Note that ~15% of Steam Players are gaming on Intel iGPUs, as awful as they are.
Intel had the ability to put more transistors on the same die size with the same power requirements. This was long after they reached thermal/clockspeed limits (with the P4). They started putting additional cores in there and bumped up the L2 and L3 caches, but there was still space left on the die.
What do you do with those extra transistors? It would be absurd to "leave them blank" as you are basically throwing money down the drain.
The iGPU needs to be manufactured anyway. And with memory controllers integrated into the CPU, it seems rather complicated to push the iGPU off the chip.
> and more RAM bandwidth
Cheap iGPUs on Motherboards would recycle the CPU's memory controller anyway. Did you work with computers in 2006 or so? Its definitely cheaper and more efficient to just integrate everything into the die in contrast to the designs of the past.
Er… you mean like every single mobile SoC?
Which means if one part is defective you can use a different part boosting yields. But, they also get much lower latency.
Now recent Phone Chips have included on die GPU's but phones are so power limited it's not an issue.
HBM is only being used on AMD's R9 Fury Graphics Card, as far as I'm aware. A rather high-end solution that simply doesn't exist in the phone market right now.
PS: I could get more accurate numbers, but with such a wide range of latency number that's pointless.
As I noted before: Mobile CPUs are using standard DDR3 or DDR4 RAM, on a motherboard drawn with copper wires to the CPU. The ONLY use of "3d stacked memory" that I know of is the HBM RAM of the AMD Fury X.
http://www.legitreviews.com/wp-content/uploads/2015/05/HBM-D...
I'm laying counter to your fact-claims. Only high-power GPUs have stacked RAM right now.
Anyone who has opened up a phone can see the DDR3 or DDR4 RAM on the phone's motherboard, with the phone SOC designed to go "inefficiently all the way" to the DDR3 RAM over "long latency" copper wires.
Read the spec sheet yourself if you don't believe me.
http://system-on-a-chip.specout.com/l/1107/Qualcomm-Snapdrag...
Anyway, I am referring to packaging like: http://cdn.arstechnica.net/wp-content/uploads/2013/04/ASIC_+... Yes, HBM also does more or less the same thing. The issues are twofold heat which is vastly less of a problem in the mobile world so with HBM it's a stack of ram chips next to a processer, with Mobil the CPU is under what might only be one RAM chip and they call it SoC.
http://arstechnica.com/gadgets/2013/04/the-pc-inside-your-ph...
Note: CPU's are already 3d with many layers depending on how you count.
Anyway after die size, heat, latency, and manufacturing costs are really the largest factors. Some PC ram chips run hot enough they really can't be stacked. You can slow them down and trade latency for total bandwidth. ed:Less of an issue with active cooling and a heat sink. However, they can call this HBM, because you can have more chips so even if each chip is slower overall you get higher speeds.
PS: Don't forget HBM is designed for a GPU/CPGPU without a large cache. CPU's have a huge cash which let's them have very different approach aka lot's of cheap but slow RAM. Benchmarks show faster ream really does not add much for a modern CPU with most workloads. You could use this for a CPU but it's harder to swap in more ram.
I am pretty sure that if keeping the memory controller, northbridge, and iGPU separate, Intel would have kept these things separate.
My bet? I think the cheaper manufacturing costs associated with motherboard construction beat-out the slightly increased cost of tighter integration. But I'm not an Intel Engineer. Just a hunch... considering that Apple A9, Qualcomm Snapdragon, AMD A10 "APUs" , and Intel have all decided to integrate everything together.
> better thermal envelope processing speeds
What, are people actually using iGPUs for difficult tasks all the time? As far as I can tell, most people surf Facebook all day and ping the CPU.
iGPU isn't slowing down anything with Facebook surfing. Besides, a crappy integrated GPU on the MoBo would still use up RAM bandwidth.
If anything, integrating the GPU has made the process cheaper for Intel. The Northbridge, Memory Controller, and iGPU are all integrated into modern CPUs, because Intel has done the number crunching and has determined that its cheaper to do it this way.
It's much easer to cut rectangular chips though. Still, even if the used the same overall design and just ignore any defects over X area would boost yield.
Thus main memory is much slower than GPU bandwidth. So, now your on chip GPU is memory starved and reducing the bandwidth avalabel to the CPU. There are some advantages if you want to pass stuff back and forth, but the standard rendering pipeline is designed for GPU's.
This ate up valuable chip real estate, RAM bandwidth, thermal overhead etc. Worse it cemented the idea that Intel was crap at graphics while slowing down the PC upgrade cycle.