Intel Previews New Architectures for 2019 and Beyond
Intel has organized an Intel Architecture Day 2018 for a limited audience, where the company revealed its plans for the future. It shared information about a new processor architecture and developments in the GPU area.
In view of the problems surrounding the shrinking of transistors, Intel seems to be bringing forward new spearheads. Instead of a focus on clock speed or core numbers, the company primarily wants to make its processors faster in the coming years by improving the architecture. In addition, things like software, security, interconnects and memory should improve the performance of Intel’s portfolio in the coming years.
During Architecture Day, which has been extensively reported by the neighbors, Intel seemed to openly admit that the 10nm Cannon Lake generation is a flop. The company would no longer plan to bring that Core generation to the market, leaving the i3-8121U sparsely released this year alone. However, the 10nm node would not be skipped, but instead a new generation of 10nm chips would appear, with cores codenamed Sunny Cove. The processors would continue to be called Ice Lake, a name we’ve seen on roadmaps before.
The Sunny Cove architecture is improved in a number of areas compared to the current Skylake architecture, which is used in all processors up to Coffee Lake Refresh. This gives Sunny Cove a ‘deeper’ architecture, so that more instructions can be executed simultaneously. This is possible due to larger caches, better branch prediction and importantly: extra execution units. Where at Skylake there are still eight per core, for Sunny Cove it will be ten per core. Intel has also developed new instructions, including for encryption and machine learning.
The tenth generation GPU, which was disabled in the i3-8121U, would be a dead end, and Intel has indicated that it is developing the eleventh generation for integrated GPUs. Thanks to a larger L3 cache and many more execution units, from 24 to 64 pieces, it would allow performance of up to about 1.1 Tflops. That would be close to the Vega 8 GPU that is in AMD’s 2200G apu. In addition, the new GPU will receive support for variable resolution rendering, a technique Nvidia calls variable rate shading in RTX cards and coarse pixel shading at Intel. The Gen11 GPU also gets a new h265 encoder, and HDR tone mapping and adaptive sync for displays.
Finally, Intel has developed a new way of assembling chips. Instead of producing one monolithic piece of silicon, it wants to use the Foveros technology. With this, different chiplets can be connected via an interposer, just like AMD announced for its Epyc chips. Intel demonstrated that Foveros technology with a 12 by 12 millimeter soc that combines two Core processors with two Atom cores, a dram chip and the necessary logic for interconnects and memory and other controllers. It is therefore a complete sip with two powerful and two economical cores. With that technology, Intel has to compete with Qualcomm, among other things, to make more energy-efficient laptops.