Blogs What can Intel do to improve its processor situation?

Stop throwing everything at the wall and seeing what sticks...

 

They could offer the same chips on a more expandable platform (more lanes etc) I'd buy it, could be on 16nm for all I care. I find mainstream platforms aren't good enough for supporting latest storage technologies.

Ryzen 3900/3950x(should it come out) is much more powerful than my current machine I'd love to jump on it for a bit of extra gaming prowess because it is cheap and lots of threads but I couldn't plug all my crap in due to same problem.

 

I think they need more SKUs with longer names.

 

I know i mentioned it in another thread but they've painted themselves into a corner with how much they've segmented their product stack, if they introduced HT on everything it would make some CPUs irrelevant, if they increased PCIe lanes it would do the same, pretty much whatever changes they make risks making something else in the product stack irrelevant.

 

Practically: continue as they have been. There are no quick fixes, and long0term fixes are just stuff-they're-already-doing.
Their bread and butter is enterprise sales, which continue to go gangbusters both due to supply chain and inertia (imaging trying to swithc an entire stack from IME-based management to a weird mixed environment incorporating some new and likely buggy middleware). Mid- and high-end Laptops continue to be sewn up through power efficiency (by the time Zen 2 arrives in the 4xxx APU series, Lakefield will already have been on the market for some time). Server and HPC are seeing major challenges more from Nvidia and Google in NN workloads, hence Nervana. There is no 'go faster' button for 10nm scaling, and the same scaling walls Intel are hitting are the ones that 5nm will hit at TSMC and Samsung (unless they pull another 7nm and perform a partial scaling). For every fab, cost/transistor is still rising as it has since 22nm so focus remains on architecture improvements rather than expecting process scaling to make things up. Designing Sunny Cove in lockstep with 10nm design rules was in hindsight a big misstep, but not one that can easily be rectified without basically branching architectures (i.e. 10nm Sunny Cove + 14nm Kinda Sorta Sunny Cove Ish).

Bigger picture, I think chasing high core counts is going to be seen in a few years as being as much of a folly as Heterogenous Compute was at the start of the decade. In the server & HPC world we've seen workloads either scale as embarrassingly parallel and jump to GPGPU or dedicated coprocessors, or stubbornly remain Amdahls Law limited. Adding more and more cores to client chips is great for benchmarks and the handful of users who are doing local-client software rendering (and RT accelerators are soon going to eat that market as they already are doing rapidly in production 3DCGI), but for 99% of users they're buying idle cores.

 

In the post "Core count war" world it will be all about stacking different types of chips:

https://bit-tech.net/news/tech/intel-reveals-foveros-3d-packaging-technology/1/

 

All they really have to do is make their upper tier CPU's price competitive...

Until a month ago I was going to build a Ryzen 7 rig but then for some bizarre reason the 9900K fell in price on Amazon to £405 and I bit their hand off at that price (now back up to £500, BTW)

...with a decent Z390 Motherboard (Aorus Pro WI-Fi) for £150 and 32GB of 3200 DDR4 for £150 it was cheaper than buying a Ryzen 3700X with a similar quality X570 Aorus board (£250) and 32GB of 3600 DDR4 (£250).

Everything changes when prices are at the level they should be.

 

But isn't that irrelevant? since if you are going to buy intel, you are also cutting off features for future upgrades, such as faster storage. Intel isn't going to support PCIe 4.0 at all.

 

Intel just need to stop changing socket for (mostly) no reason other than making money...

 

I wonder how much pci-e 4.0 will be a factor. I mean it's now available on boards about 2 years after being released.

Pcie 5.0 was already officially released this year so it won't be too long either until that's about.

I know the 5700(xt) is pcie 4.0 but when will gpus actually need it?

 

GPUs? Maybe a decade or more. PCIe 3.0 was published in 2010, and only now can you just start to see any real bottlnecking of a PCIe 8x interface... with dual GPUs.

 

That's what i was mulling over.

Soon to be building a Zen 2 but an nvme and ssds are fast enough for me on the x4xx platform - x5xx, well by the time a single gpu needs more than pcie3 i'll need a new rig anyway.

 

And if you wanted a faster SSD you can do so on PCIe 3 as well, the Intel 905p massively improves worst case scenario performance over "traditional" NVMe drives.

 

Yep, PCIe 4.0 will improve sequential performance, but does diddly-squat for the random performance that benefits responsiveness.