When Intel launched Broadwell-Y in November, branded Core M by that point, they had a 4.5W processor that was just a little slower than a 15W Haswell Ultrabook CPU. This is quite a bit of power efficiency, although these numbers are maximum draw and might not be exactly proportional to average power under load.
At CES, Intel has launched Broadwell-U, which takes this efficiency and scales it up to 15W and 28W SKUs. The idea is that the extra thermal headroom will scale up for extra CPU and GPU performance. These are all BGA-attached components, which means that these processors need to be physically soldered to the motherboards — they are destined for OEMs.
As an example of Broadwell-U's increased performance, the Core M 5Y70 has a base frequency of 1.1 GHz that can boost to 2.6 GHz; the top-end Broadwell-U has a base clock of 3.1 GHz and boosts to 3.4 GHz. From Core i3 up to Core i7, regardless of TDP, each of these processors are dual-core with HyperThreading (4 threads total). There is also a single Pentium and two Celeron SKUs, which are dual-core without HyperThreading (2 threads total).
Its GPU receives a large boost as well, particularly with the 28W SKUs receiving Iris Graphics 6100, although Iris Pro Graphics (6200 and 6300) do not yet make an appearance. If we had access to the number of execution units and we assumed the same instruction-per-clock count as Iris Graphics 5100, we would be able to calculate a theoretical FLOP figure, but that is information that we do not have. It would make sense if it were 48 execution units, twice Core M and consistent with the official die shot that Intel doesn't actually identify by product number. This would give it about 845 GFLOPs of performance, or about an OEM NVIDIA GeForce GTX 460 (the retail GTX 460 cards were about 4% faster than the OEM ones).
It is also within 2% of Haswell's Iris 5100 theoretical GFLOPs, albeit with a 15% drop in clock rate.
From a features standpoint, the GPU is a definite step-up. It has “Enhanced” hardware support for VP8, VP9, and h.265 (HEVC) video and 4K UltraHD output, wired or by Intel WiDi. Broadwell's iGPU was designed with DirectX 12 in mind and supports OpenCL 2.0 — leaving NVIDIA behind in that regard, since AMD added that API in last month's Omega driver.
Intel is slightly behind in OpenGL support however, claiming 4.3 compatibility while AMD is at 4.4 and NVIDIA is at 4.5. This could mean that these GPUs do not (unless a future driver changes this) support “Efficient Multiple Object Binding”, “Sparse Texture Extension”, or “Direct State Access”. Then again, they could support these features as extensions or something, because it is OpenGL and extensions are its thing, but you know — they're obviously missing some part of the spec, somewhere.
This leaves Broadwell-H and Broadwell-K, high performance BGA and socketed LGA respectively, to launch later in the year. These products will have significantly higher TDPs and stronger performance, at the expense of battery life (a non-issue for the desktop-bound -K parts) and heat.
Follow all of our coverage of the show at https://pcper.com/ces!