Kaveri Days continue here at HCW. First we had our full Kaveri review at launch, which compared the new APU to several CPUs from both AMD and Intel. Then we looked at how CPU performance scaled with increased clock speed, as we overclocked it to 4.7 GHz. Finally, we added a 1020 MHz GPU overclock to the mix, and compared Kaveri’s overclocked performance in gaming and OpenCL.
Today we turn our attention to the memory controller, as we discover how gaming performance scales with memory speed.
It’s common knowledge that it doesn’t take much for a GPU to be memory bandwidth limited. Even the relatively weak Intel GPUs scale up with memory speed. With Kaveri, AMD introduced the fasted integrated graphics controller to date, based on GCN architecture.
The FM2+ platform Kaveri sits on can run memory at speeds up to DDR3 2400 MHz. It may be possible to take it higher with a BCLK increase, but we’ll stick to standard memory multiplier speeds for now.
For this review, we’ll be using an 8GB kit of AMD’s own Radeon branded memory “GAMER SERIES” that runs up to DDR3 2400 11-12-12-31 at 1.65v. These modules are on the pricey side, when you compare to what’s available from other brands, but if you are going all out with an AMD gaming system, they might look great in a theme case.
|APU Specs||Radeon 7850K CPU
3.7 GHz Base
4.0 GHz Turbo
Radeon R7 Series IGP
AMD A88X Chipset
|Memory||8GB (2x4GB) Radeon Memory GAMER Series
|Hard Drive||OCZ Vector 256GB
|Motherboard Drivers||Catalyst 13.30 RC3 Chipset
Catalyst 13.12 SB
|Video Drivers||Catalyst 13.30 RC3|
|Operating System||Windows 8 Professional x64 RTM
All updates as of Dec 1, 2013
Everything will be running at stock, but with any power saving functions disabled. The memory’s JEDEC timings were used, so latencies are automatically adjusted to suit clock speed. We’ll go all the way down to DDR3-1333, although there is little reason to ever buy that speed any more. In fact, the most popular speed is still DDR3-1600.
How We Test Video Game Performance
Our gaming performance tests are probably a bit different from what you’re used to seeing on most sites. We test performance using real-world gameplay exclusively, and the results we look at are frame time measurements in addition to standard frames per second.
We have found that FPS doesn’t carry enough data to tell us truly how a given set of hardware actually feels when you play a game. A simple way to explain this is that any given 60 second segment of playthrough may have 2000-3000 individual frames shown to the user. What FPS does is take all 3000 of these data points, and combine them into one overall number. Although the use of “frame rate over time” is better than this, it still takes those 3000 data points and turns them into 60. Because the variation of speed that occurs frame-by-frame and not second-by-second is what can make a game feel ‘stuttery’, even though the “FPS” is high any given second or minute, it is important to look at game performance in terms of these individual timings. For further explanation, check out Scott Wasson’s Inside The Second article on the subject. He is the one who spearheaded the whole movement into this type of testing.
Each game is tested using a 60 second playthrough of a level with the intention of being repeatable and consistent. This is easier in some games than others, so our benchmarks are repeated 3-5 times to find the most consistent result. Once that result is found, we use that as a representation of performance. FPS is recorded, as well as individual frame times.
We have traditionally tested IGP performance at 720p with very low settings. However Kaveri has pushed performance to the point where many games (but not all) can be bumped up to 1080p, with some settings increased. The ‘easier’ games were tested at 1080p as well as our old 720p benchmark (for those who prefer 60 FPS with lower res over 30 FPS at high res). Games that already struggled at 720p were left there.
I decided to drop Metro: Last Light from our suite of tests, since it seems to have issues with the pre-release drivers we’re using.
Without further ado, let’s get it on with the benchmarking!
Assassin’s Creed IV
We tested AC4 in both 720p and 1080p, since Kaveri runs it pretty well at the lower resolution. This test is using our new benchmark path, which is very easily repeatable, and tests the engine in an open view as well as down in the streets of Havana.
Frame time charts can be clicked for a full-res version.
AC4 is an interesting game to kick things off, since there appears to be two different thresholds where the scaling takes place. At 720p, performance seems to be acceptable with memory as slow as DDR3 1600. At 1080p though, performance falls off a cliff, and you will need at least 1866 MHz DDR3 to get anything close to a playable frame rate.
This is just the first game though, let’s see how the rest compare:
We used the first part of the Singapore beach landing from single player BF4. Although the vast majority of people who play this game will be doing so in multiplayer, it would be almost impossible to do a meaningful benchmark that is repeatable. As you can see in our benchmark videos, we need everything to be specific and repeatable, and results can’t vary much. Especially when you’re looking at such minor differences in performance.
This time you will need DDR3-1866 to get within 7% of DDR3-2400 performance at 1080p. At 720p, it’s within 4.48%
Crysis 3 is one game where you won’t dare play at a resolution higher than 720p, especially in the brutally demanding section our benchmark takes place in. So this time, we only have one benchmark to show you:
Frame time charts can be clicked for a full-res version.
The trend continues here, as DDR3-1866 and 2133 perform relatively close to each other, this time 2400 nets about a 9% advantage over 1866.
Tomb Raider is the last game we’re looking at today, and another one that performed well enough at 720 that we thought we would have a crack at running it at 1080p with higher settings. This time we bumped the setting to “Normal” but disabled Antialiasing (normally FXAA).
Tomb Raider appears to scale more linearly, at least at 1080p. In this case, DDR3-1866 is at about 90% the performance of DDR3-2400.
Before we wrap things up, let’s take a look at the average performance across all the games we tested, at all resolutions, and see what kind of performance drop we get at each memory speed:
And there you go; this is how IGP gaming performance scales on Kaveri, as memory speed is increased. While performance is maxed out at 2400 MHz, it’s only about 3% faster on average than 2133 MHz. This is at stock speeds, mind you. If you overclock Kaveri, it probably makes even more use out of the extra bandwidth you give it.
Does this mean you will have to spend a lot of money to get full performance out of Kaveri? Not necessarily.
What this chart shows is the average price of the 5 cheapest modules of 2x4GB at each speed, as well as the lowest price found on PCPartPicker.com (MIR not included). As you can see, you don’t have to spend very much extra cash to get to DDR3 2133 or 2400.