Intel debuts 'Haswell' chippery: from tablets to servers
One core to rule them all
IDF 2012 Intel has unveiled details of its new "Haswell" microarchitecture, and promises that it will deliver greatly improved compute and graphics performance, drastically lower power requirements, and developer-friendly improvements when chips based on it appear next year, branded as Intel 4th Generation Core Processors.
"The great thing about this one is that it was designed with mobility in mind," said David Perlmutter, the general manager of Intel's Architecture Group, during his opening keynote at the Intel Developer Forum (IDF) on Tuesday in San Francisco.
That may have been one guiding emphasis of Haswell's development, but during the deep-tech sessions that followed Perlmutter's keynote, a series of Intel techies delved much deeper into Haswell's innards, and revealed details that show the new architecture spreading throughout Intel's product line into products ranging from tablets to servers.
"Haswell will be the foundation of a family of products, just like 'Sandy Bridge' and 'Ivy Bridge' were," said Ronak Singhal, Intel Senior Principal Engineer," at one of those sessions.
"The first thing to keep in mind about Haswell is that it is a converged core," he said. "Converged core means that we scale products that go all the way from high-end servers all the way down to low-power form factors like tablets, using the same core."
The same core means that developers can apply the same techniques to each iteration of the Haswell chip, since at their most basic levels their arcitectures are in common. "The important thing from a developer's perspective," Singhal said, "is that since we have a converged architecture across all of these segments, you can develop and optimize once, and then scale back across these other products."
Intel Fellow Per Hammarlund agreed. "You might scratch your head and think we're nuts, trying to have the same core from tablets to servers," he told the assembled übergeeks. "But in reality, if you think about it, the power levels that you have with just a few cores in a tablet at a very low power are very simlar to the power and power-efficiency requirements you have in tens of cores in a high-end server."
Of course, as important as developers are to Intel, the company also knows that a converged core architecture adds to their own bottom line. "It's actually good for us from a stingy point of view," said Hammarlund, referring to the advantages of the modularity inherent in a converged-core architecture.
Knob-twisting for fun and profit
Singhal noted that Haswell's modularity makes it easier and thus cheaper for Intel to span the new microarchitecture over such a broad range of platforms. When discussing the "knobs" of modularity – those things that can be changed without changing the microarchitecture itself – he first mentioned the number of cores, which will be 2 to 4 in the first client-level Haswells due next year.
Client-based chips were all he wanted to be specific about on Tuesday concerning core counts. "We're not going to talk today about the server-based products," he said, "but on the server products we obviously are going to have products that go to much greater than four cores."
Exactly how much greater is "much greater" we'll have to leave for another day.
Graphics are another knob. A high-performance desktop machine designed for high-frame-rate 3D gaming or multiple-stream video encoding and decoding will have higher graphics needs than will a Haswell chip churning away in some data center, serving web pages. Haswell's graphics will be available in three levels: your basic GT1, mainstream GT2, and double-your-pleasure, (almost) double-your-transistors GT3.
Other modularity knobs include cache sizes and power-level choices – both idle and active – that will determine what interconnects a Haswell chips has or what different types of platforms, such as traditional or power-optimized, to which a specific Haswell chip might be targeted.
"Modularity gives us the ability to build products that are more finely tuned to certain form factors," Hammarlund said. "That agility is great for us, and it's also good for the form-factor innovation that it helps drive."
The IDF Haswell sessions went into great detail about the compute cores themselves and the different modular graphics cores that the chips will choose among, plus their improved power-management capabilities, upgraded media capabilities, new developer-friendly capabilities, and the like.
If all that Intel promised during Tuesday's sessions comes to pass, Haswell chips should not only provide better performance than its Ivy Bridge forebears, but do so at power levels which point to future chips forcing Intel's Atom line from its client-device home, and making the li'l fellows welcome only in microservers.
The Reg promises a deep, detailed dive into all of Haswell's capabilities and architectural goodness in a few days, after IDF draws to a close and we can sit back and grok all of the Haswell microarchitecture's complexities, advantages, and advances.
Stay tuned. ®