Between our reviews of the mobile and desktop versions of the chip, we’ve spent a tremendous amount of time recently with AMD’s new Llano APU. That doesn’t mean we haven’t had some lingering questions about it, though. We always have more questions about new technologies than we have time to devote to them, seems like.
One of our remaining questions about Llano has to do with its memory support. Uniquely, the desktop versions of Llano officially support DDR3 memory at clock speeds of 1600 and 1866MHz, higher frequencies than the 1333MHz memory currently widely available on the market, with the potential for higher performance.
Now, that official endorsement from AMD comes along with some big, honkin’ caveats. Among them: those speeds are only supported “in a single DIMM-per-channel configuration.” So you’ll have to forgo half of the DIMM slots on most desktop Llano motherboards in order to use those higher speeds. Also, only “true 1.5V, JDEC-spec memory” gets official support.
In other words, AMD’s stamp of approval for the use of higher-frequency RAM with Llano is essentially forward-looking. In preparation for our review, we pinged the memory mavens at Corsair about the availability of JEDEC-approved DRAMs rated for these new frequencies. We were told JEDEC has indeed ratified the spec for those parts. However, Corsair couldn’t give us any firm date for the market availability of such DRAMs. All we know at present is that they’re on the way. AMD’s advice to reviewers concurs with this estimate, stating, “that kind of memory will likely not be widely available until a later time” and conceding that most reviewers “will likely end up testing with 1333MHz DDR3 appropriate for the price tag of this platform.”
That’s not to say one can’t find DIMMs on the market, made with current DRAMs, that will operate at 1600 or 1866MHz. Firms like Corsair perform the invaluable service of sorting and binning DRAMs and packaging up the best chips into enthusiast-class modules. Such modules can sometimes perform amazing feats that would make a JEDEC committee gasp in unison, drop the monocles from their eye holes, and whisper-talk in exasperated tones about how scandalized they are. In fact, among enthusiast brands, DIMMs rated for 1600MHz are fairly common these days and don’t carry much of a price premium over 1333MHz modules. Many of them will even operate at 1.5V. One can find modules rated for 1866MHz, as well, although they tend to come with a fairly substantial price premium attached.
The complexities of this situation forced us to make some choices when formulating our review of the desktop Llano APUs. We took a multi-pronged approach. We started by testing with 1333MHz memory in our CPU benchmark suite, since that’s the type of memory most likely to be used in value-focused systems based on this APU. Also, the architecture of Llano’s CPU cores is largely familiar, and given what we knew about it, we didn’t expect major performance gains from incremental improvements in memory bandwidth. We added 1600MHz memory to the mix in our integrated graphics testing, because that’s where we expected additional memory bandwidth to make the biggest impact. We also requested some 1866MHz-capable DIMMs from Corsair, but those unfortunately didn’t arrive in time for us to test them and include the results in our initial review.
For reasons science is only beginning to understand, the omissions of higher-speed DIMMs in the CPU suite and of 1866MHz memory in the graphics tests led to a minor outcry from some quarters. Little did the complainers know that deep in the underground expanse of Damage Labs, we were taking the wraps off of these puppies:
These two Corsair Vengeance DIMMs are 4GB each and are rated for operation at a clock speed of 1866MHz at only 1.5V with 9-10-9-27 timings. That’s pretty sweet, but at $94 for a pair, they’ll set you back about 30 bucks more than the cheapest 1600MHz modules at Newegg.
We dropped these DIMMs into our A8-3850 APU-based test system and configured them to use Corsair’s recommended timings. That’s a bit of a step back from the 8-8-8-20 timings we’d used before with 1333MHz and 1600MHz DIMMs, but with two crucial differences: these modules are running at 1866MHz, and because we’ve dropped from four DIMMs to two, we’re able to lower the DRAM command rate from 2T to 1T. The end result should be a higher bandwidth, lower latency memory subsystem than anything we’ve tested with Llano to date.
The question is: what difference does it make? To find out, we ran a broad selection of benchmarks from our CPU and IGP test suites, which you’ll find in the following pages. If you want more info on our test configurations and methods, be sure to see our original A8-3850 review for that.
Memory subsystem performance
A couple of quick synthetic memory benchmarks confirm our 1866MHz memory config is indeed both higher bandwidth and lower latency than our default 1333MHz setup. The A8-3850 with 1866MHz RAM manages more throughput in Stream than any flavor of Phenom II, and its memory latency is comparable.
However, these results also serve as a reminder of how proficient Intel’s newer CPU architectures are at wringing as much as possible from their attached memory. The Core i3-2100 manages 5GB/s more throughput with 1333MHz DIMMs—at a 2T command rate, no less—than the A8-3850 does with 1866MHz DIMMs with a 1T command rate. The Intel processors hide access latencies better, too.
These differences are, in part, a result of Intel’s use of larger caches with smart data pre-fetching algorithms that appear to make good guesses about what data our tests will request next. Llano adds a somewhat larger 1MB L2 cache per core and a tweaked pre-fetch algorithm, but it has also shed its L3 cache. In the grand scheme, the changes don’t appear to add up to much.
Among our productivity tests, 7-Zip’s decompression routine benefits the most from the faster memory config, but even there, the gains aren’t terribly impressive.
Video encoding is perhaps the one consumer application most widely considered to be limited by memory throughput, yet our trio of tests reflects only modest gains from 1866MHz DIMMs.
3D modeling and rendering
On the other hand, 3D rendering traditionally hasn’t been sensitive at all to the speed of the memory subsystem, and it proves true to form here. The needle barely moves.
Source engine particle simulation
More so than almost any other tests, our two scientific computing benchmarks have long been willing to take advantage of additional memory bandwidth to good effect, and they do show some modest improvements for Llano with 1866MHz RAM. Still, even here, the differences are fairly small.
Integrated graphics performance
Turns out Llano’s IGP doesn’t extract much more performance from 1866MHz memory, either. The most notable improvement with faster RAM is probably in Civ V‘s Leader benchmark, which works the pixel shaders pretty hard. Even there, 1600MHz memory captures most of the benefit to be had.
Power consumption and efficiency
Another question we’ve wanted to see answered about Llano is how its power consumption looks without a discrete GPU in the picture. We actually did the measurements for the A8-3850 and the Core i3-2100 prior to the publication of our initial review, but we ran out of time and weren’t able to include the results. Now we can remedy that oversight.
Note that the vast majority of the results below show total system power consumption using as similar a slate of components as possible, though the systems obviously varied by motherboard type as necessary. Just four of the result sets omit our standard GeForce GTX 460 video card—the two marked “(IGP)” and the two marked “(Brick PSU).” Additionally, the “Brick PSU” results involve low-power processors—the Atom and the E-350 APU—using a more efficient laptop-style brick power supply. We pondered putting the Llano and Core i3-2100 systems on the brick PSU, as well, but Llano’s 100W TDP probably wouldn’t mix well with that power supply’s 80W peak rating, so we stuck with our standard PSU instead. Out of necessity and because it’s only fair, we did install the Core i3-2100 in a smaller microATX motherboard, the Intel DH67BL, for the IGP power tests.
This is a pretty comprehensive win for Intel—not terribly surprising given that the Core i3-2100 has a 65W power peak. The transition to the DH67BL mobo allows the i3-2100 to achieve lower power draw at idle than the A8-3850, which is a bit of an upset. Even without a graphics card attached, the A8-3850 draws more power under load than the Core i3-2100 does with a GTX 460 installed.
Removing the graphics cards reduces the amount of energy the A8-3850 requires to complete our rendering task, but the efficiency gain isn’t terribly dramatic in the grand scheme of things.
We’re pleased Llano is capable of supporting the latest in DDR3 memory standards, and we’re happy to be able to demonstrate some performance gains when the APU is used in conjunction with faster RAM. However, I think we can say with some certainty that Llano’s overall competitive picture doesn’t change substantially when higher speed memory is added to the mix.
The CPU cores on this chip are based on, uh, a proven-technology microarchitecture, and that architecture doesn’t appear to extract much additional instruction throughput from a higher-performance memory subsystem. One might expect Llano, with no L3 cache, to benefit more than most processors from faster RAM, but these days, it seems CPUs with larger caches and aggressive pre-fetching make the best use of additional memory bandwidth when it becomes available.
More unexpectedly, even Llano’s IGP doesn’t churn out substantially higher frame rates in the presence of 1866MHz memory. At its relatively low core clock speed of 600MHz, the Sumo IGP appears to be getting most of what it needs from 1333MHz or 1600MHz memory.
Despite these results, we do expect memory bandwidth to become ever more critical as CPUs with integrated graphics—or APUs, if you must—mature. Recall that AMD has done little in the way of true architectural integration with Llano, instead opting to focus on reducing power consumption (especially at idle) and getting a product to market quickly. AMD’s next APU in this segment, code-named Trinity, should combine the brand-new Bulldozer microarchitecture with a Cayman-derived, second-generation DX11 Radeon IGP. That chip might also feature Sandy Bridge-esque architectural integration, with features like a high-speed internal bus and a shared last-level cache. Such changes should make this next APU much hungrier for memory bandwidth—and much more capable of taking advantage of it—than this first attempt.