Intel exiting the PC business as it stops investment in the Intel NUC

Thanks, that is a useful differentiation: laptops for portability and "good enough" for office work, NUCs and small form factor computers for performance in a limited space.

System76 makes a similar sized unit called the Meerkat though I’m not sure how much it compares in other specs besides runs Linux and size. https://system76.com/desktops/meerkat

I have seen R86S and variants, but as you said it's not quite there either. I understand they are achieving this by putting an OCP 3.0 slot on the bottom of their SBC (though probably with limited PCIe lanes). I'd be happy to see the trend continue.

I thought NBase-T was gonna finally save us all until Marvell bought Aquantia and they started hosing us with speed-limited variants of their chipsets. Now everyone sees a 2.5gb nic and they are stoked for the huge upgrade instead of recognizing that it's really just a crippled 10gb nic

> What Apple is doing with its combo of core performance, unified memory setup, specialized compute blocks on the SoC, and the overall thermal efficiency which allows all of this to just run and run without throttling all that much--it adds up to way more than the sum of its parts. You really have to use it in your daily life to believe it and feel the difference.

The future that fusion-HSA promised is finally here. Everyone drools over the possibility of PCs getting console-like zero-copy shared memory between the various accelerators, and people want analogous features to be ported onto the current dGPU/CPU paradigm (like directstorage). Fusion-HSA never got there itself, otherwise iGPUs would be able to do zero-copy already, but Apple has done it and the silence is thunderous.

Nobody cares, everyone is just waiting for AMD to implement something competitive and then it'll be cool. Strix Point/Strix Halo I guess.

PC enthusiasts are gonna hate this but I don't think you'll ever be able to really get to a high-performance APU without some kind of soldered memory. Consoles use soldered memory too. GDDR6 vs stacked LPDDR5X is a design call but both are clearly superior to socketed memory, you'd need an Epyc-sized socket to get an equivalent amount of memory bandwidth into an APU, and it'd pull an enormous amount of power for PHYs too. You'd probably end up with like 30w of idle power lol, meanwhile Apple is doing 25W for the whole chip. Crazy stuff.

The practical way that x86 is going to get there is what AMD is doing with Strix/Strix Halo and Intel is doing with Adamantine - you have to go to stacked cache to make up for the lack of memory bandwidth (and perhaps even still go to quad-channel like Strix Halo), and it's still going to use a lot more silicon (expensive!) and use a lot more power than just stacking some LPDDR5X on there and calling it a day. You can't get to 6-8 channels worth of bandwidth from 2 actual channels without some kind of a hack, either you run the pins super fast (GDDR6) or you move the channels on top of the package (LPDDR5X), and both of those need to be soldered. And that's a large part of why consoles and Apple Silicon can deliver a relatively large punch (3060 perf with a good CPU at 25W package power is nothing to sneeze at!) at consumer-friendly prices.

Low-end dGPUs are done for until stacking gets common, but APUs and low-end dGPUs are such an amazing impedence-match for stacking and people can't see it because it says Apple on the box instead of AMD. It grinds me when people ignore or shit-talk really cool advancements in tech just because it doesn't fit their mold or their brand, this is what everyone has been waiting 10 years for.

> I still have a mix of Intel machines and Apple Silicon machines for my work and personal life, and it's just so immediately apparent the latency difference in usage. Apple Silicon feels and runs so much better.

To be fair, some of this is the fact that it's *nix. If you run Linux on your Intel machines it'll be snappier than windows too, my 5700G SFF build is a crazy machine for linux. But honestly having a well-supported *nix ecosystem with first-class vendor support is a good value offering, I think that's why Apple is having a surprising renaissance with computer-touchers right now. Non-techies get the happy bubble OS, techies get something they can drill down to the terminal and do their thing with dotfiles/zshrc. And everyone likes the fact that they're a well-built laptop with incredible battery life and good performance while mobile.

> Sure, would I recommend more than 8GB of RAM? Yes. More SSD is better? Of course. But Apple Silicon in even its thinnest entry-level configuration is an auto-recommend for me. And the prices at that entry-level are so, so reasonable for what you're getting.

I agree with both you and the commenters you're responding to. If the entry-level models will work for your needs, they're value champs. You will not find a similar value offering to a mac mini at the $400 edu pricing for example, if your use-case fits into 8GB/256GB it blows away anything else at that price point. Often there are some deals at bigbox stores or electronics retailers (B+H, Best Buy, Costco, etc) that offer some decent prices on the higher range stuff as well. A loaded-out M1 Max 16" (MK233LL/A) is $3300 on B+H right now (and it was $3200 a couple weeks ago) and you can find 32GB/1TB manufacturer refurb (applecare-eligible) M1 MBPs on Woot pretty regularly. The refurb store also allows you a lot of the flexibility of custom configuration but especially when combined with the edu/veteran discounts it gets you close to the level of that bigbox or refurb pricing. M1 Max 10C/24C with 64GB/1TB is about $2550 for example, that's a pretty nice machine too.

I would really say that if you're a developer you probably do want at least 1TB storage. A lot of things rely on being able to store docker images/pipenvs/node packages/etc in the expected place, and while I'm sure you can configure them to run on an external, it's a pain, and they aren't configured to "float away to the cloud" like apple does with their first-party apps. A 256gb spec is a thin client/cloud terminal only - used 16GB/256GB MBAs are very very cheap compared to the higher models and I have to think some of it is because people try it and learn the hard way 256GB isn't enough for them. This includes me, and while I could never have gotten to "yes" on a loaded MBP or even a 512GB or 1TB upgrade on the MBA, 16/256 definitely did not work the way I'd hoped even as a homebody with a NAS. They really are aimed at momputers and people deeply into the icloud ecosystem where everything can be silently swapped out on-demand.

That said I definitely do feel the complaint others are making that Apple basically does "product tiering by RAM/storage envy", as I once heard someone call it. Really the difference between a MBA and a MBP or between the different chassis sizes is pretty minimal when you equalize for RAM/storage, a decently loaded 15" MBA is at least $1600 and probably closer to $2k, and that also gets you an entry-level MBP or a refurb loaded (32GB/1TB) MBP. $2k for a laptop is a lot but nobody else has the kind of laptop performance and battery life that Apple does right now, so it's kind of a question of whether you're just looking for the cheapest thing that checks the boxes or if you're looking at the offering holistically. Nothing wrong with an XPS or a Latitude or Thinkpad or whatever either, but if you're spending "premium ultrabook/business laptop/gaming laptop with dGPU money" you can definitely get a real nice macbook too and it both has unique selling points and targets a different set of needs.

And Apple has perfected the art of stacking their tiers perfectly so that you can talk yourself into getting the next higher model. That's why they're the most valuable company on the planet, lol.

> I mean I could be wrong, but Intel has gotta be looking at Nvidia's Hopper or AMD's MI300 DATA oriented architectures and wondering why the Arc team is trying to make $750 gaming computers.

That's the same strategy AMD has tried with ROCm (de-facto pro-only compute ecosystem) and it hasn't worked. Nobody is going to buy a $5000 accelerator card (or even rent AWS time) just to tinker with your ecosystem unless it's already known to be the bees' knees. NVIDIA built their success by making sure their cards were the first thing people reached for when they wanted to build a high-performance application on a compute accelerator.

Further, there's a lot of redundant work and overlapping lines of business here. You have to develop game drivers for the iGPU for anyone to take you seriously, so at that point why not make the dGPU cards and increase the ROI multiplier of your work? Or are you planning on outsourcing the whole shebang and just licensing a hardware SIP core and a driver stack? There's really only two other names with a credible Windows/Linux driver stack... NVIDIA and AMD. Imagination/PowerVR don't have any windows presence, and Intel already tried this back in the early Atom era and it fucking sucked.

Like yea you've actually pointed out the exact reason they won't cut it: Hopper or MI300 is where the HPC and performance-compute segments are going and you can't be credible in that space without an internal solution for the accelerator half. It doesn't have to be GPUs, or they don't need to have graphics pipelines, but once you are doing all the work to develop the GPGPU side, you might as well make a variant with a pixel pipeline and display outputs and sell it to enthusiasts too.

The semiconductor space has a lot of these overlaps in product verticals, and if you choose not to be in them, you're leaving revenue on the table that is relatively "cheap" to access. The same CPU uarchs that work in enterprise work in consumer, and in the grand scheme of things re-using the same uarch for a low-margin product is fine. You'd never pay to develop the product from scratch just for the consumer market, but once you've done the work, you might as well sell it to consumers too. If you're doing consumer, you probably want to be doing laptop too, but those need iGPUs. Which need drivers, and if you're doing drivers you might as well do dGPUs too.

Intel doesn't need to be exactly AMD, but AMD has gone through this and already cut to the bone on everything they didn't need. You can draw a circle around the product verticals they're in and they've pretty much identified the core business requirements for the CPU/GPU markets.

You're right though that Gelsinger is obviously stripping the business down to his own vision of those essentials. I just think right now the evidence indicates that GPUs are still a part of that vision. You can't be competitive in HPC without GPGPU, you can't do laptop without iGPU, they might as well do gaming GPUs too, especially since a lot of the GPGPU research will overlap. The drivers don't, but you have to do them for iGPU anyway for laptops.

The fabs are the rough one though. They're expensive as shit, but all intel's legacy IP uses them, and the only thing the fabs run is Intel's IP. I think he's serious about building the wall between IP and fab at intel, and about bringing in external business, but right now there is no hope of even a GloFo-style spinoff working. It would absolutely take down both halves of the company to even try.

He's definitely got a tough road, Intel's pain is only just beginning and it's going to be a long time to profitability.