Slashdot Mirror
Intel Details Silvermont Microarchitecture For Next-Gen Atoms
crookedvulture writes "Since their debut five years ago, Intel's low-power Atom microprocessors have relied on the same basic CPU core. That changes with the next generation, which will employ an all-new Silvermont microarchitecture built using a customized version of Intel's tri-gate, 22-nm fabrication process. Silvermont ditches the in-order design of previous Atoms in favor of an out-of-order approach based on a dual-core module equipped with 1MB of shared L2 cache. The design boasts improved power sharing between the CPU and integrated graphics, allowing the CPU cores to scale up to higher speeds depending on system load and platform thermals. Individual cores can be shut down completely to provide additional clock headroom or to conserve power. Intel claims Silvermont doubles the single-threaded performance of its Saltwell predecessor at the same power level, and that dual-core variants have lower peak power draw and higher performance than quad-core ARM SoCs. Silvermont also marks the Atom's adoption of the 'tick-tock' update cadence that guides the development of Intel's Core processors. The successor to Silvermont will be built on 14-nm process tech, and an updated microarchitecture is due after that."
Intel has paid sites like Slashdot to proclaim every Atom chip to be the greatest thing since sliced bread, and each Atom core has actually managed to be worse (in its sector) than the previous design. Remember, Intel is the company that spent more than the combined R+D budget of ATI and Nvidia combined (across their entire history) to produce the Larrabee GPU, the worst graphics chip ever designed, and the greatest semi-conductor flop in the history of forever.
Intel has exactly one success to its name- producing a moderately faster desktop x86 CPU core than AMD by outspending AMD by hundreds to one. Even in this incredibly limited field, AMD has actually bested Intel TWICE in x86 history. Intel is a dreadful, dreadful company, successful only because of its de facto (effective) monopoly in the x86 business.
Intel's current and coming mobile CPU cores are slaughtered in every important metric by AMDs astonishing new Jaguar cores, but it gets worse. Jaguar cores are simply part of the best x86 SoC (system on a chip) part ever produced, including the very best integrated GPU, and a massive step toward true HSA computing (Intel hasn't even begun to change its architecture to support any aspect of HSA).
Jaguar supports every current x86 instruction option. Intel ONLY supports the most recent instruction set improvements on its most expensive desktop parts. As you drop lower down Intel's line, they remove core features, ending in the utterly stripped down and crippled Atom line.
It gets worse. Intel's appalling GPU design (Intel was too thick to write drivers for the PowerVR licence it previously used- the same PowerVR architecture that powers all Apple mobile devices so successfully) stinks for any but the most basic use. Intel's integrated GPU cannot provide smooth gameplay, having massive variable frame latencies that always make rendering feel 'stuttery' (like the problem AMD has when using two ATI graphics cards together, only far far worse). Intel's GPU drivers are a compete horror, leaving most games full of visual glitches. Only games that get special attention from Intel's graphics teams have any chance or rendering acceptably of an Intel GPU.
It gets worse. Look at the coming Sony PS4- entirely designed using AMD's technology from 2014 (yes, that date is correct) desktop x86 parts. The single AMD chip that contains the EIGHT CPU cores, and the massively powerful GPU, moves almost everything the CPU might be used to process onto dedicated hardware blocks, or the GPU. Intel's CPUs are optimised around the concept of massively inefficient PC designs running a horrible Microsoft OS. Intel has a notable desktop lead over AMD because it handles these inefficiencies (like the massive over-heads when the scheduler switches active threads) better than AMD.
The PS4 architecture shows the future. A design that can finally run thousands of threads at the same time without over-heads that would use more than 100% of available CPU performance if attempted on the traditional desktop. Thread scheduling ARM based computers on the GPU itself to allow efficient COMPUTE threads to replace mega-slow FPU calculations on the CPU.
AMD and ARM are planning for a linear relationship between the transistors used, and performance. Intel's model is to double the transistor count for maybe 7% improvement in (traditional) CPU performance. It doesn't take a genius to figure out how this will work out for Intel.
Up to now, Intel has paid billions to software developers and Microsoft to focus on single-threaded performance. Such an approach has driven PC improvements into the ground, but has been incredibly successful in maintaining Intel's market share and profit margins. Behind the scenes, programmers were learning how to create 'work unit' based computing to drag every last ounce of computing from the ageing multi-core Xbox360 and PS3. This expertise, that perfectly scales to all multi-core designs, is the motivation behind the designs of both the PS4 and Xbox720, and all coming x86 parts from A