Examining Intel's Ice Lake Processors: Taking a Bite of the Sunny Cove Microarchitecture
by Dr. Ian Cutress on July 30, 2019 9:30 AM EST- Posted in
- CPUs
- Intel
- 10nm
- Microarchitecture
- Ice Lake
- Project Athena
- Sunny Cove
- Gen11
Performance Claims:
+18% IPC vs. Skylake,
+47% Performance vs. Broadwell
With every new product generation, the company releasing the product has to put some level of expectations on performance. Depending on the company, you’ll either get a high level number summarizing performance, or you’ll get reams and reams of benchmark data. Intel did both, especially with a headline ‘+18%’ value, but in recent months the company has also been on a charge about what sort of benchmarking is worth doing. I want to take a quick diversion down that road, and give my thoughts on the matter.
First, I want to define some terms, just so we’re all on the same page.
- A synthetic test is a benchmark engineered to probe a feature of the processor, often to find its peak capability in one or several specific task. A synthetic test does not often reflect a real-world scenario, and likely doesn’t use real world software. Synthetic benchmarks are designed to be stable and repeatable, and the analysis often describing how a processor performs in an ideal scenario.
- A real-world test uses software that the user ends up using, along with a representative workload for that software. These tests are usually most applicable to end-users looking to purchase a product, as they can see actual use-case results. Real-world tests can have obvious pitfalls: it can be hard to test across multiple machines with only a single license, and testing one piece of software has no guarantee on performance on another.
A typical analysis of a processor does two things: what can it do (synthetic) and how does it perform (real-world). Users interested in the development of a platform, how it will expand and grow, or engineers peering over the fence, or even investors looking at the direction the company is going, will look at what products can do. People looking at what to use, what to work with, are more interested in the performance. Reviewers should get this concept, and companies like Intel should get this too – with Intel hiring a number of ex-reviewers of late, this is coming through.
A couple of months ago, Intel approached subsets of reviewers to discuss best benchmarking practices. On the table were real-world benchmarks, and which benchmarks represent the widest array of the market. Under fire was Cinebench, a semi-synthetic test (it uses a real-world engine on example data) that Intel believed didn’t represent the performance of a processor.
Intel provided data from one of its commissioned surveys on software that people use. Their data was based on a list of all consumers, from entry-level users up to prosumers, casual gamers, and enthusiasts, but also covering commercial use cases. At the top of the list were the obvious examples, such as OS and browsers: Explorer.exe, Edge, Chrome. In the top set were important widely distributed software packages, such as Photoshop (all versions), Steam, WinRAR, Office programs, and popular games like Overwatch. The point Intel was trying to make with this list is that a lot of reviewers run software that isn’t popular, and should aim to cover the widest market as possible.
The key point they were trying to make was that Cinebench, while based on Cinema4D and a rendering tool used by a number of the community, wasn’t the be-all and end-all of performance. Now this is where Intel’s explanation became bifurcated: despite this being a discussion on what benchmarks reviewers should consider using, Intel’s perspective was that citing a single number, as Intel’s competitors have done, doesn’t represent true performance in all use cases. There was a general feeling that users were taking single numbers like this and jumping to conclusions. So despite the fact that the media in the room all test multiple software angles, Intel was clear in that they didn’t want a single number to dominate the headlines, especially when it’s from software that is ranked (according to Intel’s survey) somewhere in the 1400s.
Needless to say, Intel got a bit of backlash from the press in the room at the time. Key criticisms were that those present, when they get hardware, test a variety of software, not just Cinebench, to try and give a more overall view. Other key elements included that the survey covered all users, from consumer, commercial, and workstation: a number of the press in the room have audiences that are enthusiasts, so they will cater their benchmark accordingly. There was also a discussion that a number of software packages listed in the top 100 are actually difficult to benchmark, due to licensing arrangements designed to stop repeated installs across multiple systems. Typically most software vendors aren’t interested in working with the benchmark community to help evaluate performance, in the event that it exposes deficiencies in their code base. There was also the way in that readers were adapting over time: most focused readers want their specific software tested, and it is impossible to test 50 different software packages, so a few that can be streamlined in a benchmark suite are used as a representative sample, and typically Cinebench is one of those in the rendering arena, alongside POV-Ray, Corona, etc.
Intel, at this stage in the discussion, still went on to show how the new hardware performs on a variety of tests. We’ve covered these images before on previous pages, but Intel stated a significant uplift in graphics compared to the current 14nm offerings, from 40% up to 108%:
As well as comparisons to the competition:
Aside from 3DMark, these are all ‘real-world’ tests.
Move forward a few weeks, and Intel’s Tech Day where Ice Lake is discussed, and Intel brings up IPC.
Intel’s big statement is that Sunny Cove, a 2019 product, offers 18% more instructions per clock against Skylake, a 2015 product. In order to come to that conclusion, as expected, Intel has to turn to synthetic testing: SPEC2006, SPEC2017, SYSMark 2014 SE, WebXPRT, and Cinebench R15. Wait, what was that last one? Cinebench?
So there are two topics to discuss here.
First is the 18% increase over four years – that’s the equivalent to a 4.2% compound annual growth rate. Some users will state that we should have had more, and that Intel’s issues with its 10nm manufacturing process means that this should have been a 2017 product (which would have been an 8.6% CAGR). Ultimately Intel built enough of an IPC increase lead over the last decade to afford something like this, and it shows that there isn’t an IPC wall just yet.
Second is the use of Cinebench, and the previous version at that. Given what was discussed above, various conclusions could be drawn. I’ll leave those up to you. Personally, I wouldn’t have included it.
Aside from IPC, Intel also spoke about actual single-threaded performance about Sunny Cove in its 15W mode.
At a brief glance, I would have expected this graph to be from real-world analysis. But given the blurb at the bottom it shows that these results are derived from SPEC2006, specifically 1-thread int_rate_base, which means that these are synthetic results, so we’ll analyze them with that in mind. This test also gets lots of benefit from turbo, with each test likely to fit inside the turbo window of an adequately cooled system.
The base line here is Broadwell, Intel’s 5th Generation processor, which if you remember was the first Intel processor to have an integrated FIVR on the mobile parts for power efficiency. In this case we see that Intel puts Skylake as +9% above Broadwell, then moving through Kaby Lake and Whiskey Lake we see the effect of increasing that peak turbo frequency and power budget: when we moved from dual core to quad core 15W mobile processors, that peak turbo power budget increased from 19W to 44W, allowing longer turbo. Overall we hit +42% for 8th Gen Whiskey Lake over Broadwell.
Ice Lake, by comparison, is +47% over Broadwell. When moving from Broadwell to Ice Lake, which Intel expects most of its users to do, that’s a sizable single threaded performance jump, I won’t dispute that, although I will wait until we see real world data to come to a better conclusion.
However, if we compare Ice Lake to Whiskey Lake, we see only a +3.5% increase in single threaded performance. For a generation-on-generation increase, that’s even lower than the four-year CAGR from Skylake. Some of you might be questioning why this is happening, and it all comes down to frequency.
Intel’s current 8th Gen Whiskey Lake, the i7-8565U, has a peak turbo frequency of 4.8 GHz. In 15W mode, we understand that the peak frequency of Ice Lake is under 4.0 GHz, essentially handing Whiskey Lake a ~20% frequency advantage.
If this sounds odd, turn over to the next page. Intel is going to start tripping over itself with its new product lines, and we’ll do the math.
107 Comments
View All Comments
name99 - Wednesday, July 31, 2019 - link
That’s an idiotic chain of reasoning.ARM Macs will ship with macOS, not iOS. To believe otherwise only reveals that you know absolutely nothing of how Apple thinks.
As for comparison, the rough number is A12X gets ~5200 on GB4, Intel best (non-OC’d) gets ~5800. That’s collapsing lots of numbers down to one, but comparing benchmark by benchmark you see Apple does very well (almost matching Intel) across an awful lot.
If Apple can maintain its past pace (and there is no reason why not...) we can expect A13X to be anywhere from 20% to 35% faster, which puts it well into “fastest [non-OC’d] CPU on earth” territory for most single-threaded use cases. Can they achieve this? Absolutely.
Just process improvement can get them 10% frequency. I expect A13X to clock around 2.8GHz.
Then there is LPDDR5 which I expect they will be using, so substantially improved memory bandwidth. Then I expect they'll have SVE (2x256) and accompanying that basically double the bandwidth all the way out from L1 to DRAM.
These are just the obvious basics. There are a bunch of things they can still do that represent “fairly easy” improvements to get to that 25% or so. (These include more aggressive fusion, a double-pumped ALU, attached ALUs to load/store to allow load-ok and op-store fusion, a micro-op cache, long-term-parking, criticality prediction, ...)
So, if it’s so easy, why doesn’t Intel also do it? Why indeed? That’s why I occasionally post my alternative rant about how INTC is no longer an engineering company, it is now pretty much purely a finance company...
ifThenError - Friday, August 2, 2019 - link
Sorry, but both these comments seem mighty uninformed. The MacBooks Air and Pro currently and in the foreseeable future all run on Intel CPUs. The Apple Chips A12/13 are used in iPhone, iPad and the likes.And regarding your prediction, your enthusiasm seems way over the top. What are you even talking about? Micro-op cache on a RISC processor? Think again. Aren't RISC commands all micro ops already?
name99 - Sunday, August 4, 2019 - link
Strong the Dunning-Kruger is with this one...Dude, seriously, learn something about MODERN CPU design, more than just buzz-words from the 80s.
To get you started, how about you read
https://www.anandtech.com/show/14384/arm-announces...
and concentrate on understanding EVERY aspect of what's being added to the CPU and why.
Note in particular that 1.5K Mop cache...
More questions to ask yourself:
- Why was 80s RISC obsessed with REDUCED instructions?
- Why was ARM (especially ARMv8) NOT obsessed with that? Look at the difference between ARMv8 and, say, RISC-V.
- Why is op-fusion so important a part of modern high performance CPUs (both x86 and ARM [and presumably RISC-V if they EVER ship a high-performance part, ha...])?
- which are the fast (shallow logic, even if it's wide) and which are the slow (deep logic) parts of a MODERN pipeline?
ifThenError - Monday, August 5, 2019 - link
Oh my, this is so entertaining you should charge for the reading.You demand to go beyond just buzz words (what would be good) while your posts look like entries to a contest on how many marketing phrases can be fit into a paragraph.
Then you even manage to combine this with highly rude idiom. Plus you name a psychological effect but fail to transfer it to self-reflexion. And as cherry on the top you obviously claim for yourself to understand „EVERY aspect“ of a CPU (an unimaginably complex bit of engineering) but even manage to confuse micro- and macro-op cache and the conceptual differences of these.
I'm really impressed by your courage. Publicly posting so boldly on such a thin basis is brave.
Your comments add near zero information but are definately worth the read. Pure comedy gold!
Please see this as an invitation to reply. I'm looking forwards to some more of your attempts to insult.
Techgeek43 - Tuesday, July 30, 2019 - link
Fantastic article Ian, I for one, cannot wait for ice lake laptopsWonderful in-depth analysis, with an interesting insight into the Intel brand
repoman27 - Tuesday, July 30, 2019 - link
"The high-end design with 64 execution units will be called Iris Plus, but there will be a ‘UHD’ version for mid-range and low-end parts, however Intel has not stated how many execution units these parts will have."Ah, but they have: Ice Lake-U Iris Plus (48EU, 64EU) 15 W, Ice Lake-U UHD (32EU) 15 W. So their performance comparisons may even be to the 15 W Iris Plus with 64 EUs, rather than the full fat 28 W version.
I know you have access to the media slide decks, but Intel has also posted product briefs for the general public that contain a lot of this info: https://www.intel.com/content/www/us/en/products/d...
"On display pipes, Gen11 has access to three 4K pipes split between DP1.4 HBR3 and HDMI 2.0b. There is also support for 2x 5K60 or 1x 4K120 with a 10-bit color depth."
The three display pipes are not limited to 4K, and are agnostic of transport protocol—each of them can be output via the eDP 1.4b port, one of the 3 DDI interfaces which can support either DisplayPort 1.4 or HDMI 2.0b, or one of the up to 4 Thunderbolt 3 ports. Both HDMI and DP support HDCP 2.2, and DisplayPort also supports DSC 1.1. The maximum single pipe, single port resolution for HDMI is 4K60 10bpc (4:2:2), and for DisplayPort it's 4K120/5K60 10bpc (with DSC).
Thunderbolt 3 integration for Ice Lake-Y is only up to 3 ports.
abufrejoval - Tuesday, July 30, 2019 - link
What I personally liked most about the GT3e (48 EU) and GT4e (72 EU) Skylake variant SoCs was, that they didn't cost the extra money they should have, especially when you consider that the iGPU part completely dwarfs the CPU cores (which Intel makes you bleed for) and is much better than everything else combined together (have a look at the WikiChips layoutshttps://en.wikichip.org/wiki/intel/microarchitectu...
Of course, a significantly better graphics performance is never a bad thing, especially when it also doesn't cost extra electrical power: The bigger iGPUs might have actually been more energy efficient than their GT2 brethren at a graphics load that pushed the GT2 towards its frequency limits. And in any case if you don't crunch it on graphics, the idle consumption is near perfect: One of the reasons most laptop dGPU designs won't even bother to run 2D on the dGPU any more but leave that to Intel.
The biggest downside was that you couldn't buy them outside an Apple laptop or Intel NUC.
But however much Intel goes into Apple mode (the major customer for these beefier iGPUs) in terms of "x time faster than previous", the result aren't going to turn ultrabooks with this configuration into "THD gaming machines".
To have a good feel as to where these could go and whether they are worth the wait, just have a look at the Skull Canyon nuc6i7kyk review on this site: That SoC uses 72 EUs and 128MB of eDRAM and should put a pretty firm upper limit to what a 64 EU Ice Lake can do: Most of the games in that review are somewhat dated yet fail to reach 20FPS at THD.
So if you want to game on the device, you'd be much better of with a dGPU however small and chose the smallest iGPU variant available. No reason to wait, Whisky + Nvidia will do better.
If you want real gaming performance, you need to put real triple digit Watts and the bandwidth only GDDR5/6 or HBM can deliver to work even at THD, but with remote gaming perhaps it doesn't have to be on your elegant slim ultrabook. There again anything but the GT2 configuration is wasted, because only need the VPU part for decoding Google Stadia (or Steam Remote) streams, which is the same for all configurations.
For some strange reason, Intel has been selling GT3/4 NUCs at little or no premium over GT2 variants and in that case I have been seriously tempted. And only once I even managed to find a GT3e laptop once for a GT2 price (while the SoC is literally twice as big and the die carrier even adds eDRAM at zero markup), which I stil cherish.
But if prices are anywhere related to the surface area of the chip (as they are for the server parts), these high powered GTs are something that only Apple users would buy.
That's another reaons, I (sadly) don't expect them to be sold in anything bug Macs and some NUCs, no ChuWi notebooks or Mini-ITX boards.
abufrejoval - Tuesday, July 30, 2019 - link
...(need edit)Judging from the first 10nm generation, GPUs where the part where obtaining economically feasible yields didn't work out. Unless they have really, really fixed 10nm it's not hard to imagine that Intel could be selling high-count EU SoCs to Apple below cost, to keep them for another generation as flagship customer and perhaps due to long-term contractual obligations.
But maintaining GT2/3/4 price egality for the rest of the market seems suicidal even if you have a fab lead.
Not that I expect we'll ever be told: In near monopoly situations the so called market ecnomy becomes surprisingly complex.
willis936 - Wednesday, July 31, 2019 - link
What the hell is a THD in this context?jospoortvliet - Monday, August 5, 2019 - link
Probably full HD (True HD)?