Slashdot Mirror
Intel Confirms 8th Gen Core On 14nm, Data Center First To New Nodes (anandtech.com)
Ian Cutress, writing for AnandTech: Intel's 8th Generation Core microarchitecture will remain on the 14nm node. This is an interesting development with the recent launch of Intel's 7th Generation Core products being touted as the 'optimization' behind the new 'Process-Architecture-Optimization' three-stage cadence that had replaced the old 'tick-tock' cadence. With Intel stringing out 14nm (or at least, an improved variant of 14nm as we've seen on 7th Gen) for another generation, it makes us wonder where exactly Intel can promise future performance or efficiency gains on the design unless they start implementing microarchitecture changes.
The hope is that AMD's RYZEN will be good enough to compete with Intel in performance - not just price. That will wake Intel again, since they are always relaxing when there is no competition i.e. no motive to do something more.
Violence is the last refuge of the incompetent. Polar Scope Align for iOS
Yesterday prices were leaked on AMD Ryzen. For equal peformance, the AMD parts are abot 70 percent cheaper. Intel has been goofing off for several years now. Tweaking process improvements is not innovation. Intel's Architecture is tired and needs to be rethought. I'm really surprised that Intel has been caught with their pants down.
consumer products need more pci-e lanes AMD is doing better with ryzen. 16+4+4(chip set link?) Also USB 3 may be in the cpu as well.
ryzen server / workstation may have even more pci-e lanes and will there be 1 socket systems with 32 or more pci-e lanes + chipset link + 4? that can go after the high end consumer products form intel that are a gen behind there consumer products?
Competition is quite a bit behind Intel at the moment, so no reason to move forward while they can milk this current generation. Once competition starts getting *near* 14nm.... Intel will nudge forward to keep a few steps ahead.
What's beyond 7nm though?
It's another confirmation that Moore's Law is dead. If Moore's Law were still in effect, Intel would make their new chips at smaller geometry regardless of competition because it would be cheaper to do so and that would make for fatter profits. Cost per transistor is the driver of Moore's Law. That stalled at 28nm because that was last node that could be made without resorting to multi-patterning. Scaling worked in the past because the cost to make a wafer was roughly constant. By making features smaller, you either got more chips or bigger chips for the same cost. Multi-patterning means the cost per wafer as you scale down is going up faster than the transistor count per wafer. Performance still increases but you have to have customers willing to pay more. If the cost delta is large enough, Intel my not jump to 10nm even if AMD catches up. Process performance isn't the only knob they can turn to improve performance.
Okay fine, so 4th gen isn't literally faster than 8th gen, but I agree with what OP is getting at... What the graph you posted is best at showing is that Intel CPU performance improvements have been paltry for the past six years.
According to your graph, the new Kaby Lake 7700k is only ~55% faster than my 2nd generation Sandy Bridge 2600k. Which means that between January 2011 and January 2017, Intel performance improvements for like-for-like CPU's has been about 7.5% per year, which is pretty shitty. It's not that 8th gen is going to suck as bad as 7th gen -- it's that both 7th gen and 8th gen suck as bad as everything Intel has released fort the past six years.
Intel needs have it's ass kicked the cutting pci-e lanes on a $400 chip that in last gen had way more no you need to go up to a $600 chip to get them back and that is on the last gen workstation / server sockets. The desktop boards have been stuck on the same pci-e lanes for years and maxing out at quad core.
AMD is going have more pci-e and more cores on the desktop boards then what intel has. With the server / workstation ones like to have even more then what the amd desktop boards have.
It's actually worse than that though if you're just looking at the architecture. The 7700k has a 20% clock speed advantage over a 2700k, which means that their architectural improvements aren't even 7.5% per year. Both of those chips exist within a similar TDP bracket as well (91W vs. 95W) so it isn't as though Intel has been using process improvements to offer the performance at lower power consumption levels either. And that's only certain benchmarks as there are others where Intel's older chips perhaps only fair worse by single digit percentages once accounting for clock speed differences. Sandy Bridge was a great chip for overclocking and it wasn't difficult to get as much as 4.4 GHz without putting a lot of effort into it. Some enthusiasts have been able to get up to 5 GHz with a good chip and cooling solution. The newer Core i7 chips usually require de-lidding since Intel uses a substandard TIM which doesn't transfer heat effectively enough.
Intel needs a new microarchitecture to replace Core. Core was an exceptional design, especially considering what it replaced and how much the early performance gains were like if you bought an early Nehalem CPU. Hell, even Core itself traces its roots back to the P6 microarchitecture after Intel abandoned Prescott (which was sold as the Pentium 4 back in the wild days of the clock speed wars) which goes back decades. It's pretty clear that Core is tapped out in terms of what can be squeezed out of it and Intel needs to go back to the drawing board like AMD did and use all of the lessons they've learned to make a new architecture.
Even if AMD's offerings aren't quite as good as Intel's, they'll still be closer than they ever have before and it will allow AMD to challenge Intel in their high-margin consumer market segments or in markets were AMD hasn't been relevant in years. Intel could afford to tread water while AMD was using their failed Bulldozer architecture, as Intel would just as gladly sell you a 4 year old CPU as a new one if the prices hadn't moved much, but now AMD is going to erode those price points or offer a competing product if they don't undercut Intel. Intel will still have a process advantage with their own fabs, but they need a new architecture to widen the gap if they want to have any hope of maintaining their profit levels.
When Intel struggled to get Broadwell out, their die shrink to 14nm using the architecture that they made in Haswell, you knew that they were having at least some issues. When it turned out that Haswells almost exclusively didn't properly support the new "transaction memory", to the effect that the opcodes had to be patched out, that was also kinda depressing. Skylake, their next in line, and the newest architecture update, was the last time they have even vaguely been on schedule.
Right after skylake, they announced that, instead of a die shrink to 10nm, they would add a new "optimization" step, and continue to tweak skylake instead of shrinking it. This is kabylake, which just came out in desktop and laptop properly (Xeons lag behind normally: the full suite of Skylake Xeons should be launching in a few months). They redid all their slides to show a full new arrow, giving them effectively another year to do the die shrink. Now that we are getting close to seeing what would be the next guy ("cannon lake"), who properly should be launching later this year on 10nm, we first heard that they were going to insert a "coffee lake", which would be another optimization at 14nm, for desktop, and that only laptop and low power chips would actually be on the 10m "cannon lake". And now, we find out that the first 10nm will be out for datacenter, which means an even further push back.
Summary: their older slides used to show around a summer 2016 launch for their 10nm process. Then it became a summer 2017 launch, then that became only a partial launch, and now it is looking like a spring 2018 launch. The words change, but the message is the same: "We aren't close to having 10nm be actually profitable, or possibly even all that functional".
Intel has been talking about 10nm for at least 3 years. They "pretended" to show off a 10nm chip recently, but all indications point to maybe 2018... the launch of 10nm has been delayed at least three times (official announcements).
When did intel announce 10nm chips || date range [2015 - 2016]
It's not for a lack of effort on Intel's part. Despite the misconception, Core isn't one microarchitecture. Intel has revved the architecture several times at this point, always incorporating some of the latest theories and designs on branch prediction and the like.
The issue is that Intel's on the bleeding edge of single-threaded performance, and it's increasingly looking like there's nowhere left to go. Your options are either to increase the clockspeed - something that's proving impractical due to power consumption going crazy past 4.5GHz or so, even with FinFETs and other adaptations - or you increase the IPC.
And on the latter point, Skylake/Kaby Lake is already a wide, deep out-of-order architecture with more execution resources per core than most threads can take advantage of. The linchpin to IPC is out-of-order execution to fill these complex cores, and on that front Intel is already well into the diminishing returns stage. The re-order buffer is now 224 entries, which is deep enough that you're most often blocked by instructions you can't re-order around well before you reach the full depth of that buffer. And Intel's branch prediction is one of the best in the industry, so there's not much room left to improve performance by eliminating bad branches. Fast radix division? Already done. Low latency caches and basic instructions? Done. Decoding x86 into swiss-cheese and reassembling it into highly tuned macro ops? Been doing that for years. Meanwhile the damn backend is already 8 ports wide.
At the end of the day Skylake is already taking virtually every opportunity to increase IPC, and for most tasks they're likely close to the theoretical maximum IPC (in an information science sense) for those tasks. So what's Intel to do? It's one thing to say that Intel "needs a new microarchitecture", but what would that microarchitecture do differently from Skylake?
All that's really left for Intel right now is low-level brute forcing. This means tweaking data paths to avoid the already small numbers of times where the processor is bottlenecked by those paths waiting for data, implementing larger L3/L4 caches, etc. Otherwise the only path not yet explored is extreme speculative execution in the form of eager execution, which is massively wasteful from an energy efficiency standpoint, and still not all that great because branch prediction is so good.