Intel is seeking feedback for x86S, a 64-bit-only version of x86 for future processors.

So there's a lot to go through here, but I think I'll try my best to answer your questions. I'm a package engineer for Intel, which means I deal pretty directly with how to stick components together. This is essentially the "monster APU" problem. >Why not someone come up with some sort of a solution where CPU and GPU sit closer on the same board and graphics card is simplified. And we no longer have a huge needless motherboard? This is possible and it's done extremely often. Behold! The humble gaming laptop. The CPU and GPU are inches apart, often share a cooling solution, and are directly a part of the motherboard that breaks out all of the I/O from the chips. Going a step farther is the APU, which has the CPU and a more advanced iGPU than would be needed to just drive a display. I make this distinction to rule out Ryzen 7000 desktop CPUs and Intel's desktop CPUs, as well as most of the older mobile CPUs. On the APU front the CPU and GPU now not only share the same motherboard, but are the same physical chip. This is as close as they get. Future designs with a better unified memory setup where both can access the system RAM freely (see consoles and Apple M series) will further unify them into one thing. >How much of the graphics card other than GPU is actually meaningful? I would like to see some kind of a modular SoC design where we can change GPU and CPU on the same chip but not have a huge motherboard even on ITX size that feels mostly unused I would assume. Most of a GPU's board is power delivery and space for memory ICs. By moving the GPU hardware into the CPU, now APU, you need the power delivery and memory to support it. Memory can be brought on-package for rather extreme cost via HBM (see Radeon Vega and Nvidia's H100) but power delivery has to stay external. This moves the area required to hold your GPU's VRMs and VRAM on to the motherboard, and while it's not 1:1, you still end up taking up more room around the socket than we do currently. Better power management is possible for this single package, so the whole VRM wouldn't need to move, and the biggest space saving would come from using HBM in a unified memory setup. This means your CPU, GPU, RAM, and VRAM are all on one chip and have to be replaced at the same time. The socket for this is going to be quite large as well. Mini ITX is actually quite dense already, look at one without all the heatsinks and covers on it. If you want to go tiny, you can get mobile chips soldered directly to Pico-ITX motherboards already, which are like 1/3 the size. ​ > The point of this would be simpler and less space taking PCs. I own an 8L ITX case and some people own 4L cases with 2-slot mid to high range really good GPUs. But I would like to see PCs in the size of NUCs but actually with a decent GPU. Think about 4070 performance. Sure I guess those are called laptops but we all know how those things are. Not modular. Space heaters. LOUD. And not performant enough. This thing won't be either, and the mobile 4090 is already pretty close to that performance target if not out in front. If an 18-inch laptop weighing several kg is challenged by that chip and a fast CPU like the 13980HX, then this NUC is going to be doing it's best to replicate the surface of a star. Space constrains like this for heat loads like that are going to mean we get creative in the cooling department, and we're going to start by cranking the fans to high hell. Faster fans are effectively free thermal headroom for an engineering perspective. The next things to go are socketed RAM, standard motherboard layouts, and then socketed chips to extract every last watt with direct-die cooling. We now have the motherboard of a gaming laptop in a box. Sure, that HBM-stuffed mega APU from earlier makes the motherboard potentially smaller, but it also costs the moon to make. ​ >I really don't think all that stuff on a graphics card indispensable. This is more of a joke than anything else, but sure, go ahead, start plucking parts off the PCB and see how many aren't needed. Almost everything on the PCB needs to be there, and the large coolers exist because the current market values noise and cool temperatures over size. You can cool 350W in a dual-slot card, but you're not going to like being around it and people are going to complain that it's redlining at the thermal limit under load. This is fine for the chip btw, your CPU doesn't care if it's at 45C or 85C, it just knows it doesn't have to slow down yet. ​ >We know consoles use VRAM for system. Now I'm not sure if VRAMs can really be used for all jobs on a PC - probably not - but if it can, then you would save more space there. The latency is actually atrocious for a CPU. They're designed for different goals. GDDR is bandwidth first, while DDR is more aimed at latency and smaller chunks of data. LTT did a video with a defective PS5 APU on a motherboard some time ago and benchmarked the memory. It's not pretty. If you think of it this way, for every cycle a 2.5ghz GPU spends waiting on memory, a 5.0ghz CPU spends twice as many. This is really bad for the CPU, as all of those cycles are functionally wasted energy and time. HBM from the example APU above just brute-forces past the issues by being physically touching the main die (good for latency) and having a gigantic bus to send data over, like kilobits wide. It's not particularly hard to imagine a kiloBYTE bus width for HBM3. That's 8 stacks or 128-192GB from SK Hynix right now. The H100 has 6 stacks for comparison. You may be better off designing a quad-channel DDR5 controller and letting the APU sort out what gets what at a given time, but this makes the silicon design more complicated and now you have to pay for the DRAM for 4 channels instead of 2, either getting 4x 8GB single-rank or 4x 4GB for a more basic system. ​ >I expected AMD to try something like this when they bought out ATI considering they both make GPUs and CPUs. They do have APUs but it's just not the same thing. And they still need motherboard. Now that Intel is in GPU market, maybe they will try? I don't really understand how an APU isn't what you want. It's as close as the CPU and GPU can physically get, even sharing the VRM and DRAM. You're always going to need a motherboard. There's no way around that. There needs to be some layer between the end user and the physical silicon, or else you can't do anything with it.