the next two months while AMD’s first Barcelona chips were shipped at a disappointing 2GHz. Meanwhile, AMD’s great green hope, Phenom FX, may not debut until December. Too late to make the crucial Christmas shopping season. There’s been speculation that AMD’s problem lies in fabrication problems--with low yields being a common explanation. Matters weren’t helped when the company announced that it would release a triple-core core Phenom in 2008. While some say AMD’s tri-core plans could work because it would fill a price and demand gap between dual-core and quad-core machines, it left an opening for Intel to pee on AMD’s parade.
Is AMD doing this to fill a non-existent gap, or is AMD having problems with its native-die approach that it’s trying to cover up? AMD has made much hay of its “true quad core approach,” which puts all four execution cores on a single contiguous piece of silicon. Intel, meanwhile, still builds its quad-cores by joining two dual-core dies inside the CPU package. While AMD’s architecture is more elegant, it also can lead to lower yields as each execution core on the die must be functioning for it to be made into a CPU. Intel’s yield could be drastically better since the company only has to two find matching dual cores to make a quad core. In other words, AMD must make its quad-leaf clovers by finding perfectly formed ones while Intel can construct them using two, two-leaf clovers. Intel’s pragmatic, but architecturally inferior approach could prove to be right in the end. Intel officials say yield issues aside, it's already selling $200 quad-cores, so why would anyone want a tri-core?
Enthusiasts confidence in AMD also took a hit when rumors surfaced on the Inquirer that AMD’s much touted FASN8 dual-processor uber-enthusiast platform may have fallen under executioners axe. The company rep couldn’t confirm or deny the rumor when Maximum PC asked.
Even if FASN8 is dead, Intel says it plans to roll forward with its V8 platform. Like AMD’s FASN8, V8 is a dual-processor board offering up to 8 execution cores in a single PC. But unlike its workstation brethren, V8 would offer up to four physical and electrical x16 slots and be based on an Intel platform dubbed, Skull Trail.