If you’re reading this article eager to find out more about how AMD’s Precision Boost and XFR technology affects its performance, I guarantee that you’ve come to the right place. AMD’s new Ryzen processors went on sale earlier this month, and it comes with a brand new platform, DDR4 memory support, and a lot of new I/O connectivity options for power users to “um” and “aah” over when making their purchase. But there’s one thing in particular that might sway your purchasing decision when deciding which Ryzen processor to buy, and that’s how its boost modes work in comparison to Intel’s Kaby Lake, Skylake, and Broadwell-E families. Let’s dive into this one.
Precison Boost and Extended Frequency range go hand-in-hand
In AMD’s Ryzen press deck, there’s a slide detailing how a typical R7 1800X will move through the motions of managing its clockspeed for a given workload. If you imagine it as an actual graph showing performance over time in a heavy benchmark with a stock cooler attached, it makes a lot of sense. You start at the maximum boost clock, because thermals are still good and power draw isn’t very high with only two active cores. As the benchmark carries on, those two cores boost to 4.1GHz using XFR, but only briefly. As the chip heats up and more cores get loaded up, the clock speed drops to 3.7GHz, before lowering again to the chip’s base clock of 3.6GHz.
This carries on through the imaginary benchmark until the cooling solution on the chip is no longer able to keep it at a decent temperature. The clock speed now drops to 3.2GHz to save on power, until the benchmark completes its run and the processor idles at 2.2GHz. AMD considers the default power state of the R7 1800X to be 3.2GHz rather than its base speed of 3.6GHz, while the second power state is the 2.2GHz minimum.
Straight away, one thing is made clear – if you run an R7 1800X with a crappy cooling solution, you’re in for a bad time. The chip will almost certainly throttle to the 3.2GHz baseline when it is thermally limited, and from there Precision Boost will try increase the clock speed in increments of 25MHz to find a good balance between performance and power consumption. As I’ve noted to before, if you see reviews with something like a Noctua U12S and strange, almost unrealistically low performance, and if they don’t record clock speeds during a benchmark, then this is what likely happened. However, given that these chips shouldn’t be thermally throttling with even a stock cooler, there are other explanations, which I will briefly touch on towards the end.
When it comes to overclocking, things are a bit constrained with the current release. Neo’s engineering sample that he was reviewing couldn’t overclock beyond 4.0GHz on all cores, and some benchmarks online have shown an R7 1800X at 4.1GHz (Wendell from Level1Techs recently posted a 4.2GHz validation). According to the above chart posted by The_Stilt on the Anandtech forums, Ryzen is most efficient at just over 1.0v and a 3.3GHz clock speed. Anything above that starts to consume more and more power, and the curve gets steeper the more you increase the clock speed. How is this relevant? Well, Ryzen’s power states are set voltages, not clock speeds. The rest of the SenseMI system internally will sort those things out, so if you want to keep running at the same power state without dips, you need a good motherboard.
So, how does this compare to Intel’s Turbo Boost?
It turns out that comparing the boost clock presets of Ryzen to Intel’s products is surprisingly difficult. I have notes written from the press briefing about how the all-core boost clock scales, and AMD is pretty upfront about this. However, Intel is not. The latest charts detailing how Turbo Boost works on Intel processors is only valid for their previous generation parts, namely Skylake and Haswell-E. This might be intentional, because Intel isn’t up-front about how performance scales across cores, particularly when Turbo Boost 3.0 in Broadwell-E will boost and overclock a single core, the best one on the die, to run faster in single-core benchmarks when that application is given priority.
Taking into account that Skylake and Kaby Lake have no hardware differences between them that would necessitate a different boost model, we can assume that in desktop parts Intel drops the multi-core boost mode by 200MHz. In Broadwell-E, that ratio is also going to be similar to Broadwell on the desktop, with a 100MHz drop when all cores are loaded. This is why, for example, Skylake and Kaby Lake Core i7 processors share the same wattage, because Intel can use the available headroom from the more mature process to boost clock speeds. Do Intel processors underclock themselves when thermally limited? They probably do, but some testing will need to be done to see exactly where, and under what conditions, this may happen.
Compile these all together in a graph and you have some interesting results! To start, look at the raw clock speed advantage the Core i7-7700K has on all the other processors. It loses almost no discernible performance, and this is impressive considering that it comes listed with a 4.2GHz base clock. However, Ryzen is a lot different. It loses large chunks of clock speed when it hits its thermal limits, and you can see that for the 95W chips, both the R7 1800X and 1700X drop their clocks using the same speed bins. The R7 1700 drops clock speed much more drastically, losing 700MHz of performance going from two loaded cores to three.
The Broadwell-E processors, by comparison, lose almost none of their overall performance. A 100MHz drop for these server-grade chips is nothing in the grand scheme of things, and amounts to almost a rounding error. This doesn’t seem to gel with how the R7 1700 tends to perform when up against a Core i7-7700K in 1440p and 4K gaming benchmarks, but the lower clock speed is compensated with the R7 1700 carrying more L3 cache, which we’ll probably find makes up for that deficit.
Let’s throw one hypothetical scenario into here for good measure. For the moment, let’s assume that Ryzen and Intel processors will always try to exceed their base clock by at least 100MHz, and that we’re using stock coolers that just barely meet the minimum cooling requirements for each platform.
AMD’s first Pstate (P1) for the R7 1800X when it is thermally constrained is down at 3.2GHz, with a voltage lower than 1.0v if we refer back to the Anandtech chart. This is the lowest value that the chip will underclock itself to, in order to save on power and heat. It won’t always be 3.2GHz exactly, because Precision Boost will climb that up by at least 50MHz. The same 400MHz drop in performance might be observed on the R7 1700X and the R7 1700 as well, although it’s possible that the latter will drop another 100MHz or to stay within its 65W thermal envelope.
Where would this happen, realistically? Well, for starters, this would be the case if you happen to pair your R7 1800X overclocked to 4.0GHz to a low-tier tower cooler that’s barely suited for a sustained 95W workload. For consumer workloads at stock settings, a regular tower cooler will be just fine, but when you start overclocking on Ryzen all power savings are disabled. The only thing that would be suitable is a mid- to high-end air cooler like the Cooler Master Hyper 212 EVO, or a budget all-in-one water cooler like Corsair’s H80i or better. Custom water loops with larger radiators have many times more liquid and a larger surface area for heat to dissipate into.
AMD’s maximum thermal limit is 75º Celsius on the chip itself, which is without taking ambient temperature into account. So if you’re in Durban playing with your R7 1700 overclocked to 3.8GHz on the stock cooler, you’ll definitely be hitting the thermal limit more than someone living by the coast in Cape Town, or anyone playing in an air-conditioned room.
There may also be instances where un-optimised software tries to do something that forces the processor to hunker down to the first power state. Prime95 is a good example of this, and it will always result in a Ryzen processor thermally throttling itself because the program overrides the default power state and tries to consume as much power as possible, despite having decent air cooling. Our very own Neo Sibeko is doing some work with motherboard vendors to look into when and why this happens.
Socket AM4 is quite literally a bleeding edge platform for AMD. It has buggy UEFI implementations, and almost half of the boards released aren’t working completely. Power states are not being applied properly. AMD alluded to this in an AMA they had on Reddit, where technical marketing lead Robert Hallock listed a few things that might be causing performance drops:
In addition to Lisa’s comments, there are also some variables that could affect performance:
1) Early motherboard BIOSes were certainly troubled: disabling unrelated features would turn off cores. Setting memory overclocks on some motherboards would disable boost. Some BIOS revisions would plain produce universally suppressed performance.
2) Ryzen benefits from disabling High Precision Event Timers (HPET). The timer resolution of HPET can cause an observer effect that can subtract performance. This is a BIOS option, or a function that can be disabled from the Windows command shell.
3) Ryzen benefits from enabling the High Performance power profile. This overrides core parking. Eventually we will have a driver that allows people to stay on balanced and disable core parking anyways. Gamers have been doing this for a while, too.
These are just some examples of the early growing pains that can be overcome with time.
So, I don’t think that all of the day one reviews are entirely accurate, although that’s no fault of the reviewers themselves. The hardware is brand new. The software is old and unpatched. When you look at games like DOOM with the Vulkan engine for example, you’ll see that AMD’s performance is right where it’s expected to be in terms of IPC. But when you look at DirectX 11 games, OpenGL titles, or even some recent DirectX 12 titles with an artificially engineered CPU bottleneck, the placement of Ryzen in the graphs varies. There are a lot of variables here that are affecting performance, and one by one AMD will be tracking them down and squishing those anomalies. New microcode updates are being sent to motherboard vendors every week, and an update to the AGESA protocol will also solve those super long boot and POST times everyone is seeing.
To this end, I intend on testing Windows 7 with my own Ryzen hardware in the future to see the difference between Windows 7 before the Bulldozer patch, and the same installation straight after the patch has been applied. It might end up proving nothing, but hey – more data is good, right?