The A8-3850 heads AMD’s new line of desktop APUs with a new socket, FM1. These new Fusion chips are composed of K10-based quad core CPUs and the fastest integrated GPUs ever. Our complete lowdown on how it fits in today’s complex computing scene.
July 1, 2011 by Lawrence Lee
Socket FM1 APU
Earlier this year AMD launched their first set of Fusion processors, which joined CPU and GPU on the same chip, an APU (accelerated processing unit). It was the long anticipated union of AMD and ATI technology since the latter was acquired by the former in 2006. The new platform called Brazos consisted of a Bobcat CPU (a stripped-down low TDP version of previous AMD CPUs) and an entry level HD 6000 graphics chip, manufactured on a slightly smaller 40 nm process. The functions of the Northbridge were fused into the APU with the remainder bundled in the Southbridge in a new chip dubbed the Fusion Controller Hub (FCH). The new architecture became quite popular in the mobile space, for netbooks in particular. The design consolidation created substantial power savings making it competitive with Intel’s Atom CPU, and the presence of a full-featured Radeon GPU put in the same league as Nvidia ION graphics.
The Llano desktop formula.
The next step for Fusion is the Llano platform, with its desktop chips dubbed Lynx. This time AMD has upped the ante, packing their newly minted desktop A6/A8 APUs with much beefier central and graphics processors manufactured using a 32 nm fabrication process (down from 45/40 nm). The CPU portion has four cores, built on the same K10 architecture as the Phenom II family. Driving graphics are modified Redwood chips (HD 5000 series cores) with the latest version of UVD and improved power gating to maximize energy efficiency. They also sport whopping 320/400 Radeon cores (AKA stream processors or shaders) compared to the 80 in the HD 6250/6310 chips found in Brazos APUs, so they’ve got some teeth and can do more than play HD video. Along with these upgrades comes a new socket, FM1, which with its 905 pins is incompatible with AM3 and AM3+.
Lynx quad core APU die shot.
On quad core versions of Lynx, the CPU takes up just over a half of the die, while the GPU occupies a significant portion on the other side. Due to space, cost, and energy constraints, L3 has been omitted. Each CPU core has its own dedicated 1MB of L2 cache, twice that of the Athlon II line. The die size is 228 mm² compared to 216 mm² for Sandy Bridge and and 256 mm² for Deneb (Phenom II).
AM3 vs. FM1 vs. 1155: US$100~$150 Comparison
Phenom II X4 970 BE
Phenom II X4 965 BE
Phenom II X4 955 BE
Phenom II X4 840
Athlon II X4 645
Athlon II X4 640
Phenom II X2 565 BE
HD 6550D, TurboCore
HD 6530D, TurboCore
HD 2000, HT
HD 3000, HT
HD 2000, HT
We have one FM1 APU on our test bench, the top-end A8-3850 featuring a clock speed of 2.9 GHz and HD 6550D graphics with 400 Radeon cores. The other members of the family include the A8-3800 which is slower but supports AMD’s TurboCore dynamic overclocking feature to make up some of the performance difference. The A6 use the 320 core HD 6550D GPU and are differentiated similarly. In the future, dual core A4 chips will be released as well as enthusiast models with unlocked multipliers.
The A8-3850 and A6-3650 debut at US$135 and US$115, respectively. According to AMD, the A8-3800 and A6-3600 will be priced similarly, though they declined to gives us concrete numbers. The CPU portion of the A8-3850 stacks up closely to the Athlon II X4 640, though the X4 840 is a better match as it has priced the 645/640 out of contention. Of Intel’s Sandy Bridge processors, only the dual core models are in the same price range.
Clock Speed (MHz)
Effective Memory Speed (MHz)
1600 ~ 2000
1000 ~ 2000
1066 ~ 1866
800 ~ 1600
1066 ~ 1866
GPU Peak Compute (GFLOPs)
As for the HD 6550D, it has similar specifications to the HD 5570 with 400 Radeon cores, 20 texture units and 8 ROPs, with a slightly lower clock speed. The memory speed of the 5570 is 1000~2000 MHz depending on the type of memory used by the manufacturer while the 6550D’s memory speed depends on what system RAM is used. Using faster RAM with lower timings typically produces little CPU performance improvement, but memory speed is critical for GPU performance. As the integrated memory controller on the APU services both the CPU and GPU, high speed DDR3 with low latencies is worth more than casual consideration.
A version of Hybrid CrossFireX is also included with the new APUs called Dual Graphics which allows one to combine the prowess of the IGP with a discrete GPU. In the past, this was not much help as the makeup and speed of previous IGPs limited the cards that could be used to entry level models. Dual graphics supports higher performing cards, the HD 6670, 6570 and 6450. Not only do these pairings provide a boost to 3D applications, but as some HD 6000 series cards can support three monitors, Dual Graphics will let you drive up to five displays simultaneously.
Chipset & Motherboard
AMD has two chipsets, the flagship A75 offering essentially the same functionality of what’s currently available for AM3, along with native USB 3.0 support, a first for a desktop chipset. The lower-end A55 chipset lacks USB 3.0, SATA 6 Gb/s, and FIS port multiplication switching, but both chipsets support SecureDigital and Consumer IR.
A75/A55 chipset block diagram.
Our test motherboard is the Gigabyte A75M-UD2H, a feature-rich microATX model with SATA 6 Gb/s, eSATA, USB 3.0, FireWire and all the standard video connectors: DisplayPort, HDMI, dual-link DVI, and D-Sub. It retails for about US$110.
While this is a new platform, the board doesn’t look any different from an AM3 board aside from the slimmed-down heatsink retention bracket. Heatsink mounting remains the same, though, so AM3 supported coolers are compatible with FM1 as well. The board ships with what Gigabyte calls a "Hybrid EFI" BIOS that is basically a standard BIOS that through some technical magic, supports EFI, allowing the board to boot from hard drives larger than 2TB.
The BIOS has the standard features you would expect from a mainstream Gigabyte AMD motherboard: A variety of adjustable voltages, GPU overclocking up to 2000 MHz, CPU frequency up to 500 MHz (100 MHz is the base frequency for Llano APUs), and multiplier adjustment up to 47x, though this was a dummy setting for anything over the 29x default of our A8-3850 sample; the first crop of Lynx APUs are multiplier locked. The A75M-UD2H has a simple 4+1 power phase system so it’s not exactly built for heavy overclocking.
APU in socket.
The socket design remains very similar, though FM1 chips have 905 pins compared to AM3’s 941.
When considering the cost of a system, the CPU/APU is only part of the equation as the price of motherboards varies greatly from platform to platform. In the chart above we added the price of the chips we’re comparing today to the price of compatible motherboards from Newegg. We considered the prices (after any applicable rebates) of all Intel/Asus/Gigabyte/MSI DDR3 microATX and ATX models that had at least three basic features: USB 3.0, SATA 6 Gbps, and gigabit ethernet, and used both the lowest and average cost. For the Core i3-2100 we omitted the P67 chipset boards as they are unsuitable/unlikely choices for i3 users.
The prices turned out to be US$80/$166 for LGA1155, US$78/$118 for AM3, and US$100/$117 for FM1.
Common CPU Test Configuration:
Common Motherboard/Platform Test Configuration:
Measurement and Analysis Tools
Timed Benchmark Test Details
x264 1080p: Spaceship is a 1080p x264 clip encoded from the
Our main test procedure is a series of benchmarks, timed tests of real-world applications. System power consumption (AC) is measured with a Seasonic Power Angel during these tests (an average of the first 10~15 seconds) as well as at idle, during playback of a 1080p H.264 encoded clip, and during full CPU load. To stress the CPU we use either either Prime95 (large FFTs setting) or CPUBurn depending on which
produces higher system power consumption. The AC system power is then later converted to DC.
Certain services and features like Superfetch and System Restore are disabled
to prevent them from affecting our results. Aero glass is left enabled if supported.
We also make note if energy saving features like Cool’n’Quiet and SpeedStep
do not function properly.
Estimating DC Power
The following power efficiency figures were obtained for the
Seasonic SS-400ET used in our test system:
Seasonic SS-400ET Test Results
DC Output (W)
AC Input (W)
This data is enough to give us a very good estimate of DC demand in our
test system. We extrapolate the DC power output from the measured AC power
input based on this data. We won’t go through the math; it’s easy enough
to figure out for yourself if you really want to.
Before we jump into the test results, please note the operating voltages
of the processors tested today as sort of a disclaimer. Different samples of
the same processor often run at slightly different operating voltages which
can affect energy efficiency. Different motherboard models do not apply the exact same core voltage, either. Higher voltages don’t always equate to higher power consumption, but usually, there is a correlation.
GPU Performance: Video Playback
CPU usage during video playback was low but disappointing in Blu-ray playback. Playing through one minute sections of "Disturbia" and "Becoming Jane," the Lynx platform required twice as many CPU cycles as HD 2000/3000 and the Radeon HD 5450. It is possible the UVD 3 chip on the APU was doing more post-processing but the subjective image quality between UVD 2 and UVD 3 was about the same.
From the looks of it, power consumption during video playback is a bit of a concern, particularly when playing Blu-ray, where the A8-3850 system used 4W more than a Core i5-2500K paired with an entry level Radeon HD 5450 video card. However, note that the Intel DH67BL board used with the Core i5-2500K is freakishly energy efficient. Paired with HD 3000 graphics, it was more frugal than the Asus P8H67-M with HD 2000. The power consumption of the P8H67-M is probably closer to what the majority of third party H67 boards pull. Add 2~3 watts to the DH67BL’s numbers and it’s essentially an even playing field, at least with a Radeon HD 5450.
GPU Performance: 3D
In the last three versions of 3DMark using the default settings, the HD 6550D paired with DDR3-1600 memory fell just short of the Radeon HD 5570, a discrete graphics card that retails for around US$50. Bumping the memory speed up from 1333 MHz to 1600 MHz resulted in some nice performance gains. Intel’s HD 3000 was walloped regardless of the RAM used with the AMD’s APU.
At 1280×800 resolution and details set to low, the HD 6550D delivered comfortable frame rates in both the H.A.W.X 2 and Lost Planet 2 benchmarks. Alien vs. Predator slipped to 25 fps with DDR3-1333, while DDR3-1600 provided a substantial improvement. It appears the game is somewhat CPU bound as the APU beat the HD 5570 running in our triple core graphics card testing system.
At 1440 x 900 resolution, the difference between the two memory speeds is most evident. With DDR3-1333, the HD 6550D fell to a somewhat choppy 20 fps in both Lost Planet 2 and Alien vs. Predator. With DDR3-1600 Lost Planet 2 received a 7 fps boost while Alien vs. Predator got a 12 fps bump. These are relatively heavy gains that can differentiate playable from unplayable.
System Power Consumption vs. Competing Platforms
On our motherboard testing configuration, power consumption was similar to the Core i5-2500K and HD 5450 combination when idle and playing H.264 video. When synthetic tools were used to stress the CPU (Prime95 for Intel and CPUBurn for AMD), the A8-3850 system used about 50% more power. The relative energy inefficiency of AMD’s K10 architecture rears its ugly head; it appears that the die shrink to 32 nm hasn’t helped much in this department.
When FurMark was added to the mix to push the GPU, the system consumed 24W more, more than twice that of the i5-2500K HD 5450/HD 3000 configurations. This is expected, considering how much faster HD 6550D is in 3D applications.
System Power Consumption vs. Competing CPUs
Our CPU testing configurations feature a discrete graphics card, the GeForce 9400GT, to isolate the CPU from the IGPs.
On discrete graphics the A8-3850 was impressive, using less power than even the i3-2100, one of the most frugal chips on the market. It was particularly thrifty during video playback where it fell the i3-2100 by 9W, and beat out its Athlon/Phenom II X4 cousins by a massive 15~16W.
Encoding video with HandBrake, the A8-3850 fell in line with the Athlon II X4 635 and Phenom II X4 840. CPUBurn pushed it to X4 955 levels, though this is a synthetic test and isn’t relatable to most applications.
In Photoshop, the A8-3850 draws close to the 3.2 GHz X4 840, suggesting it is a bit more efficient per Hz than standard X4’s.
In NOD32, the A8-3850 performs as one would expect, tying the X4 635 for last place.
CPU Performance (Continued)
The A8-3850 placed last again, this time in WinRAR and used 5W more as well.
In our iTunes encoding test the new APU did a bit better, shaving four seconds off the X4 640’s time.
In TMPGEnc, it fell back down with energy efficiency between that of the X4 635 and 640.
In HandBrake the A8-3850 performed as well as the X4 640, but used 6W more.
We arrived at our overall performance score by weighing each test equally (each composing 1/6 of the total). Mathematically, a processor that finishes first in every single test receives a score of 100. The A8-3850 finished with a score of 74.0, 1.5 points ahead of the X4 635. The X4 635 has the same clock speed and shares the same architecture, but with half as much L2 cache. The cache made little difference, however, not enough to propel the A8-3850 past the 3.0 GHz X4 640.
CPU Performance, Power Consumption, & Value Analysis
After completing our benchmark suite, the A8-3850 system had used approximately 30.7 watt-hours of power, about 5% more than the X4 640. CPU performance was very similar but the APU used bit more energy in some of the tests.
We derived our average system power consumption by assuming that half the time our test system would be on low load (an average of the power consumption when sitting idle and playing H.264 video) and the other half would be spent on heavy load (the average power consumption of our five measured benchmarks). With this usage pattern, the A8-3850 consumes the least of the mainstream X4 lineup, about 0.5W less than the X4 630. [Editor’s Note: Most research studies on typical PC loads over long periods indicates that 90% of the time, idle or very low CPU load is the norm. Our assumption is heavily skewed to a very demanding user.]
To arrive at our performance per watt score we divide the overall CPU performance by the average power consumption and adjust it so the best CPU scores 100 points. The A8-3850, thanks to its excellent energy efficiency on low load, slips past the X4 640 despite its slightly poorer benchmark results.
The A8-3850 fares less well when we take into account its performance per platform cost. With mediocre CPU performance and a higher price than most of the X4 chips on the market, it falls on its face. Note that these figures are for reference only, as they completely ignore the value of the GPU. For that we need to make some adjustments to our data.
CPU Performance, Power Consumption, & Value Analysis (GPU Adjusted)
We peg the performance of the HD 6550D (with DDR3-1600) close to that of the HD 5570 which we estimate to have an idle power draw of about 8W. If we add this to our average power consumption figures for the competition processors and recalculate the performance per watt, the A8-3850’s score goes up by seven points, clearly leading the all of its AMD brethren. However, the i3-2100 still surpasses the APU by a huge margin. If our benchmarks were more biased towards threaded applications, perhaps the positions would reverse, but we feel our benchmark suite is fairly representative of typical usage patterns.
The HD 5570 retails for about US$50 and when we add this to the price-tag of the i3-2100 and the other budget AMD processors, the A8-3850’s price becomes quite reasonable.
With the price adjustment, the A8-3850 moves up a couple of spots in CPU performance per dollar.
The A8-3850 APU is an impressive accomplishment, the merging of a proper quad core desktop CPU with a GPU that is very capable and easily the most powerful integrated graphics solution on the market. The central processing cores have K10 architecture, so the 2.9 GHz APU performs similarly to a Regor/Deneb based Athlon II running at 2.9~3.0 GHz. The HD 6550D graphics chip substantially outpaces Intel’s best integrated GPU, the HD 3000 for Sandy Bridge even when using DDR3-1333 memory. We highly recommned DDR3-1600 memory if you plan on doing any gaming as it provides a significant increase in frame rates and in most cases is costs only a few dollars more (DDR3-1866 is another story). With 1600 MHz DIMMs, it’s roughly on par with the Radeon HD 5570 which retails for about retails for about US$50. Factor in the value of the GPU, and the US$135 A8-3850 is the cheapest quad core you can buy.
Power consumption is excellent but only when the APU is not being pushed. Running on discrete graphics, the A8-3850 displays superb energy efficiency when idle and playing video, surpassing even Intel’s Core i3-2100. On integrated graphics, it is similar to a Sandy Bridge CPU paired with a low-end card like the HD 5450. The improvements made in this area are overshadowed by higher energy consumption on load. When running our CPU benchmarks the system power consumption was close to that of AMD’s mainstream quad core lineup — significantly worse than Intel’s offerings.
As computing power in desktop PCs has outstripped most users’ requirements, we can easily picture Lynx-based computers being adopted en masse by large OEMs. A quad core CPU of almost any speed has more than enough power for a regular Joe, and though the Radeon HD 6550D is slow compared to most discrete gaming cards, it offers more of a punch than the onboard graphics of any current brand name system. Brand name PC makers also like to play up the media capabilities of their machines and the new APU seems perfect in that regard. We anticipate that Sabine, the mobile version of Llano, will be a big hit as well as its integrated design will allow AMD to undercut budget/mainstream laptops with entry level NVIDIA graphics cards.
We can’t picture DIY builders embracing the new FM1 socket though because of its upgrade limitations. Future FM1 APUs are unlikely to offer the kind of performance boost we’re used to seeing from new desktop processors. They have to fit both a CPU and GPU on the same die after all, so faster APU models will likely be only incrementally better than their predecessors. If you decide to upgrade to a high-end standalone graphics card, the CPU portion of the APU would likely be a bottleneck, and dropping in a faster more cost effective AM3/AM3+ CPU would be impossible. For most DIY users, we expect that Lynx-based builds will never go through major changes except for drive and peripheral upgrades. This is fine for niche builds like HTPCs, but there probably isn’t much appeal beyond that. If AMD had developed cross-compatibility with the upcoming Bulldozer family, the Lynx platform would be a more versatile and compelling product.
One of the driving forces behind the development of APUs is general purpose computing on graphics processing units (GPGPU). Applications that use the GPU’s capabilities to augment the CPU’s computational power could make great use of the reduced latency derived from the chip design consolidation. While there has been an official SDK to help develop these applications for some time, support is quite limited. We’re still waiting for software writers to catch up to the technology available. In the future, this will make APUs more than simply the sum of its parts, but for the moment they’re still just CPUs and GPUs sandwiched together for convenience.
* * *
Articles of Related Interest
Intel Core i3-2100T & Core i5-2400S Low Power CPUs
Intel Core i3-2100 vs. AMD Phenom II X2 565
Asus E35M1-M Pro: AMD Fusion Motherboard
Core i5-2400, i5-2500K and i7-2600K CPUs
Intel GMA HD 3000/2000 Graphics
AMD Athlon II X3: Affordable Compromise
* * *