Of all the metrics we commonly discuss at ExtremeTech, we’d argue TDP (Thermal Design Power) is easily the worst. If clock speed, core counts, IPC, and memory bandwidth are imperfect methods of comparing the capabilities of two processors, TDP — at least as it’s communicated on current Intel desktop processors — has moved past imperfection and is now downright bad.
Anandtech recently published a deep dive into the power consumption and behavior of two different Intel Comet Lake processors. The Core i7-10700K (8C/16T, 3.8GHz base, 5.1GHz peak, 4.7GHz all-core, 125W TDP) was compared against the Core i7-10700 (8C/16T, 2.9GHz base, 4.8GHz peak, 4.6GHz all-core, 65W TDP). The point of the comparison was to measure the performance of the 65W CPU, with its much lower base clock, against the overclocking-enabled model with nearly 2x the TDP.
On paper, an Intel CPU’s TDP is the maximum power consumed under a sustained workload at base frequency. The reason the 10700K has a 125W TDP versus the 10700’s 65W TDP is because of the huge gap in their base clock. The $52 price difference between the K and non-K version of the CPU, combined with the lower base clock and lower TDP all imply that a buyer who chooses a Core i7-10700 as opposed to the Core i7-10700K will be buying a lower-power version of the chip that trades performance for power efficiency.
In reality, the Core i7-10700 winds up drawing even more power than the Core i7-10700K:
Graph by Anandtech
This requires a bit of explaining. AMD, it should be noted, uses a different formula for TDP. AMD defines its TDP values based on the entire operating range of the CPU. 95W and 105W chips can draw no more than 142W from the motherboard unless the end user overrides it. Similarly, 65W chips are limited to 88W of power draw unless overridden.
TDP Doesn’t Imply What it Used To
The fact that I have to use the word “imply” up there is another reason why TDP deserves a nomination for Worst Metric Ever. AMD and Intel have always been careful to tell reviewers and enthusiasts that TDP isn’t a measure of power consumption. For at least a decade, however, there was a reasonably consistent relationship between the TDPs Intel and AMD claimed on their CPU specs and the maximum power consumption you’d see from the chip in a typical workload. If an Intel chip claimed a 95W maximum TDP, the CPU would typically top out around 85W-90W. At times, AMD or Intel have released a new version of a previously launched chip at a lower TDP, further emphasizing the idea that lower TDPs = lower expected power consumption.
It should be noted that the Core i5-10600K conforms to the old rule. It’s rated at 125W, and it draws 131W at peak. The 65W version of the CPU, the Core i5-10600, almost certainly suffers from the same problem as the Core i7-10700 compared to the 10700K — namely, drawing far more than 65W. We can’t confirm it without testing, but based on how the Core i7-10700 compares to the 10700K, there’s little reason to doubt the Core i5-10600 would also substantially exceed its 65W envelope using default OEM motherboard settings.
For most of the 2010s, Intel kept its typical desktop CPU power consumption at or below the CPU’s rated TDP, even at peak power draw. Once AMD launched Ryzen and Intel had to start adding more CPU cores to its desktop parts, that changed. The Core i9-10850K draws up to 265W but claims a 125W TDP. The Core i7-10700 claims 65W, but draws up to 214W under load, at motherboard defaults. The old relationship between TDP and expected power consumption no longer holds true at the high end of Intel’s market. The Core i7-10700 is guaranteed to draw no more than 65W if you disable Turbo. If you don’t, the CPU will accept whatever Turbo guidelines its handed by the motherboard.
OEMs Control Motherboard Turbo Values
Intel doesn’t dictate Turbo frequencies to the various motherboard companies. Instead, it encourages them to over-engineer their boards and define their own values for short-term maximum power draw (defined as PL2) and other, overclocking-related variables. This is one way Intel gives OEMs the opportunity to differentiate their products. If Asus, Gigabyte, or MSI want to over-engineer a motherboard to appeal to the hardest of the hardcore overclocking set, they’re allowed to do it. If that board can run a given CPU at a higher frequency for longer, Intel says it’s fine to let it do so.
AMD and Intel give OEMs freedom to define these parameters because both companies actively want to cater to hardcore enthusiasts pushing the performance edge. The difference in power consumption between Intel’s formal recommended specs and an OEM’s default values is quite large:
Image by Anandtech. The Core i7-10700K defines itself as a 125W CPU, and when limited to Intel defaults, it behaves like one.
When set to Intel defaults, the Core i7-10700K drops back to 125W after a short boost period. When set to motherboard default values, the chip will run at 217W indefinitely. As Anandtech notes, “Intel is fully supportive of motherboard vendors changing these values.” The company has suggested that reviewers should test multiple commercial motherboards before launch to compare these behaviors. Readers should be aware that no website or publication has the bandwidth to conduct such exhaustive comparisons prior to launch, particularly given the realities of late-arriving hardware and last-second UEFI updates that change the behavior of the CPU from version to version.
Intel Needs to Improve Its Power Consumption and Clock Speed Communication
While we acknowledge that OEMs ultimately control turbo mode settings, selling a CPU with a “65W TDP” when that CPU will actually draw over 200W if run at motherboard defaults is disingenuous. Intel, unfortunately, has made communication changes that make the situation worse.
Intel recently changed how it talks about its Tiger Lake CPUs. Here’s a comparison from Intel’s Ark website, where you can look up various processor specifications. This is a comparison of the embedded Core i7-1185GRE versus the mobile Core i7-1180G7. Embedded is on the left, mobile on the right.
Image from Ark.Intel.com
With the embedded part, Intel specifies a 15W TDP and a 1.8GHz base clock. This means the company is guaranteeing a 1.8GHz clock within a 15W TDP if Turbo is disabled. On the right-hand side, with Tiger Lake, Intel defines a range of TDPs (7W – 15W) and a range of base frequencies (900MHz – 2.2GHz). It’s left to the end user to guess what TDP the OEM has targeted, and what the minimum performance of the machine will be.
The nature of Turbo settings, as implemented by both Intel and AMD, allow the chip to exceed its defined TDP. AMD CPUs have a TDP of 105W and will burst up to 140W, a difference of 1.33x. An Intel CPU with a 65W TDP will boost up to 215W, a difference of 3.3x. Note that this problem seems to largely be confined to Intel’s higher-end, lower-power CPUs: The Core i5-10600K is within its TDP at peak power consumption, while the 10700, 10700K, and 10850K very much are not. The Core i5-10600, while not measured here, would likely also have a peak power consumption well above its 65W TDP.
Intel Has a Choice to Make
It’s easy to see why Intel hasn’t changed the way it defines or communicates TDP to its customers: It would make the company’s high end look vastly worse compared with AMD. A Core i9-10850K will draw up to 265W. If you want an AMD CPU with that kind of power budget you have to buy a Threadripper.
There’s nothing wrong with Intel building a chip that draws 250W+, but end customers need to know what kind of cooling solution they need and how to configure a motherboard for low-power, efficient performance. While the high-efficiency community is smaller than the high performance community, people can and do build small form factor machines explicitly intended to operate in lower power envelopes.
A buyer who picks up a 65W Core i7-10700 for this kind of build may be unhappily surprised by the behavior of the CPU. Even if they aren’t, a 3.3x gap between listed TDP and actual power consumption makes the data worse than useless for the end user. Not only does a 65W TDP mean nothing as far as the behavior the end user will see, it actively encourages the buyer to think of the 65W CPU as an intrinsically lower-power chip than the 125W CPU. As Anandtech shows, this is not the case.
Given that end users still associate TDP with some indication of how much power the CPU will draw, Intel has a choice to make. It can continue to obfuscate actual power consumption data, or it can change its metrics and begin reporting meaningful values that help customers choose the processor they want to buy. Right now, according to AT, the Core i7-10700K is $52 more expensive than the Core i7-10700, and leads the latter by about 2 percent. Power consumption between the two shows the Core i7-10700K draws about 10W less at maximum load, likely due to superior binning.
Intel is currently moving in the wrong direction on these metrics. Removing base clock reporting from Tiger Lake is dishonest. So is selling users a 65W CPU that draws over 200W at OEM motherboard defaults without informing them. These omissions self-evidently favor Intel by allowing it to report lower TDPs relative to AMD. They imply that Intel cares more about making its products look good than communicating honestly with its customers about what kind of behavior they should expect. Building a low-power system with a chip that can and will boost to 215W unless specifically configured not to do so is something people should be made aware of, up front. It’s fine to give OEMs the freedom to set power limits as they wish, but customers also need to be informed of how this freedom could impact their own system builds.
At present, there is no apparent relationship between the rated TDP of high-end Intel desktop processors and the power consumption you will actually see at the motherboard level at OEM default settings. The size of the gap will vary based on motherboard model. Some commercial boards do hold to Intel default settings, but enthusiast desktop boards typically don’t.
The problem here isn’t that Intel gives OEMs freedom to define their own Turbo operating modes. The problem is the fact that Intel’s published TDPs no longer reflect what the end user can expect.
Intel can hide behind the fact that technically TDP isn’t meant to function as a measure of power consumption, but it has directly benefited from user perceptions to the contrary. It benefits from those assumptions because those assumptions used to be true. The world didn’t end when AMD introduced the FX-9590, a CPU with a 220W TDP. Intel’s world won’t end if the company starts publishing either the peak or the typical full-load power consumption of its chips.
Intel has the right to define metrics that are only valuable to OEMs, but it doesn’t have the right to present those metrics as if they say something pertinent or useful to its customers. Right now, Intel desktop CPU TDPs are effectively worthless as far as predicting the power consumption behavior you’ll see in an enthusiast motherboard using OEM defaults. This is demonstrably true for the Core i7-10700 versus the 10700K, and we strongly suspect it would prove true for the Core i5-10600 versus the 10600K as well.
While AMD CPUs also exceed their TDPs at peak power, they do not do so to the same degree. AMD has also introduced an Eco mode with its Ryzen 3000 series, via Ryzen Master, which allows end users to reduce power consumption one step down from stock.
Intel can solve this problem in a lot of ways. It could begin publishing peak TDP or create a new metric that more accurately captures actual power consumption. It could raise the base TDP to align stated power consumption with the CPU’s all-core power consumption at Turbo in a representative (rather than a peak) workload. It could create a method for end users to set a CPU to a given TDP in software. All of these changes would improve the current status quo.
Intel needs to course-correct if it wants the press to continue to consider its published TDP values for desktop and mobile to represent valid, useful information. The only thing Intel’s TDP values currently tell you is what kind of thermal dissipation an OEM needs to provide if they only want to run a CPU at base clock, with no turbo whatsoever. This is of no use to high-end users and should be treated accordingly.
AMD’s published TDP numbers do not represent peak power draw, but measured peak power is only 1.16x (5600X) to 1.33x (5900X) above the claimed 105W TDP. Intel’s measured peak power is 1.63x higher than TDP for the Core i7-10700K, 2.12x higher for the Core i9-10850K, and 3.3x higher for the Core i7-10700. There’s no way to claim the Core i7-10700 is a 65W CPU with a straight face. Not given the practical realities of how OEMs configure their own motherboards.
A 14nm CPU with a 200W – 250W TDP doesn’t compare as well against AMD as a 14nm CPU with a 65W – 125W TDP, but Intel famously prides itself on being engineering-driven, not marketing-driven. The company has a chance to prove it here. Rocket Lake launches in March. Intel can take this opportunity to define a more honest TDP rating, provide a new alternate metric, and/or provide an eco mode that’s easily accessible in software to help efficiency-minded users.
For now, our advice is to take any published Intel high-end desktop TDP and multiply the claimed value by 1.5x – 3.3x for a more accurate estimate of peak CPU power consumption. Use 1.5x – 2.25x for regular high-end desktop CPUs, and 2.25x – 3.3x for lower power SKUs. While that’s a very rough estimate, it’s far more accurate than Intel’s published desktop TDPs. This problem likely extends at least as far down the stack as the Core i5-10600 versus the Core i5-10600K. AMD reliably lands within or relatively near to its claimed TDP values, while Intel’s smallest measured excursion is nearly twice the size of AMD’s largest.
Hopefully Intel’s incoming CEO, Pat Gelsinger will make changes that lead to better communication. As things stand today, AMD and Intel desktop TDP values are not comparable, and while both companies exceed their published values under peak load, AMD’s are at least in the ballpark of their measured values. Intel’s are not.
Now Read:
- Intel, Nvidia Deny Blocking AMD From High-End Mobile Gaming
- Intel at CES 2021: 8-Core Tiger Lake, 35W 11th Gen Mobile CPUs, Rocket Lake
- Linus Tovalds Blames Intel for Killing ECC RAM in Consumer Systems
from ExtremeTechExtremeTech https://ift.tt/39eOAmL
ليست هناك تعليقات:
إرسال تعليق