Friday, April 26th 2019
Intel 10nm Ice Lake to Quantitatively Debut Within 2019
Intel put out interesting details about its upcoming 10 nanometer "Ice Lake" CPU microarchitecture rollout in its recent quarterly financial results call. The company has started qualification of its 10 nm "Ice Lake" processors. This involves sending engineering samples to OEMs, system integrators and other relevant industry partners, and getting the chips approved for their future product designs. The first implementation of "Ice Lake" will not be a desktop processor, but rather a low-power mobile SoC designed for ultraportables, codenamed "Ice Lake-U." This SoC packs a 4-core/8-thread CPU based on the "Sunny Cove" core design, and Gen11 GT2 integrated graphics with 64 execution units and nearly 1 TFLOP/s compute power. This SoC will also support WiFi 6 and LPDDR4X memory.
Intel CEO Bob Swan also remarked that the company has doubled its 10 nm yield expectations. "On the [10 nm] process technology front, our teams executed well in Q1 and our velocity is increasing," he said, adding "We remain on track to have volume client systems on shelves for the holiday selling season. And over the past four months, the organization drove a nearly 2X improvement in the rate at which 10nm products move through our factories." Intel is prioritizing enterprise over desktop, as "Ice Lake-U" will be followed by "Ice Lake-SP" Xeon rollout in 2020. There was no mention of desktop implementations such as "Ice Lake-S." Intel is rumored to be preparing a stopgap microarchitecture for the desktop platform to compete with AMD "Matisse" Zen 2 AM4 processors, codenamed "Comet Lake." This is essentially a Skylake 10-core die fabbed on existing 14 nm++ node. AMD in its CES keynote announced an achievement of per-core performance parity with Intel, so it could be interesting to see how Intel hopes 10 "Skylake" cores match up to 12-16 "Zen 2" cores.
Source:
AnandTech
Intel CEO Bob Swan also remarked that the company has doubled its 10 nm yield expectations. "On the [10 nm] process technology front, our teams executed well in Q1 and our velocity is increasing," he said, adding "We remain on track to have volume client systems on shelves for the holiday selling season. And over the past four months, the organization drove a nearly 2X improvement in the rate at which 10nm products move through our factories." Intel is prioritizing enterprise over desktop, as "Ice Lake-U" will be followed by "Ice Lake-SP" Xeon rollout in 2020. There was no mention of desktop implementations such as "Ice Lake-S." Intel is rumored to be preparing a stopgap microarchitecture for the desktop platform to compete with AMD "Matisse" Zen 2 AM4 processors, codenamed "Comet Lake." This is essentially a Skylake 10-core die fabbed on existing 14 nm++ node. AMD in its CES keynote announced an achievement of per-core performance parity with Intel, so it could be interesting to see how Intel hopes 10 "Skylake" cores match up to 12-16 "Zen 2" cores.
46 Comments on Intel 10nm Ice Lake to Quantitatively Debut Within 2019
It tells us 10nm is good enough for one of the most important CPU lines Intel makes. The next step (and final confirmation that all is well) will be when (if) Intel puts it in Xeon Platinum. -U SoCs are expensive anyway. Even a mainstream i5-8250U (4C/8T) costs $297. Top of the range i7-8665U costs $409.
But you buy a laptop, not a SoC. You never really know how it impacts the final price.
Now, the supply of these CPUs is still a mystery, so lets just wait for that.
If you cannot deliver higher performance, you are just better to stay on 14nm. That's why Intel's 10nm is a failure judging from present data (lower frequency, worse thermals).
That's why Globalfoundries just abandoned the plans to upgrade to 7nm. No one will ask GF to manufacturer chips if GF could not bring up a competitive node. From what Intel presented recently, they don't have a new faster architecture. The so-called Sunny Cove is just a tweaked Skylake with minor updates to architecture. The IPC gain is likely less than 10%.
Unless Intel can magically clock their 10nm parts to 5.5GHz (very unlikely), I don't think their lead on single-thread performance is large enough to warrant the disadvantages on thermals and multi-core performance.
AMD said HBM was gonna change everything how many years ago ? .... and here we are in 2019 and it's and GDDR is still crushing it ... it's been a non-ffactor. Why care if a CPU has 12, 18, 24, 32 cores, if the competition's 6 core completes what ya wanna do faster, it don't matter. Why care if it's 7, 10, 12 nm, if the competition's unlucky 13 nm CPU completes what ya wanna do faster, it don't matter. Focusing on the technology before it's advantages have actually been demonstarted is fun perhaps, but little else.
Sunny Cove is the largest overhaul in many years, and is larger in scope than both Skylake and Haswell. Performance gains remains to be seen, but it does change/improve almost everything across the design, including major overhauls of cache (capacity and bandwidth), more than doubling of int mul/div performance, improved memory address calculation, doubling of load/store bandwidth and more. AMD recently stated that they expect clock speeds to decrease over the next years. We are probably at or near the peak of clock speed on the current type of semiconductor materials, so don't expect to see much in the 5-6-7 GHz range. Both Coffee Lake and Zen+ are pushing their respective chips into throttle territory, and it should be obvious that there are not huge gains to be expected in the future. I'm more interested in achieving good base clocks across many cores rather than max boost.
The good thing about IPC changes, is that even a modest IPC boost with a modest increase in die size, is likely to outperform the previews architecture while operating at lower clocks and achieve better thermals. And when it comes to multicore scaling, single core performance is actually very important, as most non-server workloads are synchronous, fewer faster cores is always going to outperform many slower cores. And going into the next decade, IPC is only going to get more important.
So far what we have is a crapload of powerpoint slides and vague marketing terminology, alongside the occasional stab at the competition. What we don't have is Intel showing to be proud of what's on the horizon. We're seeing damage control here.
So far Intel hasn't managed a 'major' IPC boost since Haswell, really. They've been lots of baby steps and Skylake wasn't all that earth shattering either besides the move to DDR4. The biggest gain in Core since Sandy Bridge has been that we got our high clocks 'back' after all those IPC bumps and being 'stuck' at 4.2~4.6 Ghz for years (Best-case). That is also the only thing giving Intel the lead for many a use case.
With limited production capacity, especially for higher clocked chips, it's no point in wasting that on mainstream mid-range and low-end CPUs. With a hypothetical 5% IPC gain across the board, they could easily cut 200 MHz max boost and still have higher performance, just as an example.
images.anandtech.com/doci/13699/Ronak28.jpg
To summarize,
1. Increase buffer sizes
2. 50% increase in L1 data cache
3. Larger L2 cache size
4. Larger microops cache
5. Larger 2nd level TLB
6. 2 more execution ports (from 8 to 10)
7. Double store data width
8. Execution port more capable
Haswell improvements
Exists: 1, 5, 6 (from 6 to 8), 8, double cache bandwidth, double load / store bandwidth
Skylake improvements
Exists: 1
Is Haswell much better than IvyBridge? I don't think so.
Is Sunny Cove introducing more improvements than Haswell? I don't think so too.
The improvements I talked of was of Sunny Cove vs. Skylake.
Here is another image to compare with Skylake:
images.anandtech.com/doci/13405/Ronak20_575px.jpg
As you can see, it replaces one Mul unit with two Mul units and a dedicated Div unit which will more than double integer mul/div performance.
Sunny Cove expands/extends pretty much every aspect except int ALUs (which they increased to 4 in Haswell), and vector units is only marginally improved over Skylake-X/SP, but significantly improved over Skylake-S of course.
If you know what the ports means.
There are also at few other typos or misplacements in here, so I wold take some of these details with a grain of salt.
I suppose they are just giving up on straight wins on desktop. They are basically following the same strategy AMD has in number of occasions - screw the efficiency as long as performance is there. Whether this will work in this day and age remains to be seen.