What makes a CPU fast?
WHETHER YOU’ RE BUYING OR BUILDING A NEW PC, THERE’ S A WHOLE SPREAD OF CPUs TO CHOOSE FROM. DARIEN GRAHAM-SMITH EXPLORES WHAT MAKES A TOP-PERFORMING CPU– AND WHETHER YOU NEED ONE
Whether you’re buying or building a new PC, there’s a whole spread of CPUs to choose from. Darien Graham-Smith explores what makes a top-performing CPU – and whether you actually need one.
C
PUs may all do the same broad job, but they’re not all alike. As the benchmark scores in our reviews confirm, the chips found in lightweight laptops are nowhere near as powerful as the heavyweight processors used in graphical workstations and gaming desktops. In this feature, we’re going to explore exactly why.
Perhaps the simplest reason is that different models run at different speeds. If one processor runs at 3GHz and another runs at 4GHz, it’s pretty obvious that the latter will be faster to complete a given task.
This raises the question of why anyone bothers making processors that run at slower speeds. One reason concerns manufacturing challenges. There’s never any guarantee that every chip that comes off the assembly line will be stable at its target speed: tiny, sub-microscopic variations between chips are almost inevitable, and that can have a significant effect on how hot a particular unit gets. This matters because if the chip overheats, the whole processor – and the computer it’s housed in – is liable to crash.
To deal with this, chip manufacturers test their silicon before selling it; chips that are perfectly stable at 4GHz will be sold as such for a pretty penny. Others might not work reliably at 4GHz, but run just fine at lower speeds – so rather than junking them, they’re locked to a lower maximum speed and sold at a discount.
This process of testing chips after manufacture and categorising them according to their observed capabilities is called binning – a slightly ironic name, since it means that sub-standard chips don’t get thrown away. In fact, since midprice CPUs usually sell better than cutting-edge models, a manufacturer might bin chips down to a lower speed than they’re capable of just to meet demand. In theory, there’s nothing stopping you dialling up the clock speed on any processor to get some free extra performance – all you need is a motherboard that gives you access to the required timing settings. However, if you stick with a standard heatsink and fan then overclocking your CPU by more than a few hundred megahertz will probably cause it to become unstable. With more powerful cooling solutions you can go much higher: the current overclocking record was set in 2012 by enthusiast Andre Yang, who used a liquid nitrogen cooler to overclock his AMD FX-8350 processor from a standard speed of 4.2GHz to 8.8GHz.
GOING TURBO
We tend to think of processors as running at a set clock speed, but that isn’t really the case. To minimise heat build-up and power consumption, both Intel and AMD’s chips reduce their clock speeds when there isn’t much work to do, then quickly rev back up to full speed as needed. In fact, Intel’s “Turbo Boost” technology allows the CPU to temporarily accelerate above its nominal clock speed under heavy load, and AMD chips use a similar system called Turbo Core.
As to how fast the CPU will go, and for how long, that depends. The manufacturer normally sets a maximum Turbo frequency for single-core operations, with a lower limit for multicore jobs. The processor will try to maintain these speeds for as long as there’s work to do, but once a predefined temperature limit is reached, it will drop back to standard speed to cool down.
This creates additional performance differentiators for CPUs. Some chip models have big Turbo Boost ranges, while others will offer only a small speedup – and some low-end designs might not support Turbo Boost at all.
The upshot of this is that you don’t necessarily need to overclock your CPU to get a performance boost: if it supports Turbo Boost, a better cooler will allow it to sustain its maximum speeds for longer. Conversely, a powerful processor with inadequate cooling will only be able to achieve short bursts of peak performance, holding back overall performance. This is why thin and light laptops are often slower than chunkier models with better airflow and larger fans – even if they’re using the same CPU.
FLUSH WITH CACHE
The clock speed of a processor determines how quickly it can execute incoming instructions, but this relies on there being a continuous flow of instructions to process. This isn’t something you can take for granted.
To see why, let’s imagine a processor that loaded its instructions one by one from the main system memory bank, and executed them in turn. This would be a very inefficient way to proceed since it takes several clock cycles to fetch a value from RAM. This processor would spend much of its time sitting idle waiting for the next instruction to arrive.
To avoid such delays, modern
CPUs include fast on-chip memory caches that can keep the processor cores continuously supplied with instructions and data. Typically a three-level arrangement is used, with the level one (L1) cache comprising a small amount of extremely fast memory, while L2 and L3 get progressively larger and slower. Clever logic is used to anticipate which data is likely to be wanted in the near future, and to pre-emptively transfer it into the appropriate cache.
The trouble is, it’s not always possible to predict which data will be needed next – for example, a program might do different things depending on which button you press. And if you’re doing heavy multitasking, the cache might simply not be large enough for all the information your CPU cores want to access at once.
Chip designers use excruciatingly clever and complex models to try to maximise the accuracy of their caching predictions. But if the budget allows, it can also help to simply build more cache memory into the chip. Thus, Intel’s humble Celeron G5905T processor has 4MB of cache RAM, while the mighty Core i9-11900 has 16MB and the server-grade Xeon Platinum 8358 has 48MB. AMD’s
mighty Ryzen Threadripper 3990X, meanwhile, boasts a massive 256MB of L3 cache, helping it keep all 64 cores well fed with data.
CORE S, OF COURSE
On that note, it’s time to acknowledge the not very secret fact that almost all modern CPUs have more than one core. This allows them to perform multiple operations at once – a big step up from the old single-core approach, which required the processor to continually switch its attention back and forth between different processes.
Multicore computing can be a big help for multithreaded workloads – that is, ones that divide into a set of independent sub-processes that can all be performed in parallel. Not every task can be broken down in this way, however, and doing so increases the complexity of the program, so even applications that could take advantage of multithreading don’t always do so.
The other benefit of having multiple cores is for multitasking. Even if you only have one application in the foreground at a time, the Windows Task Manager will show you dozens of little processes all running behind the scenes – some connected to the operating system, others carrying out background duties for applications that may be open but not active. Although it may appear that everything is running on one CPU core, the individual execution threads will be distributed across all available cores to keep everything flowing smoothly.
For this reason, having two cores is a prerequisite for a modern multitasking operating system; going up to four helps give web browsers a boost when they need to render multiple tabs at once, or allows games to generate and manage a whole onslaught of enemies. Beyond that, though, you’re likely to hit diminishing returns outside of specific multithreaded workloads.
HYPER-THREADING
We implied above that a processor core can only handle one operation at a time. The truth is a little more complicated. Modern CPUs feature custom hardware that lets them parallelise certain tasks: for example, Intel’s SSE instructions (Streaming
Single Instruction, Multiple Data Extensions) allow programs to perform a single operation on multiple values simultaneously.
What’s more, many Intel models make use of a design enhancement called Hyper-Threading, while AMD chips use a similar system called SMT (simultaneous multithreading). In a Hyper-Threaded chip, each core has two instruction pipelines and two sets of registers (the internal variables that a CPU works with). As a result, each core appears to the operating system as two virtual processors – although in reality, these are different interfaces to a single execution core.
Although the chip can still only execute one instruction at a time, its available resources can be shared between the two pipelines, so one thread can take advantage of resources that the other isn’t using. And Hyper-Threading really comes into its own when a program has to fetch a value from the system memory: in a regular chip this would just leave the processor sitting idle for ten or 20 clock cycles, but with Hyper-Threading a different thread can have full use of the core during that time.
Hyper-Threading clearly isn’t the same as having twice as many physical cores, but it provides extra headroom for everyday desktop computing at a much lower cost. Even in hardcore multithreaded workloads it can often provide a benefit of around 20%.
WHAT’ S IN A NAME?
With all these factors contributing to processor performance, it’s not reasonable to expect a normal person
to work out which of two models will be faster. From 2008, Intel set out to simplify the market by dividing its Core processors into three families – Core i3, Core i5 and Core i7, targeting lightweight, mainstream and high-performance systems. It’s a good, clear system, and AMD has shamelessly followed suit, dividing its own chips into Ryzen 3, Ryzen 5 and Ryzen 7.
The branding isn’t as transparent as it may seem, however. For one thing, Intel cranks out a new generation of chips almost every year, and the capabilities of the chips are constantly changing. In some recent years a Core i3 chip has meant two physical cores and no Turbo Boost, while the current roster includes quad-core i3 models with full Turbo capabilities.
Even within a generation the branding doesn’t necessarily tell you a lot, because each chip family spans all the way from lightweight laptop chips right up to chunky workstation processors. The latest 11th-generation Core i5 lineup comprises 21 different models, with power consumption levels between 7W and 125W. Needless to say, between the top and bottom of that range there’s a huge variance in performance.
As if that weren’t confusing enough, the original three-tier lineup has been extended with even more powerful Core i9 processors, in both regular and top-end “X-series” configurations, while AMD now offers Ryzen 9 chips. Intel’s old Pentium and Celeron brands have been kept alive for budget CPUs that rank below Core i3 too.
You can’t even be sure about a given chip’s graphical capabilities.
Some Intel chips use the lightweight UHD Graphics GPU, while others have the latest Iris Xe silicon, and different models have different numbers of graphical execution units (EUs). The “G” rating appended to the model number gives you a hint of comparative GPU power: the Core i3-1125G4 has UHD Graphics with 48 EUs running at 1.25GHz, while the Core i5-1155G7 features Iris Xe with 80 EUs at 1.35GHz (note that the “11” at the start of the processor name indicates it’s an 11th-generation chip). As with the main Core branding, however, these numbers don’t promise anything consistent, especially not across generations.
AMD vs INTEL
We’ve talked a lot about the generalities of processor design; this may be interesting, but it doesn’t necessarily tell you which processor you want in your next computer.
If you intend to run Windows then there are two main chipmakers to choose from: AMD and Intel. You don’t have to agonise too much over this choice: thanks to long-standing licensing agreements, the two companies both make processors that fully support the x86 architecture, and can, therefore, run Windows and all popular desktop apps. And the truth is, unless you’re looking at real bottom-of-the-barrel hardware, almost any processor you buy today will do for everyday computing.
Chips from the rival manufacturers tend to have different strengths and weaknesses, however. Historically, Intel’s processor cores are individually faster than AMD models for single-threaded tasks, and often run at faster clock speeds. AMD compensates for this by building more cores into its processors – meaning that, while Intel’s chips mostly offer between six and eight cores, you will see plenty of Ryzen CPUs with 16 cores or even more. To keep all of these cores fed with instructions and data, there’s normally more on-chip cache memory too.
As we’ve intimated above, more cores doesn’t always equal better performance. While some applications can benefit enormously from the extra silicon, an everyday desktop might be better off with fewer, faster cores. We recommend that you work out your own requirements, and check both the single and multithreaded benchmark scores to identify the best options within your price range. It should quickly become apparent which CPU models are right for you.