AMD E-350: Fusion and Mini-ITX solutions - BeHardware
Written by Guillaume Louel
Published on February 22, 2011
Announced in summer 2006 on AMD’s buyout of ATI, the Fusion strategy has at last concretised itself with the release of the first APUs. What they’re called has changed (back in the day AMD was talking about Media Processing Units) but the basic concept is the same, integrating CPUs and GPUs on the same die.
The IT market has nevertheless developed considerably over the intervening period. First there was the arrival of netbooks in 2007, followed by Intel’s introduction of a specially designed netbook processor, the Atom, in 2008. In 2010, two other phenomena, the arrival of the tablet and the levelling off of netbook sales, have again moved the market in a new direction. ARM architecture, which already almost exclusively dominated the smartphone and tablet market, received another boost at CES with Microsoft’s cryptic announcement that it would be introducing ARM architecture support with Windows 8.
It shouldn’t be forgotten that, technically speaking, with the Sandy Bridge architecture, Intel is also marketing x86 processors with the CPU and GPU on the same die.
It’s in this context that AMD has launched its first APUs, the Brazos platform (Zacate and Ontario processors) designed mainly for netbooks and entry level or PCHC type PCs. Note that this isn’t AMD’s only ‘Fusion’ project. Llano, which groups Athlon II type cores with a DX11 type architecture is expected in the second quarter and will target mainstream laptop PCs. The chip will be manufactured at 32nm by Global Foundries. The Bulldozer cores will also be available in Fusion versions, in 2012.
The Ontario/Zacate chip
As things stand four distinct APU models have been launched, all based on a new architecture and, of course, with a common die:
Before looking at platforms and performance, we’re going to go into more detail on the architecture choices made.
- E-350, 2 cores, 1.6 GHz, Radeon HD 6310 at 500 MHz, 18 watts
- E-240, 1 core, 1.5 GHz, Radeon HD 6310 at 500 MHz, 18 watts
- C-50, 2 cores, 1.0 GHz, Radeon HD 6250 at 280 MHz, 9 watts
- C-30, 1 core, 1.2 GHz, Radeon HD 6250 at 280 MHz, 9 watts
Atom, Bobcat: x86 architectures
Before coming to the AMD APU architecture, let’s take a look at the architecture of their main competitor, namely Intel’s Atom. To obtain a low energy architecture, Intel made certain radical choices which mean that the Atom is often compared to the first generations of the Pentium (those prior to the launch of the Pentium Pro). One of the particularities of the Atom is that it is aimed at both the mobile segment (under 2 watts, smartphone type) and netbooks and other entry level machines (nettop, PCHC).
While still a superscalar architecture, it is limited to two simultaneous instructions (dual-issue). The comparison with the Pentiums comes from its in-order architecture. This means the execution units process instructions in the exact order in which they are decoded and and placed in the pipeline. The disadvantage of this type of architecture is that the pipeline can be stalled if an instruction is waiting for data from the cache or the memory.
Out-of-order architectures handle this problem with a more complex pipeline and by allowing the order in which instructions will be processed to be changed. This was one of the major developments introduced by the Pentium Pro. The disadvantage of an out-of-order architecture is that it makes the pipeline significantly more complicated. In the case of the Atom, it wasn’t so much to economise on transistors as to limit chip energy consumption that Intel went for an in-order architecture. This compromise was justified by the initial vocation of Atoms (that hasn’t really come into effect) for use in smartphones.
Intel compensated for in-order slow downs with its hyperthreading technology. In practice, the scheduler feeds two instruction lines in parallel (one per hardware thread) which generates more opportunities for processing two instructions simultaneously. The use of hypterthreading is welcome here and although it doesn’t make up for all the faults of in-order architecture (weaker monothreaded performance), there are big gains in multithreaded scenarios:
The impact in monothreaded situations is however inexistent. Each Atom core has a level 1 data cache of 24 KB (+32 KB for the instruction cache), with a 512 KB level 2 cache.
x86 cores used on the AMD APUs are not derived from K8 architecture, but from an architecture (codename K14) specially designed for consumption of around 10 watts and above. In contrast to the Atom, Bobcat has not been designed to run on 2 watts CPUs. This is a disadvantage in terms of product positioning – AMD doesn’t have any products for those markets – but a technical advantage: AMD doesn’t have to make the same drastic technical choices as Intel.
x86 cores and L2 caches on the left. The GPU uses the majority of the die in the middle
The K14 architecture is superscalar dual-issue, like the Atom. AMD has however gone for an out-of-order design. On paper this represents a significant difference with the Atom in monothreaded applications. At 15 stages, the pipeline is short. Note the use of physical register files (PRF), a technique that has recently been used in GPUs and reintroduced by Intel on Sandy Bridge (see our explanation here). AMD also says that using PRFs has an impact on energy consumption (you avoid one data copying operation, pointers to the PRFs are used instead). If there is such a gain, it seems limited to us.
The E-350 viewed in hwinfo32.
Note, among other differences with the Atom, the execution units dedicated to integers can carry out multiplications. On Atom, the floating point units (SSE) take care of this. Another point is that Bobcat includes hardware support for virtualisation, which isn’t the case on the Atom. Note that the caches are also relatively similar, 32 KB for the L1 and 512 for the L2, per core.
Atom, Zacate/Ontario: GPU, interconnexions
On their introduction, Atom chips only included the x86 core (one or two depending on the model). The memory controller and the graphics core were ported to a third party chipset, the two chips being linked by an FSB. Then with its ION platform NVIDIA brought in a full replacement chipset with a memory controller, graphics chip and peripheral support.
At the end of 2009, a new Atom revision (Pineview) altered the landscape: memory controller, graphics core and x86 cores all grouped on the same die. The way all these elements were integrated however remained pretty basic as the graphics controller and memory controller still communicated with the x86 cores via an FSB type bus internal to the chip. This meant the platform’s performance remained similar at equal clocks to the previous generation (Diamondville).
While it simplified the platform and reduced energy consumption (the Diamondville chipset was engraved at 65nm), the Atom Pineview was de facto incompatible with third party chipsets such as the ION. The only solution for NVIDIA was to market a GPU which was connected on the southbridge (the NM10 chipset) using a PCI Express 1x link. Not really ideal and quite limited.
The integrated Atom graphics core is a GMA 3150. It’s DirectX 9.0c and OpenGL 1.5 compatible and an equivalent generation chip to that of the GMA 3100s which were integrated in the G33 chipsets (for Core 2, May 2007). In spite of its name, which indicates an improved version, it's actually lower performance. The chip only has two pipelines for processing pixel shaders (in 2.0 mode), while the vertex shaders are farmed out to the processor (Intel has a very high performance CPU code in its drivers for processing vertex shaders).
This ‘GPU’ is clocked at 400 MHz and includes a video decoding block which only decodes MPEG-2. In terms of drivers, the Atoms use a different driver to Intel’s other graphics chips. Unsurprisingly, support for current games is limited – you won’t be able to run Far Cry 2 or Crysis Warhead for example. The chip does nevertheless support acceleration of the Windows Aero desktop. The graphics part is obviously the weak point on the Atoms.
Radeon HD 6310
Although AMD hasn’t given its APU any particular name, the chip’s graphics part does have a name. Or even two, as in the case of Zacate (the 18 watt version) AMD is talking about the Radeon HD 6310 while for Ontario (9 watt version) it’s the Radeon HD 6250.
In practice, this chip is based on the Radeon HD 5450 architecture. This is, then, a DirectX 11 and OpenGL 4.1 compatible GPU. There are 80 unified shader units accompanied by 8 texturing units and 4 ROPs. Access to the memory is via the common APU memory controller and, in contrast to Sandy Bridge, there’s no shared memory cache between the x86 cores and the GPU.
Zacate/Ontario implementation in detail.
The difference between the HD 6310 and HD 6250 is simply the clock, 500 MHz against 280. To recap, the Radeon HD 5450 is clocked at 650 MHz.
Most importantly, the chip includes UVD3 derived technology which handles the acceleration of video decoding for MPEG 2, VC1, H.264 and Xvid formats. However and in contrast to the ‘real’ UVD3, MVC (the 3D version of the H.264/AVC decoding used for 3D Blu-rays) is not supported. Note that in terms of drivers, the Radeons used use AMD’s unified Catalyst drivers.
The processors, the platforms
The AMD platformsAMD are introducing four models:
These four APU models use a common die, manufactured by TSMC on a 40nm process. These processors can be used both in mobile and desktop platforms. AMD itself does not split the range up into mobile and desktop.
- E-350, 2 cores, 1.6 GHz, Radeon HD 6310 at 500 MHz, 18 watts
- E-240, 1 core, 1.5 GHz, Radeon HD 6310 at 500 MHz, 18 watts
- C-50, 2 cores, 1.0 GHz, Radeon HD 6250 at 280 MHz, 9 watts
- C-30, 1 core, 1.2 GHz, Radeon HD 6250 at 280 MHz, 9 watts
In terms of specs, the APUs support DDR3 memory on a single channel at clocks of 800 to 1066 MHz. 1333 MHz memory, mentioned at the Bobcat architecture presentation last year, is only supported in overclocking mode by motherboards and won’t therefore be exploited commercially speaking. The APU has a PCI Express 2.0 4x port for those who would like to link up a separate GPU.
The APUs must be linked to a southbridge for peripherals support. This time, AMD has gone for a proprietary bus, the UMI, which functions, technically speaking, like a PCI Express 4x between the processor and the chipset.
There are two distinct southbridges (FCHs in AMD parlance), the A50 and A45. The A50 has a Serial ATA 6 Gb/s disk controller while the A45 uses a Serial ATA 3 Gb/s controller. The UMI bus can, on both chips, function both in Gen1 mode (PCI Express 1.0) or Gen2 (PCI Express 2.0). AMD says that as Gen 1 mode reduces energy consumption it's advisable to use it on laptop platforms, with Gen 2 mode for desktop. AMD has not yet been able to confirm in which mode the link was running for the motherboard supplied by Gigabyte.
Otherwise the specs are the same, namely four PCI Express 2.0 1x links and up to 14 USB 2.0 ports. The chipset TDP varies from 2.7 to 4.7 watts according to AMD, depending on the number of ports activated.
The Gigabyte E350N-USB3 motherboard.
For our test, AMD supplied us with a Mini ITX Gigabyte E350N-USB3 motherboard. It uses an E-350 APU accompanied by the AMD A50 M chipset. The APU is soldered onto the motherboard with a BGA fixture. Note that while Gigabyte has chosen active cooling, ASUS has gone for a fanless solution, which makes more sense.
We had a look through the Intel range for something to serve as a performance comparison for these new APUs. Intel’s Atom range is divided into three. First you have the Atom Ds, designed for desktop. These are the fastest, available in single and dual cores with a maximum TDP of 13 watts. The top end Atom D525 has two cores and hyperthreading technology and is clocked at 1.8 GHz. This is the model we tested opposite the AMD E-350.
Intel has two other dedicated mobile ranges. The first, the Atom Ns were designed for netbooks and have the Speedstep technology (clocks reduced at idle – the Atom D clocks are fixed). The chip clocks are lower and there’s only one dual core (the N550 clocked at 1.5 GHz). The TDPs vary between 6.5 and 8.5 watts.
The last range are the Atom Zs designed for mobile usage (smartphones, MIDs and so on). These chips use an Imagination Technologies (PowerVR) SGX 535 graphics core that is identical to those you can find in the Apple iPhone 4, for example.
To check the performance of solutions using the D525, we turned to two motherboards, the Intel GMA 3150 and the NVIDIA ION GPU:
- Gigabyte GA-D525TUD, Intel Atom D525, GMA 3150
- Asus AT5IONT-I, Intel Atom D525, NVIDIA ION
The Atom-based motherboards
As a comparison with faster processors for these entry level solutions, we also added the performance of the Core i3 2100s that we tested recently and the 2100T. To recap, here are their specs:
- Core i3 2100T, 2.5 GHz, 2C/4T, 3 MB LLC, IGP 650-1100 MHz, TDP 35 watts
- Core i3 2100, 3.1 GHz, 2C/4T, 3 MB LLC, IGP 850-1100 MHz, TDP 65 watts
The Intel DH67CF motherboard.
Neither the pricing, nor the TDPs of these solutions are comparable. The motherboard used for these tests was the Intel Mini ITX DH67CF.
Energy consumption, latency, memory
Energy consumptionWe measured the energy consumption of our configurations at the wall socket. We took two readings in load, with Prime95 and Prime95 + Furmark respectively. Other components such as the hard drive were at idle when we took these readings. The power supply block used was an 80 watt Mini ITX Akasa Enigma model. This is an external 12V power supply coupled with a PCB within the casing to handle voltage distribution.
The first surprise here was that the integrated Atom graphics core couldn’t run Furmark. The latest driver for the GMA 3150 supplied by Intel (although updated in 2010) is only compatible with OpenGL 1.5, Furmark requiring Open GL 2.0 (introduction of the GLSL shader language, a standard ratified in 2004!). This is unfortunately far from being the only limitation of the driver as we’ll see further on.
At idle, the Sandy Bridges are actually more efficient. It’s easy to see why in comparison to the Atom D525, which doesn’t have Intel’s SpeedStep technology in the desktop version. Its clock is fixed at 1.6 GHz. At idle our E-350 platform also consumes three watts more.
In processor load, while the Atom draws 4 watts more, energy consumption on the E-350 is up 10 watts. This difference corresponds to the TDPs given for these chips (13 and 18 W respectively). Processor load is much higher on the Core i3s, as is performance of course…
Note finally that the Gigabyte motherboard used for our tests consumes, according to AMD, between 5 and 7 watts more than the competitor MSI model and therefore handicaps the Brazos platform.
Latency, memory bandwidthTo reduce costs and simplify the platforms, the Intel Atom and AMD APU solutions have just a single channel memory controller.
Although the memory controller is integrated onto both chips, with the Atom this is simplified as the old northbridge on the Diamondville Atoms has been added to the die on the Pineviews. However it is still linked to the processor part of the die by an FSB internal to the chip. This limits the gains you can get as a result of integrating the memory controller onto the same die. In terms of clocks, the Atom is limited to 800 MHz in its DDR3 version.
AMD hasn’t really told us in detail how its memory controller is linked to the cores. In terms of clocks, the Ontarios/Zacates were announced last year as supporting DDR3 800, 1066 and 1333 GHz memory. 1333 MHz is no longer officially supported, though it was in practice on our test motherboard, an overclocking board.
For our tests we used DDR3-1333 MHz memory at timings of 9-9-9. We retained the SPD settings, which gives us the following timings (depending on clocks):
- 1333 MHz, 9-9-9
- 1066 MHz, 8-8-8
- 800 MHz, 6-6-6
First we measured the memory latency and the latency of the level 1 and level 2 caches (in nanoseconds).
The automatic SPD configuration guarantees relatively close memory timings, independently of the memory clock. Here Brazos is slightly more efficient than the Atom for memory latency, but both are a long way behind what you get with the Core i3s under the same conditions. At 1333 9-9-9, the E-350 had a latency of 86.8 ns. The L2 cache on the AMD solution has a slightly lower latency but we’re a long way off what you get with traditional processors.
We used AIDA64 to measure memory bandwidth during monothreaded usage of our various solutions:
Even in the absence of dual channel, which significantly favours the Core i3s, the results for the Atom and APUs are low and we’re a long way off theoretical bandwidths (6.4 GB/s for DDR3 800, 8.5 for DDR3 1066). More surprisingly, at 1333 MHz we noted no gain in reads/writes on the E-350. Only copying operations were improved by 2%.
We wanted to confirm this with the RightMark multithreaded memory test which uses a different memory load (128-bit SSE registers) and gives different results to AIDA:
As with the Phenoms, AMD’s memory controllers are significantly more efficient with multithreaded loads. In reads, the difference with the Atom is very small, while the E-350 is systematically more efficient in writes. Again, using DDR3-1333 doesn’t give much of a gain, except in writes. This is probably why AMD doesn’t officially support this memory clock on its platform.
The Atom is down on the monothreaded version when it comes to read efficiency. This relatively standard behaviour can also be observed on traditional processors, with the overhead on multiple threads limiting the performance of a controller that is already very efficient on a single thread. For information, the L2 cache bandwidth is similar on both platforms, in the region of 6 GB/s for reads.
Graphics performance, video playback
First we looked at the performances of the various GPUs integrated into our solutions. They are the NVIDIA GT218 on the ION platform, the Radeon HD 6310 in the E-350 APU, the GMA 3150 in the Atom alone and HD 2000 in the Core i3s.
We used three different graphics modes, a minimal mode at 1280 x 720 (720p low) and a medium mode at 720p and 1680 x 1050. This allowed us to get an idea of the limitations of the integrated GPU and processor.
Far Cry 2 1.03
We used Far Cry 2’s low and medium modes, which are both exclusively DirectX 9.
It was impossible to run FarCry 2 on the GMA 3150, which isn’t really surprising. In comparison to the Atom + ION, the AMD APU did 10% better, though the scores were still too low to consider the title as playable. The HD 2000s are out front but are still far from being playable.
Crysis Warhead 1.1
The Low and Medium modes correspond to the ‘mainstream’ and ‘gamer’ modes in Crysis Warhead :
The faster Core i3 processors give no advantage here, with all our GPUs similarly limited. The ION solution has a slight advantage, but none of the solutions is really playable. Crysis Warhead won’t run at all on the GMA 3150, showing that the chip is not supported.
H.264/MKV video playback
Although the integrated graphics chips on the Mini-ITX solutions weren't designed for gaming, it surely isn't too much to expect them to manage video playback. Once again, this comes down to a software issue. We used two MKV files, encoded at H.264 using x264 at 1080p and 720p respectively. These files have an average bitrate of 17 and 4 Mb/s.
We used version 1.5 of Media Player Classic which supports video decoding acceleration (DXVA) on all platforms. We measured fluidity, noting any jumpiness, as well as average and maximum processor occupation. We used the rendering mode that is enabled by default, EVR Custom Pres:
At first there was some jumpiness during playback of MKV files on the ION platform. We managed to work out that the problem was linked to playback of subtitles that were integrated in our files. There is a school of thought that advises installing Haali Media Splitter and deactivating the Matroska filter in MPC-HC, the theory being that the H.264 decoder used in Media Player Classic doesn’t work with DXVA if subtitling is activated.
In practice and even when we reduced the size of the texture used for subtitles, the rendering of subtitles still posed a problem on the ION platform, both with the internal and external Matroska filter, a problem that we didn’t come across on the other platforms.
Playback on the Atom on its own, without DXVA support for H.264 on the GMA 3150, is handled by the processor alone and jumpiness was then omnipresent as soon as processor occupation of one core got up to 100%.
At 1080p the problems got worse, with playback on the Atom on its own virtually impossible. We didn’t succeed in getting smooth playback with subtitles on the ION platform. Without subtitles, processor usage on the platform was much lower, as of course was the workload.
The E-350 stands out here with perfect playback, with and without subtitles.
YouTube H.264 video playback
We tried to play a YouTube HD video at 720p and 1080p on our various platforms. We used the latest version of Adobe Flash 10.2 and Firefox 4 beta 10 in our tests. We measured the number of frames that weren’t displayed as well as smoothness:
The ION platform achieved perfect fluidity here, as did the E-350 once YouTube had updated its video player for Stage Video (full DXVA acceleration).
Update 22/02/11: There is still a problem with version 10.2.152.26 of Flash, verified by AMD and Adobe when it comes to acceleration of videos for which the H.264 playback component hasn’t been updated on the server side for Stage Video. YouTube has however updated its component for Stage Video and YouTube videos now play correctly, including in HD. This doesn’t however mean that the problem has been resolved definitively for video sites which haven’t updated their video playback component. This is for example the case for DailyMotion where HD video playback still doesn’t work in fullscreen (DXVA inactive, processor occupation 90% and over). Finally, while Chrome uses a slightly more recent version of Flash (10.2.154), this made no difference to the bug in our tests.
It should be noted that without ION, the Atom on its own plays the video without any apparent jumpiness and the Flash stats tool shows a framerate oscillating between 24 and 25 frames per second. Visually however it looked as if only one frame in two or three was displayed.
At 1080p, the results were pretty much the same, once again with reduced fluidity on the Atom on its own, even though there was no apparent jumpiness.
BluRay playbackWe tried to play several Blu-rays encoded in various formats and at various bitrates with a beta version of PowerDVD 10 that is more recent than the one that is publicly available. Unfortunately the software wasn’t very stable and this made it difficult to make a good comparison. In spite of this, here are our impressions. We didn’t bother to test the Atom without ION as the GMA 3150 wouldn’t have a hope with this task on is own:
There was just a very little jumpiness with the E-350 during playback of The Dark Knight in scenes where throughput sometimes got high. Playing the same scene several times didn’t however reveal the same problem repeatedly. It may therefore be linked to the fact that the version of PowerDVD that we were using was very much a beta.
With REC things were more complicated as this Blu-ray is one of those with maximum average bitrate at the same time as being encoded at 1080p50i. We noted some occasional jumpiness which didn’t appear on the ION platform or on the CPU rendering on the Core i3 platform. Once again, the fact that the problem also transpired when using the Radeon HD 5450 indicates that it may well be a more general ATI driver problem or one of PowerDVD support for ATI cards, probably in terms of handling of deinterlacing.
We will re-evaluate Blu-ray playback in more detail on these platforms once we have a more stable version of PowerDVD.
Next we looked at CPU performance for each of the solutions. We added the following common components to the platforms:
- Western Digital Velociraptor WD3000HLFS hard drive
- Radeon HD 5770 graphics card
- Windows 7 64 bits
The discreet graphics card, of course very powerful for this type of configuration, allowed us to check CPU potential independently of the integrated GPUs.
We compressed a series of files of various sizes and compressibility using the LZMA2 multithreaded mode of the open source 7-Zip compression tool:
We noted previously that hyperthreading gave a 47.8% performance gain on the Atom in this test. The Atom therefore finished ahead of the E-350, but was nevertheless 3.5x slower than the Core i3 2100 in this test.
The 3D Cinema 4D rendering software benchmark allows us to evaluate the performance of these processors in demanding tasks, both in mono and multithreaded mode:
It was no surprise to see the Atom 40% down on the AMD E-350 on a single thread. The Atom shows a considerable gain in multithreaded mode: to recap, hyperthreading on its own improves Atom dual core mode performance by 60%. The E-350 however still has the advantage.
We used Avidemux to compress an extract from an HD video in the iPhone format at 640x480:
Although four times slower than our traditional processors, the E-350 is 15% faster than the Atom.
FarCry 2, Crysis Warhead
FarCry 2 1.03
Here we used the same graphics modes as in our tests for the integrated GPUs, using this time a Radeon HD 5770 graphics card, so as not to be limited by the GPUs of the respective solutions tested. In this way we isolated the gaming potential of the processors.
Far Cry 2 is quite resource heavy and although the E-350 is 34% faster in the least demanding graphics mode, once we moved up to medium settings neither processor was able to respond, irrespective of resolution. In this very limited case, the E-350 remained 22% ahead.
Crysis Warhead 1.1
We finished our tests with Crysis Warhead, the graphics load on which is significantly higher.
Once again there was an enormous difference between the Core i3s and the rest, even in the most basic rendering modes. The Atom and the APU are a long way back though the E-350 is nevertheless 32% ahead of the Atom.
We calculated a performance index, giving the Atom 525’s performance in each test a value of 100. Each test was given equivalent weight and for games we used the ‘720 low’ performance for our average.
For pure processor performance, the AMD E-350 scores 23% higher than the best of the Atoms, in spite of being clocked 200 MHz down. The gap to the Core i3 Mini ITX solutions is simply enormous.
To give a comparison for performace per watt, we took the above performance index and set it against the energy consumption of the platforms in processor load. This index does not take into account the energy consumption of the processor alone, but of the platform as a whole. Of course this favours those platforms where the processor consumption is highest, the overall energy consumption (hard drive, memory and so on) working against the lowest energy processors.
Even though it consumes 4 watts less in load, the Atom platform has a slightly lower yield than the AMD platform. The two platforms are still very far from the performance to watt ratio you get with the Mini ITX Sandy Bridge solutions.
Note also that the Gigabyte motherboard used for our tests seems to consume most among the Brazos platforms, drawing 5 to 7 watts more than the equivalent MSI motherboard. We’ll try and check this with a model from another manufacturer.
AMD took a long time to finally bring out a Fusion version of its processors. While it took some time to finalise the common Ontario/Zacate die manufactured by TSMC, on paper it represents a real step forward for AMD.
In terms of processor performance, the new Bobcat architecture does pretty well. Using an out-of-order pipeline does have an impact on the chip’s energy consumption, but the result makes monothreaded performance less frustrating than with the Atom. The Atom is however saved by slightly better memory management and, of course, hyperthreading which can sometimes work miracles, giving, on its own, a 60% jump in performance. The relatively limited Bobcat memory performance, as well as the fact that there isn’t any gain with DDR3 1333 is no doubt why DDR3 1333 is no longer officially supported, though it was initially given as being so.
Of course, on this type of platform, as video playback can be a significant element, the GPU also comes into play. Here, in comparison to the Atom on its own, the AMD APU has an obvious advantage. The support given by the GMA 3150 drivers is very limited, with most modern games not accepted. With video playback, there is no acceleration support which limts playback of ‘simple’ MKV files.
The addition of ION improves things a good deal, although here again, MKV playback wasn’t ideal when it came to subitling support. The AMD APU is undermined on the software side as although MKV playback is no problem, Flash 10.2 struggles to launch video acceleration on sites on which the playback component hasn't been updated. This software bug can be traced back to Adobe/AMD. We are however reserving our judgement when it comes to Blu-ray playback on the AMD platform, noting nevertheless that there were no issues with the ION platform.
From a purely product point of view, the Mini-ITX cards mainly target the PC Home Cinema market and the fact that there isn’t universal software support for the acceleration of hardware decoding is a reminder of the difficulty of creating such a solution for PCs. Where a Core i3 (new or old generation) will, if worst comes to worst, always manage to decode video using the x86 cores, the APU E-350 and Atom + ION solutions are at the mercy of the support or otherwise given by their GPUs. The fact that theres’ no MVC and HDMI 1.4 decoding on these platforms is something else which doesn’t work in their favour, preventing 3D Blu-ray playback.
In the end, what AMD has done here is open a door onto the netbook market with a more flexible platform than the Atom platforms thanks to a considerably more powerful GPU solution. The performance advantage on the processor side also marks AMD’s APUs out from the Atom + ION solutions but here you’ll need to consider things on a case by case basis according to which processors are used: here we only tested the highest performance Atom and APU models.
Copyright © 1997-2013 BeHardware. All rights reserved.