Intel’s 10nm Ice Lake CPUs Achieves 18% IPC Improvement Over Skylake


Intel has dropped a ton of information on its upcoming 10nm Ice Lake processors and it is clear that this is going to be one of the biggest updates to its cadence so far - the first major change since 2015 by its own admission. Ice Lake, or more specifically Sunnycove, is going to be a major upgrade path for consumers that have been on 14nm for a while now and is going to be the first "tick" on the 10nm process. There had been rumors about the company's 10nm process getting delayed (or even cancelled) but all of those fears seem to have been mostly unfounded.

Intel's 10th generation Ice Lake processors detailed in Computex 2019 disclosure: 18% IPC, AI performance, WIFI6, Gen 11 graphics and more

Intel's Ice Lake processors are going to be the first 10nm commercial processors shipping in volume from the company and the second (hopefully successful) attempt at 10nm after Cannonlake. The details and technical information the company disclosed at Computex 2019 is very impressive to say the least as this major architectural improvement achieves 18% IPC improvement over Skylake and almost 40% in terms of raw performance. This is going to be a major upgrade and could trigger a new cycle of upgrades for Intel's customers if things go well for the company.

The biggest feature of Ice Lake is of course the fact that it is manufactured on the 10nm process but apart form that, the company is telling us about significantly upgraded capabilities to handle AI and machine learning-related workloads which could prove to be a selling point for certain customers. According to the slides, the Ice Lake processors achieve an almost 2.5x speed up in AI workloads as compared to Skylake thanks to DL Boost.

Thunderbolt 3 is integrated and 4K 60 with full HDR10 support is included. Gen 11 graphics mark the first serious iGPU inclusion in Intel chips, you can read our full coverage of that over here, and 1080p gaming (in the casual sense) is now viable. HEVC (up to 8K) encoding is also supported and WIFI 6 has been added as well. All in all, this looks like a great successor to Skylake and one that has been 4 years in the making.

Ice Lake features a new converged chassis fabric design which increases bandwidth and reduces while keeping the cores scalable. Remember the Sunnycove disclosures a while back? this is the same thing. Display pipelines for the graphics have been increased and now support up to 5K 60 or 4K 120.

As far as the raw specs go, the preliminary lineup of ICL (Ice Lake) CPUs are going to be 9W, 15W and 28W designs for the mobility market (as per usual) and will feature clock rates up to 4.1 GHz boost. Cores will max out at 4 with hyper threading enabled (Ice Lake has hardware level mitigation for the spectre family of attacks). The lineup will be broken down into the familiar Core i3, i5 and i7 segments and graphics will be clocked at 1.1 GHz. DDR 3200 is now supported by default without the need for overclocking of the memory controller. L2 cache size is up to 8MB.

Interestingly, the message Intel is conveying is one of controlled optimism. Having signaled the end of process-focused ideology last year, the company refers to 10nm as 'new transistor technology' instead of making it the star of the show which is good news, under promising and over delivering is always the right way to go. Ice Lake will also feature new platform integration using a brand new PCH - something we will talk about in detail.

The new Ice Lake PCH is going to be manufactured on the 14nm process and will house the integrated WIFI 6 controller that is supposed to provide a 3x speed up over the last generation. The ICL PCH features a new FIVR (fully integrated voltage regulator) setup that should ensure better power delivery. It has a standard IO set of 6x USB 3.1 (or 10x USB 2) x16 Gen3 PCIe lanes and 3x SATA 6 Gbps.

The packaging is broken into two main categories: 15W and 9W. The 15W Type 3 package will have 1526 balls with a pitch of 0.65mm while as the 9W Type4 package will have 1377 balls across a pitch of 0.43 mm. To the uninitiated, the balls in a BGA packaging are the same thing as pins in the standard LGA socket. Both types feature a new thin magnetic inductor array.

Intel Ice Lake CPU (Sunnycove uArch) performance numbers and benchmarks

Intel also provided some performance numbers for Sunnycove, Ice Lake's first micro-architecture, and they do not disappoint. Most of the preliminary information is something we have already covered in the Sunnycove disclosure that Intel made a few months back. This includes the Deeper, Wider and Smarter strategy of designing the processor and among other things, includes larger cache sizes, wider allocation and execution ports as well as hardware level mitigations for speculation-based attacks (Spectre etc).

Some new details, however, have also been given this time and these are of the juicy, technical kind. The L1 cache size has been improved from 32KB to 48KB and the L2 cache size has been doubled to 512 KB. The op cache has been increased to 2.25K vs 1.5K in Skylake and the out-of-order window has also been increased to 352 from 224. In-flight loads are now 128 versus 72 on Skylake while in-flight stores are 72, up from 56. New instructions for crypto performance have been added as well as vector capabilities that include DL boost. All of this makes for a much more snappy processor.

As we mentioned earlier, Intel is boasting about the AI performance of the new cores and they do not disappoint. In Resnet 50, one of the standard workloads for AI inferencing, the Ice Lake CPU with Sunnycove cores scores 2.5 times more than a Skylake processor. a 2.5x speedup could be invaluable to someone on the go. And finally we have the Sunnycove IPC performance benchmark: an 18% markup over Skylake and almost a 1.4x speedup in raw terms. Single threaded performance scores are also given below.