Six months ago, AMD introduced its Kaveri APU design to the desktop market. It represented a lot of firsts for the company. For instance, it was the first processor to feature what AMD calls a truly Heterogeneous System Architecture (HSA). Its x86 execution cores are the first to leverage the Steamroller architecture, replacing the Piledriver design leveraged in both Trinity and Richland. Kaveri is also the first desktop APU armed with AMD's Graphics Core Next (GCN) architecture, featuring many of the same building blocks at the heart of today's discrete Radeon graphics cards, offering TrueAudio functionality and support for AMD's Mantle graphics API.
On paper, and in practice, Kaveri is a compelling mainstream, budget-oriented SoC. You can read more about it in our launch coverage: AMD A10-7850K And A8-7600: Kaveri Gives Us A Taste Of HSA.

Unfortunately, the APU's life in the desktop space isn't easy. Kaveri is intended to operate within a broad range, from 15 to 95 W. AMD has already talked about sweet spots inside of that thermal envelope; right from the start, company representatives said this design does its best work at TDPs under 45 watts.
We also know that Kaveri's 28 nm SHP bulk silicon process optimizes density at the expense of higher clock rates, which could be another reason the mobile version proves more capable than anything AMD could enable on the desktop.

Indeed, when it comes to the desktop, the 45 W A8-7600 (an APU that we reviewed in January, but is still not available to purchase, by the way), serves up performance results that look impressive compared to the 95 W A10-7850K. In other words, you get relatively little extra in the benchmarks for the 50 W-higher power ceiling. In fact, in many cases, the Kaveri-based A10-7850K barely improves on the previous-generation A10-6800K, which offers more overclocking headroom thanks to its 32 nm SOI process.
Yes, Kaveri's Steamroller-based cores do get more work done per clock cycle than Piledriver. However, the architecture's advantages are largely masked by lower clock rates at comparable power levels. When the playing field is normalized by battery power, though, AMD says its modern APU design is at its best.

And so we now have fairly aggressive expectations, shaped by AMD's triumphant marketing message. At the end of the day, though, engineering has to live up to the PR team's promises. It's time to determine whether practice lives up to theory.
The first, most obvious change from AMD's mobile Richland design is DDR3-2133 support. That's technically not a new feature for Kaveri, which already accommodates fast data rates on the desktop. But it's certainly something we haven't seen before in the mobile space. Unfortunately, there's only one 35 W mobile APU with DDR3-2133 support; the rest cap out at 1866 MT/s. And the 17 to 19 W versions are validated for up to DDR3-1600.

Of course, the mobile version of Kaveri demonstrates different thermal behavior compared to the desktop implementation. AMD is actually leveraging the same kind of optimizations it exposed in the low-power Mullins and Beema APUs (Mullins And Beema APUs: AMD Gets Serious About Tablet SoCs). Many aspects of the SoC's utilization are monitored on a constant basis, and the goal is to provide quick, high boosts in frequency that dynamically respond to user input. Once a snappy response is achieved, the clocks quickly drop to lower levels, cutting power consumption and preserving thermal headroom.
AMD is trying to give the impression of fast, responsive performance, and our sample does seem to achieve that. Of course, this model falls apart when sustained, intense workloads are applied to the hardware. But all chip makers share the same challenge when it comes to high performance in small form factors.
New Brands: Mobile FX, and The AMD Pro Series APUs
Applying the FX brand to mobile APUs will no doubt raise eyebrows amongst enthusiasts. That moniker was already feeling watered-down after the Bulldozer generation let us down. In truth, this is simply a new top-tier sub-class for AMD's notebook-oriented hardware. Previously, A10 was used to suggest top-of-the-line. Now FX does that job, implying the highest available clock rates, functional units, and performance. Again, the FX-7600P distinguishes itself as the only APU with DDR3-2133 support, while the 19 W FX-7500 is limited to DDR3-1600.
In addition to the FX brand, AMD adds a new line of products to address the low-voltage commercial and business space. It dubs these the AMD Pro-series APUs, most of which appear to be re-badged versions of the ULV 19 W consumer models. The exception is the only dual-core Kaveri mobile model, the AMD A6 PRO-7050B. Of course, the Pro marketing message focuses on attributes that IT departments like to talk about: performance, longevity, and stability. It's a bummer that the Pro line-up doesn't include FirePro driver certifications, though AMD representatives concede this could become possible if the market (or certain customers) demand it. Why did AMD choose to bring a commercial version of its APU to market? AMD's Bernard Fernandes told us that it's a response to requests spurned by the increase in visually-oriented workloads. "We have made commitments to the image stability, product life cycle, and performance required in commercial environments with the AMD Pro A-Series APUs."
Here's what the initial stack of Kaveri-based mobile APUs looks like:
| CPU Cores | Base/Max CPU Freq (GHz) | Shader Cores | Max GPU Freq (MHz) | L2 Cache | Max DDR3 Speed | PCIe Lanes | TDP (W) | ||
|---|---|---|---|---|---|---|---|---|---|
| AMD A-Series SV APUs | |||||||||
| FX-7600P | 4 | 2.7/3.6 | 512 | 686 | 4 MB | DDR3-2133 | 1x16 Gen3 | 35 | |
| A10-7400P | 4 | 2.5/3.4 | 384 | 654 | 4 MB | DDR3-1866 | 1x16 Gen3 | 35 | |
| A8-7200P | 4 | 2.4/3.3 | 256 | 626 | 4 MB | DDR3-1866 | 1x16 Gen3 | 35 | |
| AMD A-Series ULV APUs | |||||||||
| FX-7500 | 4 | 2.1/3.3 | 384 | 553 | 4 MB | DDR3-1600 | 1x8 Gen2 | 19 | |
| A10-7300 | 4 | 1.9/3.2 | 384 | 533 | 4 MB | DDR3-1600 | 1x8 Gen2 | 19 | |
| A8-7100 | 4 | 1.8/3.0 | 256 | 514 | 4 MB | DDR3-1600 | 1x8 Gen2 | 19 | |
| AMD A-Series Commercial ULV APUs | |||||||||
| A10 PRO-7350B | 4 | 2.1/3.3 | 384 | 553 | 4 MB | DDR3-1600 | 1x8 Gen2 | 19 | |
| A8 PRO-7150B | 4 | 1.9/3.2 | 384 | 533 | 4 MB | DDR3-1600 | 1x8 Gen2 | 19 | |
| A6 PRO-7050B | 2 | 2.2/3.0 | 192 | 533 | 1 MB | DDR3-1600 | 1x8 Gen2 | 17 | |
That top-of-the-line FX-7600P is a completely functional Kaveri APU, with both of its Steamroller modules (four integer cores) and all 512 of its shaders enabled.
As with the benchmarks we ran on AMD's Mullins APU last month, the company provided us with a prototype platform for evaluation. Also similar to last time, we were given a few short hours to squeeze as much comparative data out of the system as possible. Consequently, we don't have as many results as we'd like, though we did come away with some meaningful results.

The early sample was equipped with AMD's FX-7600P APU and outfitted with 8 GB of memory in a dual-channel configuration. CPU-Z reported it as 1890 MT/s, though we're inclined to believe that could have been a detection error.

With AMD's marketing message honed in on 19 watts, we were a little surprised to test the 35 W APU. Then again, in a story about performance, it makes sense to put the best foot forward.
While AMD appears confident in the ULV Kaveri's ability to compete against Core i7-4500U, the higher TDP meant we needed to find an Intel-based platform with a comparable thermal ceiling. I tapped the 37 W Core i7-4702MQ within Acer's Aspire V3. Although it includes a discrete GeForce GPU, I was able to disable that component using Nvidia's Optimus technology, verifying that it was off before running our tests. Special thanks to Jared and our friends at iTech Computers in Winnipeg, who were kind enough to facilitate our testing. Surprisingly, it was difficult to find a mobile Haswell-based processor rated at 35 W.
We begin with 3DMark's synthetic graphics tests. More than likely, you already guessed that AMD's graphics engine would trump Intel's HD Graphics 4600 solution. But can the updated Steamroller architecture stand up to the efficient Haswell design?

Futuremark's Cloud Gate benchmark humbles this specific implementation of Kaveri, which features two Steamroller modules in what AMD likes to call a quad-core configuration. Because the Core i7 is a Hyper-Threaded quad-core solution, scheduling eight threads at a time, it has little trouble dominating. The Ice Storm Unlimited metric is a little more forgiving to AMD.

In both tests, AMD presents a compelling graphics story. However, Intel gets the aggregate advantage due to its superior execution cores.
AMD claims that Kaveri can hang with Haswell when it comes to PCMark's Home suite, though. How does that one pan out?

Indeed, the FX establishes a slight advantage. This is probably a fairly accurate reflection of the user experience in simpler tasks that don't come close to touching the APU's thermal ceiling.

Sandra's Arithmetic module helps quantify the theoretical superiority of Intel's x86 logic, which we already know to be quite fast in single-threaded and more parallelized workloads. But...

If we go the other way to show what AMD's OpenCL efforts enable, the combined potential of host processing and graphics together heavily favor the FX processor.
Granted, there still aren't many OpenCL-accelerated apps. However, the situation is changing slowly. Though it's safe to say that most software benefits more from Intel's efficient architecture, we've seen a number of tasks sped-up dramatically by optimizations for general-purpose computing on graphics hardware. AMD needs to continue beating that drum if it hopes to change the way software is programmed.
Battlefield 4's cutting-edge graphics engine is probably too demanding for on-die graphics engines, but I wanted to give it a shot anyway. The game also gives me the chance to test with DirectX and Mantle on the Kaveri APU. Benchmarking began at Battlefield 4's lowest detail settings and 1280x720.


You'll notice that the Core i7 and its HD Graphics 4600 are missing from the charts. Unfortunately, Intel ran into significant artifacts, including stuttering and textures popping in and out. Strangely, the benchmark reported smooth framerates, though the output certainly wasn't playable.
The FX-7600P never achieved smooth frame rates, though I consider its best effort respectable for a mobile APU. I was more surprised that there wasn't much difference between performance under Mantle and DirectX; both APIs enable output between 21 and 26 FPS during our benchmark sequence.
Frame time variance is acceptably low, though unplayable frame rates keep variance from even becoming a consideration. You'd need higher performance before thinking about the consistent delivery of output.

In contrast, Grid 2 tends to offer smoother performance on lower-end graphics hardware. My first thought was that I could get away with the game's Medium quality preset at 1920x1080. Let's see if that was a good guess:


AMD's FX-7600P scores a solid win as Intel's Core i7-4702MQ dips below the 30 FPS threshold. In contrast, the Kaveri-based APU doesn't fall under 34 FPS.

The Core i7 exhibits less unwanted variance through our test. It's only a shame that the frame rate we measured is so low. We can't consider it a viable platform for playing Grid 2 at these settings.

Valve's Dota 2 is one of the most-played games online, so it's important to gauge how each platform handles this prolific title.
Dota 2 is attractive. But its graphics engine isn't particularly demanding, so we're able to try the highest detail settings at 1920x1080. I apply a few tweaks to ease the burden on integrated graphics processors; ambient occlusion and the additive lighting pass are disabled.


Intel's Core i7-4702MQ flexes its muscle in this platform-limited sequence. With that said, AMD's FX-7600P remains playable, only dropping under 30 FPS a handful of times.

Although I couldn't pick out problematic stuttering in Dota 2, my frame time variance figures on the FX-7600P get disturbingly bad. Due to the nature of strategy games, where the camera is placed above the battlefield in a static location, it's harder to identify lag between frames. I'd still call the AMD APU playable, though Intel's Core i7 has a clear advantage.

The last game I'm testing is also still widely played, though its popularity is arguably waning. Nevertheless, World of Warcraft remains a relevant metric that could conceivably be a viable target for gamers with laptops. We played it using the Good detail preset with 4x MSAA enabled at 1920x1080.


Although Blizzard's MMO gives us the flexibility to increase fidelity using anti-aliasing on AMD's FX-7600P, that option simply crushes Intel's HD Graphics 4600 engine. There's no comparison as the Kaveri-based APU never drops below 30 FPS.

The tables turn compared to Dota 2. The difference is that the Core i7's dismal frame rate is painfully obvious, regardless of the ugly variance figures.

With little time to spare, we ran a few Web-based tests on AMD's prototype laptop.



While the scores overwhelmingly favor Intel's Core i7-4702MQ and its efficient CPU architecture, I will go on-record saying that any user sitting in front of either system and browsing the Internet wouldn't notice a difference between them. Where Intel would no doubt pull ahead, indisputably, would be in the threaded tasks we typically run in a CPU review. Those tests require more time, and I'd want a shipping product to look at from AMD, too. Neither of those luxuries were afforded this time around, unfortunately.
Kaveri is finally being introduced to the mobile space, and I believe it has more potential there than on the desktop. Where the previous generation of Richland-based APUs offer greater headroom for higher clock rates, that advantage is largely neutralized when you're more worried about battery life than raw performance.
Of course, AMD's mobile APUs aren't competing against the products they replace; they're doing battle against Intel's best effort, which is manufactured using more advanced technology. In that light, the mobile flavor of Kaveri fights an uphill battle to prove its mettle, particularly when the comparison points are benchmarks of popular applications.
The same arguments crop up over and over. Will you notice a difference between platforms while you're banging out emails, working in Excel, or browsing the Web? Probably not. But that's not a good enough reason to adopt a slower or less energy efficient system. Yes, OpenCL and the HSA initiative have wonderful potential, but it's still potential. Neither effort is yet prolific, and that's what we really want to see.
How about graphics? Score one for AMD there, though in certain host processing-bound applications, Intel's x86 cores alleviate bottlenecks that Kaveri must suffer through. But let's make this clear: in the four games we tested, AMD's FX-7600P established a clear win over the Core i7-4702MQ in three. Intel's HD Graphics 4600 engine was either unplayable or it dipped below the threshold of playability, while AMD proved more likely to deliver tolerable performance numbers. If you're a mobile gamer, that's a notable distinction between low-power processors.
The final piece of the puzzle is price. We won't know the specifics until AMD gets some mobile Kaveri design wins on the shelves. At a similar cost, you'd choose a Core i7-4702MQ or FX-7600P based on your preference for intense computing tasks or graphics ability, respectively. But I'd be a little surprised if those two chips ended up in similarly-priced laptops. If you could save a significant amount of money by choosing the AMD option, the value proposition could be compelling. As always, we'll have to wait until we have commercially available laptops to test in order to do a real value analysis. But no matter how you slice it, the mobile version of Kaveri is a stronger opponent than its predecessor.
