Originally Posted by whatever00
Does that mean Apple will be forced to offer discrete GPUs on all Nehalem-based models to promote the adoption of OpenCL given that GMA is too slow and the fact that third party IGPs are not viable? That's good news then.
Well that is the hope since DMI is to low-bandwidth to hang an IGP off of and QPI equipped CPUs aren't the right market for IGPs.
The other option is to do like Sideport in the ATI 780G, 790G, and 790GX. Basically have an IGP with a small amount of dedicated memory, say 128MB, with the rest shared with system memory. Of course, by the time you start giving IGPs dedicated memory, you probably should consider whether you might as well go with a discrete GPU to avoid thermal and transistor budget limitations of sticking the IGP with the rest of the chipset.
Originally Posted by nvidia2008
As someone pointed out above, Apple may be Intel's bitch... Well, just like Apple had a long running exit strategy with compiling for x86 (codename Marklar or something)... Apple had better keep in touch with AMD, ATI and PASemi etc. to avoid being painting themselves into a corner with Intel and/or Nvidia.
Well I don't view the decision to integrate the northbridge on the CPU as particularly malicious since it makes sense from a performance perspective. Using only low-bandwidth DMI-links for most chips also makes sense since high-bandwidth QPI is mainly needed for multi-socket setups. In terms of licensing, Intel is probably just holding out hoping to get concessions from nVidia. Afterall, nVidia has agreed to license SLI now which is important since PCIe controllers are going to be integrated on the CPU. If the bus license agreement said FSB, then nVidia will be in the wrong and will need to negotiate a new agreement for QPI and/or DMI.
I don't doubt that Apple will keep their options open. Realistically, ATI is a bigger supporter of OpenCL, since they've said they will support standards like OpenCL and DX11 Compute Shaders over proprietary GPGPU frameworks. nVidia however, is going to continue to promote their proprietary C for CUDA language alongside OpenCL and DX11 CS.
It should be noted too that while Apple may have a lot riding on Intel, Intel has a lot riding on Apple too. With Dell now selling AMD CPUs, Apple is actually the lone major OEM that is Intel exclusive for computer CPUs. With Apple's marketing power, Intel losing Apple as a customer would be a critical blow to Intel's image even if it doesn't directly effect their bottom-line. It's in Intel's best interest to keep Apple if not happy, then at least satisfied.
Originally Posted by nvidia2008
Do we know if those Clarksdale and Arrandale on-chip GPUs come even close to a 9400M level of performance?.
The GMA X3100 was produced on a 90nm process and the GMA X4500 was produced on a 65nm process. The desktop GMA X3100 was clocked at 667MHz with 8 stream processors and the GMA X4500 was clocked at 800MHz with 10 stream processors with some internal redesign and optimization. Usually, a shrink from 90nm to 65nm would allow the doubling of stream processors, so of that was given up for higher clock speed, but Intel was definitely conservative with the GMA X4500's design. The GMA X4500 is still about twice faster than the GMA X3100.
The IGP on Clarksdale and Arrandale will be built on a 45nm process. I think the most likely configuration will be 16 stream processors clocked at 1066MHz (8 times Nehalem's CPU bus of 133MHz). That's probably conservative and achievable, which is good since this is the first time an IGP has been put on a CPU even if it is on package as a separate die from the CPU. This configuration won't beat the 9400M but will come close, especially if Intel is serious about their drivers which you'd think they will be in the lead-up to Larrabee. Being close to the CPU, and the raw power of Nehalem will only help performance. Intel was also on the working group for OpenCL and made a whole bunch of flowery talk about it in press releases so it's very likely future IGP's will have OpenCL support.
I'm sure if Apple put up enough fuss, Intel may even make versions with overclocked IGPs, like how current iMacs use overclocked Santa Rosa platform chipsets and CPUs specifically made for Apple rather than the newer Montevina platform chipsets and CPUs.
Originally Posted by allblue
Thanks for taking the trouble to write such an informative post (and winterspan likewise.) Much appreciated.