October 2018 - Intel has finally launched its long-awaited Ice Lake architecture, the second processor generation manufactured on the company's 10nm node following Cannon Lake (CNL). Ice Lake (ICL) is the next step in Intel’s Tick-Tock product cadence and the first new architecture on the 10nm(+) process node (“Tock”).
Cannon Lake introduced the AVX-512 instruction set to the mainstream, as well as featuring an improved Gen10 IGP doubling the execution units. Being a “Tick” the underlying tech is mostly a slightly improved Skylake (SKL) architecture. After all, the main purpose was to iron out the rough edges of the new 10nm process before introducing a larger architectural improvement while reducing the die size and power consumption over previous 14nm products. Ice Lake, being the “Tock”, brings much larger changes to the table, like larger caches, further improved IPC, small additions to the AVX-512 instruction-set but, most importantly, up to 8 cores with turbo clock speeds up to 5 GHz. All this within a reasonable 95W tTDP (typical TDP) and 160W pTDP (peak TDP) under AVX-heavy loads.
If you're wondering what the hell you're reading here, as this is obviously not reality, this is what could (and probably should) have been the introduction to the product launch that Intel anticipated for 2018 just 2-3 years ago. Instead, Intel has just launched yet another iteration of the Skylake architecture from 2015, the 8-core "Skylake Coffee Lake", following its 6-core sibling from last year. In essence, this is really just the Skylake architecture we all know since 2015, just with double the number of cores added to the ring bus. The IGP saw a small update with Kaby Lake, supporting newer video formats, but is in most aspects still the same as in 2015. This means it is still lacking native HDMI 2.0 support and its relative performance has become even poorer in comparison to IGP products from AMD.
Let’s be fair though, even 3 years after its introduction the Skylake architecture is still a strong performer, regardless of the numerous variants of security flaws named Spectre and Meltdown that introduced various performance regressions. While AMD has a great product with Ryzen in terms of IPC, even the 2nd generation fails to trump Skylake, which is amplified by the fact that Intel’s products can reach higher clock speeds. Those higher clocks come at the cost of a significantly higher power draw at the upper end though.
It’s rather impressive how far Intel could improve their 14nm process since the first Broadwell processors in 2014, which were far from being trouble-free in the beginning either. Obviously they never planned to be on 14nm with their core products for 5-6 years but Intel has made the best of it. With the latest Core i9-9900K Intel is going all-in on its (hopefully) last 14nm based mainstream flagship to make sure that its latest introduction can proudly claim the mainstream performance crown for the next few months; which it does. Slightly higher IPC in most cases, higher AVX throughput and higher clock speeds are more than enough to keep AMD’s best mainstream entry, the Ryzen 7 2700X, in check. While there are rumors over a slightly faster Ryzen 7 2800X, a look at the overclocking results of the 2700X shows that there is not that much headroom left for a 2800X. This is even more true when we take into consideration that the best dies are reserved for the more expensive Ryzen Threadripper models. The real competitor for the 9900K is hopefully the next generation Zen2 architecture, manufactured at TSMC in their latest 7nm node.
8 core rivalry: Intel Core i9-9900K and its direct competitor, AMD Ryzen 7 2700X
Going full throttle in a car goes hand in hand with increased fuel consumption. The same is true for the i9-9900K and to a slightly lesser extent for the i7-9700K as well. These chips are pushed to the max, way beyond the sweet spot that manufacturers usually aim for in order to keep the power and cooling requirements in check. This is not necessarily new, both Intel and AMD have done this multiple times in the past (e.g. Pentium III 1.13GHz, Pentium 4 (Prescott), Athlon64 X2 6400+, FX-9590) when they needed to squeeze out the last bit of performance for whatever reason. Needless to say, those chips were usually not particularly great overclockers either, since they were by default closer to the limit that could be safely reached with conventional cooling.
In the last couple of years Intel didn’t have to push that hard since competition from AMD was rather weak. Intel could play it safe with lots of margin left in case they would need to push, saving on manufacturing (thermal paste instead of solder, less aggressive binning) and still making good money with pricier K-series models that allowed enthusiasts to find out what the chips could really do when paired with more expensive Z-chipset motherboards. Now, with Intel being stuck on 14nm for at least another year and faced with an unexpectedly agile product lineup from AMD, it is again forced to go full throttle.
The cost is a power draw under heavy loads we have only seen before with the lackluster AMD FX-9590, both reaching 220W in real world applications, when well implemented AVX2/FMA3 code is used. Without doubt that is a lot to ask for on mainstream platforms, especially considering that a high-end 16-core Threadripper comes with a real CPU power draw of just 180-200W. On the upside, the Core i9-9900K does indeed deliver when it comes to performance, while the FX-9590 did not.
Temperatures and Power Consumption
While such a high power draw is not great, this would be perfectly fine if it would be openly communicated by Intel. Instead Intel is still stating a TDP of 95W, which is marketed everywhere and often believed to be the maximum power draw by many non-enthusiasts. Of course, this plays into Intel’s hands and it is highly unlikely that it is a simple oversight that the real peak power values are not stated by a second TDP value.
Yes, somewhere in the fine print one can find the remark that the TDP is only valid for the base clock, which is usually much lower than the stated Turbo speeds. Yes, there is pretty decent documentation available in non-public areas at Intel, describing how the power limits really work and how they should be used. Good luck finding that information at price comparison sites, in shops or in the information sheet of your newly bought system. To be fair, AMD’s Ryzen processors with the X suffix will exceed their TDP by 10-20% as well when using the XFR Performance Boost. It's an entire different ballpark, though, when Intel's 9th gen 8-cores exceed their specified TDP by up to 130%.
Due to the huge power draw and cooling requirements of those CPUs the motherboard vendors had to make crucial design choices for their existing Z370 models and the newly launched Z390 variants. Up to the i7-8700K 6-core it was perfectly fine to let the CPU take whatever wattage it needs and thus profit from a few extra percent of performance by going over the stated TDP specification value. With the arrival of highly clocked mainstream 8-cores vendors couldn't go down that path anymore — or at least they shouldn't. Even if the VRM-design can handle the burden, which is definitely not the case for most Z370 and low-end Z390 boards, let alone cheaper H370, B360 or H310 models, the CPU cores will go easily above Intel's recommended 80°C max on every air cooler out there. So out of necessity desktop PCs get introduced to a concept that is only too well-known from mobile devices: power limits. These limits put a restriction on the time the CPU is allowed to exceed a certain power limit before the (turbo) clock must be reduced to a dynamic value that allows the CPU to stay within the limit. The peak wattage is limited by Power Limit 2, also called "Short Duration Package Power Limit", while prolonged load is capped by the Power Limit 1 or "Long Duration Package Power Limit" in combination with a time value called Tau or "Package Power Time Window". Additionally an overall Current Limit called IccMax is applied to various parts of the CPU. This is were things get interesting: Motherboard makers are not obligated to follow any of Intel's recommendations for these values, in fact Power Limit 2 is officially recommended to be chosen appropriately for the motherboard design (maximum capabilities of the VRMs). On the other hand, using a cheaper motherboard that sticks to the recommended Intel settings you will not get the performance indicated by many reviews.
With all that said you can already guess the consequences of these design choices. There is a multitude of different settings out there, which are perceived or treated as being "stock" when in reality they are not. It is quite telling that even reviewers failed to mention these inconsistencies in their benchmark results. It's even worse when they don't document their test setup properly, let alone discuss Power and Current Limits so readers are aware of the problem at hand. In the end there are many reviews out there with a spread as wide as 15% on various benchmarks, all advertised as Intel's Stock Profile. It's a mess. That's why we jumped in at the deep end and found out what Power and Current Limits are about. You'll find it here: A Messy Situation: Intel's Stock Profile and Power Limits
As stated in the beginning of this chapter, all this would be fine and is indeed necessary, but it's not openly communicated by all parties involved. Some motherboard vendors even fail to address the issue with BIOS updates for 9th gen compatibility on their existing Z370 models. Using the top models i9-9900K and i7-9700K this results in overloaded VRM designs, VRM-temperature induced down-clocking or even sudden shutdowns; just because of unrestricted Power Limits. The GIGABYTE Z370 AORUS ULTRA GAMING WIFI is a perfect example for the worst case scenario:
The GIGABYTE Z370 AORUS ULTRA GAMING WIFI can't handle an i9-9900K on default settings. Even with MCE disabled (all cores on max turbo ratio, which is the default setting as well on this board).
As explained in the last chapter it's more important than ever before to specify the testing parameters for benchmarks. Please have a good look at the following profiles to understand what has actually been tested.
9900K @Stock: PL1 95W, PL2 210W, IccMax 193A: The Stock Profile for ASUS mainboards like the MAXIMUS XI EXTREME.
9900K @5 GHz: PL1, PL2 and IccMax unlimited, "Multicore Enhancement" disabled. This is an overclocking profile that can only be achieved with a good mainboard and CPU cooler.
2700X @XFR2: XFR2 enabled, Performance Enhancer: Level 1 with clock frequencies of up to 4.36 on a single core, 4.16 GHz on all cores. This is very close to the maximum overclock (4.2 on all cores).
GPUPI for CPU 100M
GPUPI for CPU 1B
3DMark - Timespy (CPU only)
Fire Strike Extreme
1920x1080, DirectX 11
Grand Theft Auto V
1920x1080, OpenGL 4.5
The Witcher 3
Far Cry 5
To show the performance impact of streaming a game with our test candidates, we used the popular OBS Studio to stream a 60 FPS live video to Twitch in the same resolution we benched the game in.
If you don’t mind having a 180-225W beast in your system, want the best single-core and gaming-performance you can get for (somewhat) reasonable money, then the new Core i9-9900K and Core i7-9700K models are definitely worth a look. The same goes for consumers that need strong AVX/FMA3 instruction performance for tasks like media encoding or 3D rendering on the CPU. To get the most out of the top-end 9th generation Core processors a higher-end motherboard with a strong VRM is a must, so don’t expect to be able to go cheap here or reuse lower-end Z370 boards. A top-end cooling solution and an airflow-friendly chassis with proper fans are recommended as well, especially when the system is going to be paired with a high-end video card.
Total Points of 57 results
i9-9900K @5 GHz
R7 2700X @XFR2
The total points are calculated as a sum of all benchmark results. The product with the best result is awarded 100 points, while a product exactly half as good will receive only 50 points. The achieved points in each benchmark are therefor always in relation to the best result.
Overall the Intel Core i9-9900K is unbeatable in terms of performance
However, if AVX performance is not that important or you don’t need the last 5-10% when it comes to gaming (the video card is often the bottleneck anyway) then the Ryzen 2700X and its smaller siblings are currently the more attractive choice. As a bonus, they need less attention when it comes to cooling, airflow and choice of motherboard. This is even more true when one considers the current supply constraints at Intel, which cause the new models to sell for inflated prices, if you can even get them.
A Ryzen 7 2700X is currently available for about €320, whereas the Core i9-9900K is listed with about €600 while the i7-9700K without Hyperthreading comes in at €490. The performance difference does not warrant such a big gap and from a rational perspective, when you are a gamer, it is wiser to go for the Ryzen 7 2700X and spend more money on a video card than the other way round.
R7 2700X @XFR2
i9-9900K @5 GHz
The price of each product is taken from best listed price on geizhals.at, an Austrian online price comparison platform. For products no longer available on the market we will choose an appropriate second hand pricing found on ebay and other sales sites. Time: 25.11.2018
The strength of Ryzen: For each performance point achieved in our benchmarks, you have to pay double if you go with Intel!
In case you have an application that scales well with the number of available cores, you should have a look at the 12-core (1920X, 2920X) and 16-core (1950X, 2950X) Ryzen Threadripper models. The power draw at AVX/FMA3 loads is lower or on par and the higher number of cores might be more than able to compensate the 9900K's AVX advantage and higher core clocks.
To summarize: As of November 2018 the Core i9-9900K will get you the best mainstream performance currently available but at the cost of a very high power draw, higher component cost and more effort to keep things silent under full load. Unless you really need exactly what this CPU has to offer, there are cheaper options on the market.
Intel’s accomplishments with the 8-core Coffee Lake chip aside, let’s not forget that it was AMD with Ryzen that forced Intel to launch the Coffee Lake 6-core about 10 to 11 months earlier than originally planned. Now Intel's hand was forced again due to cheap 8-core Ryzen processors and the ongoing struggle with the 10nm process.
AMD disrupted other parts of Intel’s plans and roadmaps too, with the initial Ryzen Threadripper launch driving Intel to introduce 12-18 core parts to the LGA2066 platform, using the Skylake-SP HCC die, which was originally not intended for anything apart from high-margin Xeon models. The impending launch of 32- and 24-core 2nd generation Ryzen Threadripper models shook up Intel’s roadmap even more, resulting in desperate measures like a water chilled demo of the highest-end Skylake-SP XCC based 28-core Xeon at Computex 2018. This LGA3647 based product will make it to the market as Intel’s super high-end platform, sitting above the current LGA2066 platform — a direct response to the 32-core Threadripper model from AMD.
Closing the circle to the alternate reality introduction, if Intel would have been able to execute their roadmaps as originally planned, or at least with lesser disruptions, AMD’s Ryzen series would now face a much stiffer competition and probably would not be seen as favorable as it has been since its introduction one and a half years ago. Yes, AMD would not be as far behind as they have been with their ill-fated Bulldozer series, but they would probably have a 10-15% IPC performance gap to the newly launched Ice Lake architecture with no sign of AVX-512 instructions for mainstream CPUs. By the time Ice Lake will be launched, which will probably happen either in late 2019 or in the first half of 2020, AMD has to improve their shortcomings against the current generation with their upcoming 7nm Zen2 based products. AMD claims a 10-15% IPC improvement over their current Zen(+) based products and probably even more for AVX workloads due to two 256-bit capable FPUs per core. The Chiplet approach should also help getting more consistent performance from future high-core Threadripper and EPYC parts. As it looks right now, AMD might tackle the right areas with their next iteration of Zen, which is good because it will face stiff competition once Intel gets more traction again.
Let’s not underestimate the blue giant: While Intel is currently struggling on various fronts, mostly because of their 10nm production woes, which caused most of the other issues they are currently having, they have the resources to weather the storm they are currently in. AMD has a good window of opportunity to snatch market share from Intel in 2019, be it in Desktop, Server and maybe even Mobile, as long as their 7nm Zen2 products remain on track. A slip into the second half of 2019 could already be a serious blow to AMD’s ambitions, especially in case early performance numbers indicate a serious performance uplift for Ice Lake and Intel might be able to introduce it earlier than expected.
Still, kudos to AMD for having such a strong product in 2017 and forcing Intel to react instead of setting the pace. Just as in 2003, when the Athlon 64 was introduced with many firsts, AMD has the right product at a time when Intel is struggling. From what we currently know, 2019 looks quite promising, maybe even allowing AMD to finally gain a noticeable market share in the Server market with EPYC. One can only hope that the manufacturers of Laptops will finally put some serious effort into products with AMD processors and not hamper them by cheap and/or intentionally, artificially crippled setups, as it is still the case with Raven Ridge based Laptops.
As we have seen in the last 18-21 months, we, the customers, benefit greatly from healthy competition with serious choices available. The computing power now available at mainstream prices is amazing and finally a real reason to upgrade from your old AMD FX or Intel 2nd, 3rd or 4th generation based system.