Hard to believe it's been 24 years since Y2K (2000) And it feels like we've come such a long way, but this decade started off very poorly with one of the worst pandemics the modern world has ever seen, and technology in general is looking very bleak in several ways
I'm a PC gamer, and it looks like things are stagnating massively in our space. So many gaming companies are incapable of putting out a successful AAA title because people are either too poor, don't want to play a live service AAA disaster like every single one that has been released lately, Call of Duty, battlefield, anything electronic arts or Ubisoft puts out is almost entirely a failure or undersales. So many gaming studios have been shuttered and are being shuttered, Microsoft is basically one member of an oligopoly with Sony and a couple other companies.
Hardware is stagnating. Nvidia is putting on the brakes for developing their next line of GPUs, we're not going to see huge gains in performance anymore because AMD isn't caught up yet and they have no reason to innovate. So they are just going to sell their next line of cards for $1,500 a pop for the top ones, with 10% increase in performance rather than 50 or 60% like we really need. We still don't have the capability to play games in full native 4K 144 Hertz. That's at least a decade away
Virtual reality is on the verge of collapse because meta is basically the only real player in that space, they have a monopoly with them and valve index, pico from China is on the verge of developing something incredible as well, and Apple just revealed a mixed reality headset but the price is so extraordinary that barely anyone has it so use isn't very widespread. We're again a decade away from seeing anything really substantial in terms of performance
Artificial intelligence is really, really fucking things up in general and the discussions about AI look almost as bad as the news about the latest election in the USA. It's so clowny and ridiculous and over-the-top hearing any news about AI. The latest news is that open AI is going to go from a non-profit to a for-profit company after they promised they were operating for the good of humanity and broke countless laws stealing copyrighted information, supposedly for the public good, but now they're just going to snap their fingers and morph into a for-profit company. So they can just basically steal anything they want that's copyrighted, but claim it's for the public good, and then randomly swap to a for-profit model. Doesn't make any sense and just looks like they're going to be a vessel for widespread economic poverty...
It just seems like there's a lot of bubbles that are about to burst all at the same time, like I don't see how things are going to possibly get better for a while now?
Power consumption numbers like that are expected, though.
One thing to keep in mind is how big the die is and how many transistors are in a GPU.
As a direct-ish comparison, there's about 25 billion transistors in a 14900k, and 76 billion in a 4090.
Big die + lots and lots of transistors = bigly power usage.
I wouldn't imagine that the 5000-series GPUs are going to be smaller or have less transistors, so I'd expect this to be in the die shrink lowers power usage, but more transistors increase power usage zone.
You can also get big power consumption from turning up the voltage and cranking the clock speeds well past their efficient zone. You see that right now with most 40 series cards where turning the clock speeds down a smidge gives you huge power savings at almost no loss in performance.
Cost per MM^2 of die space has only gone up with each process node these last 10 years, so unless you're paying big money don't expect big chip.
Conversly, the apple silicon products ship huge, expensive dies fabbed on leading TSMC processes which sip power relative to contemporaries. You can have excellent power efficiency on a large die at a specific frequency range, moreso than a smaller die clocked more aggressively.
You're not wrong (and those are freaking enormous dies that have to cost apple a goddamn fortune to make at scale), but like, it also isn't an Apples-to-Apples comparison.
nVidia/Intel/AMD have gone for the maximum performance and fuck any heat/noise/power usage path. They haven't given a shit about low-power optimizations or investing in designs that are more suited to low-power usage (a M3 max will pull ~80w if you flog the crap out of it, so let's use that number) implementations. IMO the wrong choice, but I'm just a computer janitor that uses the things, I don't design them.
Apple picked a uarch that was already low power (fun fact: ARM was so low power that the first test chips would run off the board's standby power and would boot BEFORE they were actually turned on) and then focused in on making it as fast as possible with the least power as possible: the compute cores have come from the mobile side prior to being turned into desktop chips.
I'm rambling but: until nVidia and x86 vendors prioritize power usage over raw performance (which they did with zen5 and you saw how that shit spiraled into a fucking PR shit mess) then you're going to get next year's die shrink, but with more transistors using the same power with slightly better performance. It's entirely down to design decisions, and frankly, x86 (and to some degree so has nVidia) have painted themselves into a corner by relying on process node improvements (which are very rapidly going to stop happening) and modest IPC uplifts to stay ahead of everyone else.
I'm hoping Qualcomm does a good job staying competitive with their ARM stuff, but it's also Qualcomm and rooting for them feels like cheering on cancer.
This outlines several issues, a key one is outbidding apple for wafer alloc on leading processes. They primarily sell such high margin products that I suppose they can go full send on huge dies with no sweat. Similarly, the 4090's asking price was likely directly related to it's production cost. A chunky boy with a huge l2$.
I like the way Mike Clark frames challenges in semi eng as a balancing act between area, power, freq and performance (IPC); like a chip that's twice as fast but twice the size of its predecessor is not considered progress.
I wish ultra-efficient giga dies were more feasible but it's kind of rough when TSMC has been unmatched for so long. I gather Intel's diverting focus in 18A, and I hope that turns out well for them.
I'm not sure that arm as an ISA (or even RISC) is inherently more efficient than CISC today, particularly when we look at Qualcomm's latest efforts in notebooks, more that Apple have extremely proficient designers and benefit from vertical integration.