CES is always good for a peek at what companies are planning for the future, and the show in 2022 was no exception. Intel and AMD both updated their respective mobile roadmaps for 2022 last week. Apple hasn’t made any major announcements recently, but the company’s plan for ramping its M-class processors into desktops and laptops is a little easier to see now that multiple variants of the CPU are shipping.
This story will focus on the lower-power mobile market because the differences we’re going to discuss are more visible in the 9W – 15W TDP brackets. Intel’s 28W – 64W and especially its 45W – 115W mobile chips more closely resemble their desktop counterparts.
To quickly review: When ARM introduced big.Little, it claimed it would be more efficient to deploy two different CPU microarchitectures in the same physical silicon rather than running one core architecture at various clocks. Many mobile devices today use SoCs that combine a “big” core to deliver high performance and a “little” core to maximize power efficiency. Data is shared between the two core clusters and the OS is capable of scheduling work on both cores depending on the nature and urgency of the task at hand.
Ten years ago, Intel argued that it could avoid the need for big.Little by using Dynamic Frequency and Voltage Scaling (DFVS). With Alder Lake, Intel decided to move away from using just one CPU microarchitecture. Alder Lake uses two different microarchitectures in what Intel calls a “hybrid” core design.
This slide is from Intel’s Lakefield launch, but it still works as a conceptual example of what Alder Lake is intended to accomplish, namely: Achieving higher overall performance at higher power efficiency.
AMD has not launched a hybrid SoC or announced any plans to do so. Its APUs are heterogeneous with respect to CPU+GPU but homogeneous as far as CPU microarchitecture.
Power and Performance Segmentation: Apple Versus x86
Intel and AMD both measure thermal dissipation using a metric called TDP, or Thermal Design Power. TDP is not equivalent to total system power consumption. It refers to the amount of power a CPU is expected to dissipate over a period of time in a typical workload and is intended to help manufacturers design thermal solutions for laptops. Intel and AMD define their base TDPs at base clock, not boost clock, and CPUs are allowed to exceed their rated TDP for short periods of time when operating in turbo mode.
Manufacturers have substantial control over how the CPU in a given system behaves. One company might define a very low maximum skin temperature to guarantee a good user experience, even when this prevented the CPU from using turbo mode. Another might choose to emphasize maximum performance over temperatures or battery life. These differences are part of why laptops with identical CPUs sometimes perform quite differently.
While TDP is an imperfect metric, it does allow for a certain amount of high-level comparison.
Apple does not refer to CPU TDP in its documentation, so we’ll be examining actual power consumption for various M1 systems. Data from Anandtech suggests the M1 Max draws between 30-50W under heavy CPU load, while the M1 in a Mac Mini tops out around 21W. This compares reasonably well to Intel’s 15W – 55W and 28W – 64W TDP segmentations.
Data and chart by Anandtech. The last two rows also test GPU power consumption and are beyond the scope of this comparison.
We’re primarily focused on x86 CPUs with a baseline TDP of 28W or less. Intel’s 45W+ laptop CPUs more closely resemble their desktop counterparts more than the lower-power SKUs.
AMD: Staying the Course
AMD’s path to better mobile performance and increased competitiveness against Intel kicked off with the launch of Ryzen 2000 Mobile three years ago, but the company’s Ryzen 2000 and some 3000 Mobile SKUs still lagged against Intel’s best. Ryzen 4000 Mobile and Ryzen 5000 Mobile were more evenly matched. AMD is now a genuine competitor in mobile, not just desktop and server.
AMD’s 2022 product stack. Note that the majority of chips are either 8/16 or 6/12 APUs, whether at 45W or 15W.
AMD’s strategy for 2022 is to iterate and improve Ryzen, and the company’s market share gains over the past five years make that a hard strategy to argue with.
Discussion of Intel’s hybrid cores (exemplified by Lakefield and Alder Lake) drove some understandable curiosity about what AMD would do with the hybrid idea. Thus far, AMD’s stated opinion is that it believes Zen is flexible and capable enough to address both the low-power and high-power markets for x86. This is what we would expect AMD to say, given that it doesn’t have a hybrid solution ready to ship, but the Ryzen 5000 family is quite power efficient and the Ryzen 6000 Mobile family is expected to improve in that regard. The shift to 6nm in 2022 is not expected to change much on the power consumption side of things, but AMD is claiming some significant improvements thanks to silicon optimizations and refinements.
Clock speeds on chips like the 6800U (8C/16T, 2.7GHz base / 4.7GHz boost) compare well against the 5800U (8C/16T, 1.9GHz base, 4.5GHz boost). AMD’s TDPs are calculated against base clock, so the tick up from 1.9GHz to 2.7GHz should bode well for the 6000 family.
AMD does not segment its mobile stack as aggressively as Intel does. The vast majority of Ryzen 6000 Mobile APUs are either 8C/16T or 6C/12T. Eight-core CPUs can be purchased in 45W, 35W, 15-28W, and 15W power bands, with lower clocks on the lower TDP chips. As the chart above shows, the difference between the U-class CPUs and the H-class chips are the listed base and boost clocks. There is only minimal variation on core count.
Apple: Pump Up the (Big) Core Count
Apple currently sells three M1-based SoC’s — the M1 Pro, M1 Max, and the original M1. The original M1 is a symmetric design, with four “big” cores (FireStorm) and four “little” cores (IceStorm). Testing of devices like the M1-equipped Mac Mini suggests that total device power consumption doesn’t rise above 30W in this system during a CPU-centric workload. M1-equipped laptops should match or exceed this power efficiency.
The M1 Max’s CPU-centric testing shows wall active power consumption of as much as 62W, but this is likely caused by heavy memory accesses, according to Anandtech, rather than strictly by the CPU itself. (Intel and AMD’s TDP ratings do not account for RAM) Outside of that, total wall power tops out at ~50W.
The CPU differences between the M1 and the M1 Pro / M1 Max mostly boil down to core counts. Where the M1 is a 4+4 design, the M1 Pro and M1 Max are either 6+2 or 8+2. What’s noteworthy about Apple’s decision here is that it chose to add more “big” cores when it moved to the M1 Pro and M1 Max and to reduce the size of its “little” core cluster at the same time. Apple also runs its CPUs at much lower core clocks than Intel or AMD. The M1 Max reportedly clocks from 600MHz – 3.2GHz on its FireStorm cores and 600MHz – 2GHz on its IceStorm cores. CPUs tend to be more efficient at lower clock speeds, and this likely helps the M1’s relative efficiency.
Intel: Emphasizing Efficiency Cores
Intel’s approach to power management on Alder Lake depends on which TDP bracket one considers. Intel’s 9W and 15W chips are all about little cores. In the 9W segment, Intel’s Pentium 8500 and Celeron 7300 recall Lakefield, with just one high-performance core and four efficiency cores. Performance should be superior to Lakefield, however, thanks to the higher TDP (9W versus 4.5W) and IPC improvements to both CPU microarchitectures.
The higher-end Core CPUs opt for two performance cores and as many as eight efficiency cores. The primary difference between the Core i3-1210U (2P+4E) and the Core i5-1230U / Core i5-1240U is the addition of four more efficiency cores for a 2P+8E CPU.
Intel’s 15W-55W chips have the same core counts as its 9W-28W processors. They difference between the two families are their base frequencies, boost frequencies, and (probably) how long they hold their turbo clocks under load.
The 55W maximum TDP on these chips is close to the maximum wall power consumption Anandtech measured on the M1 Max, but with a reversed CPU configuration: 2+8 for Intel, as opposed to 8+2 for Apple. Keep in mind that all comparisons between measured wall power and TDP should be treated as approximates. If Intel specs a CPU for a 55W maximum TDP, it means the CPU is capable of dissipating up to 55W of heat for a specific period of time. It does not mean total system wall power will be no more than 55W. While the CPU’s power consumption under load is a significant percentage of the total as measured at the wall, there is always additional overhead related to the display, storage, and RAM.
Intel’s hybrid core design is better described as “big+bigger” as opposed to “big+Little,” and the company is betting its smaller cores are key to hitting its low-power targets. The fact that Intel only starts integrating more than two performance cores in the 28W – 64W bracket tells us a lot about the relative strengths of the efficiency versus performance cores. It is apparently more effective to scale up the E-cores in the 9W – 15W envelope than to add new performance cores.
Spec sheet comparisons are not the same as bench testing, but the fact that AMD, Intel, and Apple have positioned their products this way gives us some evidence for where each company hopes to go.
Apple scaled up from the M1’s very low power envelope by adding more FireStorm cores and reducing the number of IceStorm cores. There is no sign that Apple is working to specifically improve low-wattage performance by respinning its “little” core, though we expect general improvements over time. The most recent Apple rumors suggest the company is working on 20-core and 40-core Mac Pro SoCs.
Apple’s M2 refresh is expected this year, but a report from NotebookCheck suggest the M1 Pro and M1 Max won’t get a refresh until 2023. The M2 will reportedly use TSMC’s 4nm process node and will adopt the same SoC cores that Apple will debut this year with the A16. N4 is an extension of 5nm in the same way that N6 is an extension of 7nm. There is no word on any change to core configurations and Apple is historically conservative when it comes to updating core counts.
AMD continues to argue that it does not need a hybrid x86 core to match its competition. 7nm Ryzen CPUs are already regarded as very energy efficient processors and AMD believes it can continue to deliver meaningful generational improvements without needing two different CPUs. Most of the gains for Ryzen 6000 Mobile are likely to be on the GPU side of things thanks to the adoption of DDR5 and RDNA2, but AMD is forecasting for some CPU-side gains as well.
It’s hard to judge mobile Alder Lake until low-power silicon hits market. What’s certain is that this is a significant shift from any previous Intel CPU family.
Eleven years ago, Intel announced that it would begin emphasizing lower-power TDPs in mobile chips and would standardize on 17W parts as opposed to treating 35W – 45W as the “default” product. By Haswell, the company was putting its low-power efforts and ultrabook launches front and center:
Intel’s Haswell launch documentation.
With Alder Lake, the company has launched 15 CPUs in the 9W and 15W TDP brackets with 1-2 performance cores and 13 higher-end SKUs with 4-8 performance cores. The most common Intel CPU configuration at 15W and below is 2P/8E. At 28W base TDP, 4P/8E is common. At 45W, we see a mix between 4P/6E and 6P/8E.
It’s going to be very interesting to see how these solutions compare once silicon is in-market. We’re especially curious to see the impact of Intel’s hybrid cores on gaming. A number of its 9W and 15W chips also feature either 80 or 96 EUs. This is similar to Tiger Lake mobile, which also offered a number of relatively high-end integrated GPU configurations. In Alder Lake Mobile’s case, however, the typical 4C/8T “big” core configuration has been swapped for 2P/8E. The impact this will have on gaming versus an equivalent 11th-Gen system (if any) is unclear.
The best-case outcome for Intel is that the combination of higher IPC from Alder Lake’s P-core improvements and higher power efficiency from E-Cores allow the company to improve performance within a given power envelope relative to previous 10nm CPUs. The big question to watch this year is how AMD’s power efficiency improvements compare to Intel and how both x86 manufacturers compare to Apple’s M-class CPUs.
Now Read:
- Intel Poaches Apple Engineer Who Lead the M1 Development Team
- Analyst: Apple’s Upcoming Headset Will not be Metaverse-Friendly
- RISC vs. CISC Is the Wrong Lens for Comparing Modern x86, ARM CPUs
No comments:
Post a Comment