We first reviewed an AMD Kaveri processor back at the start of the year, but since then, AMD’s new APU has been in kind of a weird place. The A8-7600 chip we reviewed has been scarce in retail channels, evidently because AMD succeeded in selling them elsewhere—likely to big PC manufacturers. Some of the chips were surely set aside for use in laptops, too. As a result, PC hobbyists just haven’t had very good access to AMD’s latest APU.
Happily, that situation is finally changing, and Kaveri-based chips are starting to make their way into the market. AMD is putting an exclamation point on that fact today by filling out its APU lineup and making some tweaks to its pricing. The headliner of the bunch is a brand-new model, the A10-7800, that may just be the most desirable Kaveri-based desktop processor yet.
Here’s a look at AMD’s updated lineup:
The brand-spanking-new A10-7800 nearly matches the top-of-the-line A10-7850K in terms of clock speeds and unit counts, but it does so in a much smaller 65W power envelope. And it costs less than the 7850K. Given everything, I’d say the A10-7800 is the Kaveri chip to get, as long as you don’t plan on overclocking your processor. (Only the K-series parts have unlocked multipliers.)
Also new today are official retail editions of the A8-7600 and A6-7400K. The A8-7600 is the same basic product we reviewed in January, while the A6-7400K is an unlocked K-series part. At $77, the 7400K matches up against the unlocked Pentium G3258, but going directly against an overclocking titan like that one would probably be suicidal. The 7400K is better suited to providing an attractive CPU-IGP combo for truly low-end systems.
At $155, the A10-7800 is priced in a gap between Intel’s Core i3 and i5 desktop parts. That’s a clever tactical move by AMD. The company’s marketing materials clearly position the A10 against the Core i5, so the firm is looking to be the lower-cost alternative. As we’ll see, though, the A10-7800 will have to deal with some of the top Core i3 offerings in order to stake that claim.
Meanwhile, the new Kaveri-based APUs face some unusally formidable competition from a familar source: past generations of AMD APUs, specifically those based on 32-nm Richland chips.
As we noted in our initial review, AMD had some tough choices to make with Kaveri. The 28-nm process provided by its chipmaking partner, GlobalFoundries, offers some potential advantages, including increased power efficiency and the ability to pack more logic into a given area. AMD has used the extra gates to cram in lots of graphics horsepower—specifically in the form of the GCN graphics architecture used in the latest Radeons. GlobalFoundries’ 28-nm process is not, however, tailor-made for CPUs like its older 32-nm SOI process was. The transistors in Kaveri chips can’t switch quite as quickly as those in AMD’s 32-nm chips, and as a result, CPU clock speeds are somewhat lower.
For instance, the A10-7800 ostensibly replaces the Richland-based A10-6700. Both are 65W quad-core processors. The A10-6700 has a base clock of 3.7GHz and a Turbo peak of 4.3GHz. By contrast, the A10-7800 runs at 3.5/3.7GHz.
AMD has attempted to make up this deficit by tweaking the “Steamroller” CPU cores in Kaveri to increase per-clock instruction throughput. As we’ll soon see, those improvements have paid off to some degree. Still, this isn’t the best time for AMD to be treading water when it comes to CPU performance, given how big a lead Intel holds in this department.
AMD hopes to paper over its relative weakness in CPU performance by pushing for more desktop applications to use the GPU side of the chip to help with computing tasks. The concept makes sense—and heck, tablets and phones are using GPU acceleration pretty consistently these days—but unfortunately, Windows applications that take advantage of GPU computing have been slow in coming. Speaking of which…
What about that 12-core APU?
No, Virginia, AMD is not releasing a 12-core processor any time soon.
The Internets have been afire with a rumor about a 12-core APU lately, prompted by some AMD marketing materials that focus on the number 12.
I suppose the enthusiasm is natural; 12 cores is a lot of cores. I’m not sure what folks expect to do with them all, but whatever. Here’s the thing, though: in its push for “converged” computing, AMD has taken to calling its graphics compute units “cores.” By this reckoning, the A10-7800, with four CPU cores and eight graphics “cores,” would be a 12-core processor.
So I guess AMD really is introducing a 12-core processor. They’ve also had another one on the market for a while now.
Mind blown. Poof.
Meanwhile, Intel sells a Core i5 chip with four CPU cores and 20 graphics execution units, so it has 24 cores, right?
Hrm. I’m not so sure about this new math.
The competition has gotten more potent since Kaveri’s initial release, too. Intel has refreshed its Haswell lineup from top to bottom, raising clock speeds while generally keeping prices the same. Here’s a relevant sampling of current Haswell models.
As I’ve mentioned, the A10-7800 is priced in between the Core i3 and i5, so we don’t have an exactly direct competitor to test against. Instead, we’ll bracket the 7800 with higher- and lower-priced Intel CPUs.
The very cheapest Core i5 on Intel’s price list is the i5-4460 for $182. There are some slower models, but Intel has priced them at $182, too—a sign that they’re on the way out. Thing is, you can also grab the Core i5-4590 for only $10 more than the i5-4460. The 4590 has a 100MHz faster base clock and a 300MHz higher Turbo peak. I figure I’d take that deal if I were buying a CPU in this range, so I chose the 4590 as our representative from the Core i5 lineup.
The Core i3-4360 stands in for the Core i3 camp. The $138 price you see for it in the table above comes from Intel’s price list. Intel CPUs typically stay close to list, but we paid $153 for our i3-4360 when we ordered it from Amazon recently. AMD probably wouldn’t want to admit this, but the Core i3-4360 may be the A10-7800’s closest competition, truth be told.
We also have the A6-7400K lined up against a stock-clocked Pentium G3258 in a pitched battle among budget chips. We fully intended to overclock both chips as part of a larger battle, but… well, many things didn’t go as planned in the making of this review.
I started out with big plans. It’s been a while since we’ve had one of our epic, full-scale CPU reviews, and I figured it was time to produce another one with updated tests, games, and CPUs of every stripe. I collected a ton of chips and threw myself into the task. Nearly everything would be new, and we’d try all manner of intriguing tests, including our famous inside-the-second analysis of frame-by-frame gaming performance. I tested the A10-7800 versus the Core i5-4590 in some particularly difficult gaming scenarios: the “Welcome to the Jungle” level of Crysis 3, wandering the city in Watch_Dogs, Battlefield 4 with and without AMD’s Mantle API. And much, much more.
Then I looked up. Four days had passed. My weeked was gone. So was Monday. And I’d only managed to test two of the planned eight or nine CPU on my list. Whoops.
The review I’d envisioned was going to be glorious. But it was also going to kill me—and severely delay our coverage of the new Kaveris—in the process.
Ultimately, I had to pull the ripcord and slim down the selection of CPUs and tests. What you see on the following pages is just a down payment on a larger CPU roundup that’s still in the works, but it should be sufficient to tell the story of AMD’s new APUs. We’ll get into more detailed gaming coverage and the like, with a broader selection of CPUs, once I’ve had more time to prepare.
You’ll see results for the AMD FX-8350 and the Core i5-2500K in the following pages. Consider them a bonus. I had the sockets open, so I was able to test them alongside the other processors. The FX-8350 is AMD’s fastest CPU—save for the crazy 220W parts that require a water cooler. With eight integer cores and four FPUs, the FX-8350 is an interesting reference, at least. The Core i5-2500K, meanwhile, was introduced at the start of 2011 and was an enthusiast favorite from the start. When the 2500K went end-of-life in early 2013, it cost $224, a little more than the Core i5-4590 costs now. I’m intrigued to see how today’s chips fare against it.
Our testing methods
The test systems were configured like so:
|Processor||AMD FX-8350||Athlon X4
|Motherboard||Asus Crosshair V Formula||Asus A88X-PRO|
|North bridge||990FX||A88X FCH|
|Memory size||16 GB (2 DIMMs)||16 GB (4 DIMMs)|
|Memory type||AMD Performance
|AMD Radeon Memory
|Memory speed||1600 MT/s||1866 MT/s|
|Memory timings||9-9-9-24 1T||10-11-11-30 1T|
|AMD chipset 13.12||AMD chipset 13.12|
Realtek 184.108.40.20633 drivers
Realtek 220.127.116.1133 drivers
|OpenCL ICD||AMD APP 1526.3||AMD APP 1526.3|
|Motherboard||Asus P8Z77-V Pro||Asus Z97-A|
|North bridge||Z77 Express||Z97 Express|
|Memory size||16 GB (2 DIMMs)||16 GB (2 DIMMs)|
|Memory speed||1333 MT/s||1333 MT/s|
|1600 MT/s||1600 MT/s|
|Memory timings||8-8-8-20 1T||8-8-8-20 1T|
|9-9-9-24 1T||9-9-9-24 1T|
|INF update 10.0.14
|INF update 10.0.14
Realtek 18.104.22.16833 drivers
Realtek 22.214.171.12433 drivers
|OpenCL ICD||AMD APP 1526.3||AMD APP 1526.3|
They all shared the following common elements:
|Hard drive||Kingston HyperX SH103S3 240GB SSD|
|Discrete graphics||XFX Radeon HD 7950 Double Dissipation 3GB with Catalyst 14.6 beta drivers|
|OS||Windows 8.1 Pro|
|Power supply||Corsair AX650|
Thanks to Corsair, XFX, Kingston, MSI, Asus, Gigabyte, Intel, and AMD for helping to outfit our test rigs with some of the finest hardware available. Thanks to Intel and AMD for providing the processors, as well, of course.
Some further notes on our testing methods:
- The test systems’ Windows desktops were set at 1920×1080 in 32-bit color. Vertical refresh sync (vsync) was disabled in the graphics driver control panel.
The tests and methods we employ are usually publicly available and reproducible. If you have questions about our methods, hit our forums to talk with us about them.
Power consumption and efficiency
We tested total system power consumption at the wall socket by plugging our test rigs into a power meter. What you see above is the power consumed during a set period of time. During that span, we asked each CPU to encode a video with the x264 encoder. You can see that some processors took longer than others to finish—and the various systems drew different amounts of power as they worked.
Both Intel and AMD have made some nice strides in recent years at reducing power consumption at idle by integrating more components onto the CPU, where they’re under the control of a power management policy. All of the CPUs we tested participate in those advances except for AMD’s FX-8350, which still plugs into a socket with discrete, chipset-based PCIe connectivity.
We removed the discrete Radeon HD 7950 for some of our tests. Those results are labeled “IGP” at the end. As you can see, the configs with integrated graphics alone tend to be the most efficient overall.
Although the A10-7800 has a lower peak power rating than the Core i5-4590—65W to 84W, respectively—the 7800-based system requires higher wattage at peak in this workload. Manufacturers’ TDP ratings are just peak numbers, and they aren’t terribly useful for comparisons with other brands.
Adding up the total power used for the duration of the test period is one means of measuring efficiency. We can also look at the power used only as the encoding work was being done, which involves shorter spans of time for the better performers.
Intel rules the efficiency rankings in this workload, thanks to a combination of lower peak power draw and shorter encoding times. The good news for AMD: Kaveri is clear progress over Richland. The A10-7800 requires less power than the A10-6700 to encode the same video. The generational improvement isn’t huge, but it’s definite progress.
Discrete GPU gaming
For now, we have only a single game test, Thief‘s built-in benchmark, to show us discrete gaming performance. No matter. The results tell a familiar story.
AMD’s latest CPUs just aren’t terribly good at cranking out frames quickly in PC games. Even the stock-clocked Pentium G3258 outperforms the A10-7800 in this game’s default Direct3D mode. With its higher clock frequencies, the A10-6700 is slightly faster than its Kaveri-based counterpart, too. Since it has only two relatively pokey integer cores (and a single, shared FPU), the A6-7400K suffers even more.
Switching over to AMD’s Mantle graphics API, with lower CPU overhead and apparently better threading, seems to help somewhat. All of the CPUs get faster, and the A10 APUs are at least able to maintain a 60 FPS average (although FPS averages alone won’t tell you much about true smoothness.) Oddly, the two dual-core CPUs wouldn’t start Thief properly in Mantle mode, for whatever reason, which is why those results are missing above.
Gaming with integrated graphics
In the test above, we used a discrete GPU to remove any graphics bottleneck from the picture. Now we’ll consider what happens when we switch to the graphics processors integrated into each of these chips. In this case, the IGP is much more likely to be the primary performance constraint—which puts us on Kaveri’s home turf.
Yeah, that changes things. Not only does Kaveri have more graphics grunt than Richland and Haswell, but AMD has also officially blessed the use of DDR3-2133 memory with the A10-7800. As a result, the 7800 clearly outperforms both its predecessor and the competition from Intel in these tests.
This outcome is a big part of AMD’s pitch: if you care about both graphics and CPU performance, Kaveri can offer the best mix of the two. I can see the logic, but there’s still not enough bandwidth going into the CPU socket to allow graphically intensive 3D games to run terribly smoothly on this IGP. Hmm.
Let’s run through a quick sampling of some desktop-style applications that rely on the CPU cores to do their work.
Remember what I said about AMD’s CPU performance treading water from one generation to the next? It’s on display above. The A10-6700 and 7800 trade spots from test to test without any clear overall victor. The A10-7800 also exchanges blows with the Core i3-4360, and to my eye, it looks like the i3-4360 comes out a bit ahead in the overall mix. The more expensive Core i5-4590 is clearly much faster than either of them.
The A6-7400K, meanwhile, suffers through a series of embarrassments against the Pentium G3258. Putting a single Steamroller module up against a dual-core Haswell is not a good proposition. Then again, the integrated graphics results on the last page were almost as lopsided in the 7400K’s favor.
LuxMark OpenCL rendering
LuxMark is a nice example of GPU-accelerated computing. Because it uses the OpenCL interface to access computing power, it can take advantage of graphics processors, CPU cores, and the latest instruction set extensions for both. Let’s see how quickly this application can render a scene using a host of different computing resources.
I found that AMD’s APP software driver was faster on both Intel and AMD processors than the latest Intel ICD. Even so, Intel’s FPUs outperform AMD’s here. The dual FPUs (each shared between two integer cores) on the A10-7800 can’t keep pace with the dual cores on the Core i3-4360.
The A10-7800’s potent IGP comes out on top here, well ahead of the pre-GCN graphics units in the A10-6700. Intel has made some nice progress on the GPU compting front, too, though. The Core i5-4590’s IGP isn’t that far behind the A10-7800’s.
Perhaps due to some changes in the plumbing in Windows 8.1, combining the CPU cores and integrated graphics to work on a shared task is much more effective now than it used to be. There’s a clear speedup in each case. Unfortunately for AMD, the Haswell chips appear to have the better combination of CPU and GPU computing power for this application. The A10-7800 trails the Core i3-4360 slightly.
Heh, well. The numbers get much larger when you farm out the work to a proper discrete GPU like the Tahiti-based Radeon HD 7950 (same thing as a Radeon R9 280, if you want to feel minty fresh.) The faster CPUs wring a little more performance out of the discrete GPU, but the differences aren’t that large.
Ask the CPU and discrete GPU to team up, and the sample rates go even higher. The A10-7800 just barely outperforms the $72 Pentium G3258 in this case, though.
For our final benchmark, we have an expanded series of results sourced from my recent overclocking experiments. Cinebench also renders a scene, but it uses only CPU power to do so.
The single-threaded results above emphasize the issue with AMD’s CPU performance right now. Nothing AMD has to offer, even a chip overclocked to 4.8GHz, can match the per-thread performance of the 3.2GHz Haswell Pentium. At least in this workload, AMD has taken a step backward with Kaveri versus Richland.
Today’s changes, including the addition of the A10-7800, bring AMD’s desktop CPU offerings up to date with the latest Kaveri silicon. Our direct comparison of the A10-7800 versus the A10-6700 illustrates what that shift means. Kaveri is somewhat more power-efficient than Richland, and its integrated graphics are substantially improved. Unfortunately, the greatest single weakness of AMD’s current APUs, their per-thread performance in CPU-intensive tasks, hasn’t gotten better. In fact, in some cases, the 7800 is a bit slower than the chip it supplants. The ground lost isn’t worth getting worked up over, but who could blame us for wishing for something more?
I’d like to think AMD’s push for GPU-accelerated computing could help make up the difference, but there are two problems with that prospect. One, the applications that use the GPU to accelerate common tasks have been terribly slow in coming. AMD has been talking about these things for years, but everday Windows programs just haven’t made the transition to GPU acceleration in any great numbers. Two, the results we saw in our LuxMark OpenCL rendering test suggest the Core i3-4360 offers a more potent combination of CPU and GPU computing power than the A10-7800. Whoops.
One area where AMD’s APUs clearly do have a lead is integrated graphics performance in intensive 3D games. The A10-7800 has the world’s best integrated graphics in a socketed desktop processor. The question is: does that matter? There’s not enough bandwidth going into a CPU socket—or, more importantly, enough IGP performance coming out of it—to make hard-core PC gaming on an APU an attractive proposition. If you plan to play games, our recommendation remains the same as ever: get a decent CPU and plug in an inexpensive discrete graphics card like the GeForce GTX 750. You’ll have a much better experience.
In light of everything we’ve just said, we’re led back to a familiar conclusion: that AMD APUs make a lot more sense for environments constrained by size, power, and thermals than they do in big desktop systems. That shouldn’t be much of a revelation. Kaveri is clearly a mobile chip first and foremost. Many of AMD’s engineering choices were driven by that fact.
There are desktop systems where Kaveri could play well: all-in-one PCs, small-form-factor systems, Steam machines, and so on. The A10-7800 and the A6-7400K both have “configurable TDP” modes; they’ll sacrifice clock speeds to fit into a 45W power envelope. That option could be interesting for a quiet home theater PC or the like. Still, finding the right niche where an AMD APU clearly makes more sense than a competing Core i3 or i5 processor will require some creativity.
Fortunately, we now know AMD is committed to improving this situation. Kaveri is the first CPU that’s fully compliant with AMD’s HSA accelerated computing architecture, and the APUs based on it can serve as the development platform for HSA-enabled applications. That is just a start, really, but it’s an important piece that was missing. Second, the firm has an all-new, x86-compatible CPU core in development. If all goes as planned, we could see products based on this new core in 2016.