I think Intel showed us with Lunar lake and Arrow lake that a lot of that energy efficiency has to do with SOC packaging. When you move the ram, storage, etc on die, you save system power and improve performance at increased production and switching cost and sku numbers. After all- now you need to pay for ram and storage in addition to the CPU, and can't swap them out.
I think the efficiency benefits of on-package memory are severely overrated, because of Apple fanboys or autisms. It can reduce power consumption and allow increased memory speed somewhat but it's not a magic bullet. Also, the storage isn't moving on there. That's something to think about when we finally get universal memory in 10+ years.
There are a lot of optimizations and design changes in Lunar Lake, from Meteor Lake-U. Very notably it goes from 4 chiplets (not counting Foveros base) to just 2. It moves graphics back onto the compute tile, using TSMC N3B instead of Intel 4 for CPU and TSMC N5 for graphics. I don't know how N3B compares to Intel 4 but I imagine it's better. Lunar Lake surprisingly drops the "low power E-cores" after just one generation, having the massively improved (according to PowerPoint slides) Skymont E-cores handle low-power situations. Instead of 2 P-cores, 8 E-cores, and 2 LP E-cores (2+8 has been the config since Alder Lake), it uses 4 P-cores and 4 E-cores. Lessons were learned and have been applied in Lunar Lake.
I don't think we know enough about Arrow Lake mobile/desktop to say a lot, or I need to pay more attention. But I did read that
there will still be an Arrow Lake-U, which means Lunar Lake is not a replacement for the "-U" die if that's true. I do hope we see "Adamantine" L4 cache during this half of the decade.
Realistically speaking, where will we go with the 5000 series? Everything since the 2000's feels like dimishing returns.
Only thing I can think why I'd want to upgrade my GPU is to run more local AI models and maybe Helldivers, since that's the only game I struggle to run. Everything else is just mediocre and most games I play are older than 5 years anyways.
We can expect that the 5090 and 5080 will launch first, so for those who have no interest in spending $1,000+ on a GPU, you will have to wait much longer into 2025 to care.
From the info that has been leaked about the dies used, it looks like Nvidia is continuing the strategy of making the 90-class card get an impressive 50% or better generational gain, while other cards below that will be shuffling tiers around again. We can be pretty sure that the 5080 can't beat the 4090, because sanctions would prevent it from being sold in China. Maybe it can win in gaming from the bandwidth but not compute?
If you are feeling no pressure to upgrade for gaming, you can safely ignore GPU launches besides glancing at them.
For AI, the 5090 is rumored to use a 448-bit bus (cut down from 512-bit which will be used with pro cards) for 28 GB of memory instead of 384-bit / 24 GB this time. And it will use much faster GDDR7. That could translate into some great improvements for AI, although I wouldn't be surprised if having under 32 GB is a trick to keep it from running some models as good as the professional cards can. -4 GB doesn't seem like a lot but 32 GB is ~14% more than 28 GB, and there have already been 32 GB pro cards. Some models are probably designed to use that much VRAM.
If you're looking for something cheaper, based on those memory buses, you're likely to see the 5070 with 12 GB, 5080 and maybe a 5060 Ti with 16 GB. Same as before. But all of those will use GDDR7 and the increased memory bandwidth should be helpful. The lowest die listed for 8 GB budget cards uses GDDR6.
Cyberpunk 2077 was the first and last game release I felt hyped for.
I want to feel something for TES VI, but a lot has changed for the worse since 2011. So I'll keep my expectations low, maybe pirate it 5 years from now and be pleasantly surprised.
