AMD Trinity A10-4600M APU Review: Jumping the Shark?

[siopao1984]

---

IntroductionAMD's 2012 Consumer top to bottom VISION product positioning with FX CPUs for the enthusiasts, Trinity APUs for the mainstream and Brazos for the entry level.

Summary of Trinity APU improvements - now using "Piledriver" modules (2nd generation of Bulldozer, still the obtuse 2 integer/1 FP unit per module design), featuring DDR3-1866 memory support (1600 for notebook) and depending on SKU, up to 384 Radeon "Cores".

Technical Specifications and block diagram of the partitioning inside the die

The desktop APUs are still rated at 100W TDP like Llano models (not too surprising considering that they are both on the same manufacturing process), with 17W/25W/35W models for mobile variants. Operating frequencies are expected to be between 2.0-3.8GHz on the CPU and 424-800MHz on the GPU. FMA3 instruction set (which is the one adopted by Intel on the upcoming Haswell, instead of FMA4 first seen on Bulldozer) makes its debut here. Improved scheduling, branch prediction and prefetcher also account for higher IPC than the Husky cores used in Llano.

Integrated northbridge and memory controller

The new Turbo Core 3.0 now balances power envelope of the CPU/GPU to fit within the TDP, dynamically throttling the clocks to fit the workload profile of the active running application. There is no indication on whether this is handled by hardware or software driver (ahem quack3.exe).

These are the 32nm Trinity mobile SKUs that are officially announced today:

AMD A-Series Mobile APUs (FS1r2 package, 35W TDP, DDR3-1600 support)

Model	GPU	CPU Cores / Modules	CPU Clock (Max/Base)	L2 Cache	Radeon Cores	GPU Clock (Max/Base)
A10-4600M	HD 7660G	4 / 2	3.2 / 2.3 GHz	4MB	384	686 / 497 MHz
A8-4500M	HD 7640G	4 / 2	2.8 / 1.9 GHz	4MB	256	655 / 497 MHz
A6-4400M	HD 7520G	2 / 1	3.2 / 2.7 GHz	1MB	192	686 / 497 MHz

AMD A-Series Mobile LV / ULV APUs (FP2 package, DDR3-1333 support)

Model	GPU	CPU Cores / Modules	CPU Clock (Max/Base)	TDP	L2 Cache	Radeon Cores	GPU Clock (Max/Base)
A10-4655M	HD 7620G	4 / 2	3.2 / 2.3 GHz	25W	4MB	384	497 / 360 MHz
A6-4455M	HD 7500G	2 / 1	2.8 / 1.9 GHz	17W	2MB	256	424 / 327 MHz

Expect to see desktop Trinity APUs surfacing in the next quarter, with the chips using the new FM2 socket and motherboards.

Read more: AMD Trinity A10-4600M APU Review: Jumping the Shark? by VR-Zone.com

[siopao1984]

Graphics and Heterogeneous System Architecture



Recommended Dual Graphics Configurations (Mobile GDDR3 Northern Islands)




Trinity's graphics is an improvement over Llano's but it still employs the VLIW4 architecture of the Radeon 6000 series instead of the more modern Graphics Core Next on recent Southern Islands GPUs (UVD engine on the Radeon 7000 was backported though).







Double Precision is capped at 1/16 SP.





Heterogeneous System Architecture is a developer initiative by AMD to seemlessly utilize CPU and GPU (hence APU) to accelerate applications with scalar processing on CPU and parallel processing on the GPU (as a co-processor).

[siopao1984]

The Comal Platform (Mobile Trinity APU)

'Trinity' A10-4600M APU silicon measuring 35x35 mm and features two "Piledriver" 32nm HKMG 128-bit FPU/compute modules, 128KB L1 and 4MB L2



722-pins



Socket: FS1r2 722-pin lidless uPGA





A70M FCH (65nm, 2.7-4.7W, 6 x SATA 3.0, 4 x USB 3.0, 10 x USB 2.0, 2 x Internal USB 1.1)





The 14-inch reference whitebox notebook applied by AMD looks exactly like Dell's Vostro series and comes with a bog standard 1366x768 screen







A standard 50wh battery was included in the package





System specifications dump





Using the balanced productivity/entertainment setting on PowerMark, we got a result of 3 hours and 16 minutes, which is within range of most Ultrabooks that we've encountered.





The included Samsung 830 128GB SSD performed 50% below expectations on the reads, and we did check whether SATA 3.0 6Gb/s was negotiated at both ends of the drive and controller. None of our tests later in this review is storage I/O bound.

[siopao1984]

Test Setup / The Contenders





System 1: AMD Trinity A10-4600 (35W Quad Core) with Radeon HD 7660G

AMD Comal Reference Platform

2 x 2 GB DDR3-1600 @ 11-11-12-28

CPUID Validation Link: CPU-Z Validator 3.1



System 2: Intel Ivy Bridge (Dual Core) 2.4GHz with HD 4000

Core i7-3770K with half the cores disabled and downclocked to 2.4GHz (to emulate the i7-3517U for ultrabooks)

ASUS P8Z77-V Deluxe

2 x 4 GB DDR3-1600 @ 9-9-9-24

CPUID Validation Link: CPU-Z Validator 3.1



System 3: Intel Sandy Bridge i7-2637M (17W Dual Core Ultrabook) with HD 3000

Dell XPS 13 Ultrabook

2 x 2 GB DDR3-1333 @ 9-9-9-24

CPUID Validation Link: CPU-Z Validator 3.1



System 4: Intel i5-450M (Year 2010 Dual Core) with Radeon HD 5650M (D)

Acer TimelineX 3820TG

2 x 2 GB DDR3-1066 @ 7-7-7-20

CPUID Validation Link: CPU-Z Validator 3.1



Notes:

We regret that our roundup does not include an equivalent Llano system for comparison - all our FM1 chips and motherboards decided to become expensive paperweights in the form of multiple bent pins and blue smoke from VRM on one motherboard.

All systems were running Windows 7 SP1 (with all patches as of 12/5/2012, core parking/scheduling fixes included), latest possible drivers for all hardware components and power profile set to performance to extract maximum potential out of each test system.

[siopao1984]

Synthetic Benchmarks - CPU and Memory Subsystem

AIDA 64 v2.30.1900



Despite using faster memory (Dual Channel DDR3-1600 @ 11-11-12-2, the Trinity A10 gets blown away by the Ivy Bridge and Sandy Bridge test systems in the synthetic memory tests by an embarassing margin, which says alot about the memory controller efficiency on the AMD chip. It did well in some integer (Zlib) and hashing tests but got let down badly by its archiac floating point engine implementation.





Sandra 2012 SP4



Another benchmark run with similar conclusions, with Ivy Bridge taking most of the accolades this time





Nuclearus Multi Core v2.0.0 RC1



This benchmark seems to favour higher clock speeds (the Sandy Bridge has a high turbo frequency) and IPC, and once again Trinity has some multi-core integer wins.

[siopao1984]

CPU Benchmarks



3DMark 11 Physics Test (Futuremark - world's most popular benchmarks and PC performance tests)

• Rigid body physics simulation with a large number of objects
• Scales linearly with core count and frequency

Cinebench R11.5 (MAXON | 3D FOR THE REAL WORLD: CINEBENCH)

• Based on CINEMA 4D
• Shows near linear scaling with CPU clock speed.
• Performs a fair amount of floating point operations.

Deep Fritz 12 (Chess News, Chess Programs, Databases - Play Chess Online)

• Test of branch prediction/cache hits and IPC

FRYRENDER one SP2 (Technical Specifications)

• Photo-realistic render engine where all the elements involved in the generation of the final image (materials, lights and cameras) are based on physically accurate models
• Test of SSE/IPC

LuxMark 2.0 (LuxMark - LuxRender Wiki)

• Rendering engine based on SLG2 with Metropolis Light Transport, Multiple Importance Sampling, Image reconstruction done on the GPU, etc;
• Test conducted with ~2,000,000 triangles (Complex Room) setting

UCBench 2011 (UnRAR-crack benchmark)

• RAR password cracker
• AES/SHA-1 units for decryption and support for SSE4/XOP instruction sets
• Scales with clock speed and cores

[siopao1984]

Productivity Tests



PCMark 7 1.04





Peacekeeper (Peacekeeper - free universal browser test for HTML5 from Futuremark)





SPECviewperf 11







Description of application tests (taken from SPEC.org):

The snx-01 viewset is based on traces of the Siemens NX 7 application. The traces represent very large models containing between 11- and 62-million vertices, which are rendered in modes available in Siemens NX 7.

The tcvis-02 viewset is based on traces of the Siemens Teamcenter Visualization Mockup application (also known as VisMockup) used for visual simulation. Models range from 10- to 22-million vertices and incorporate vertex arrays and fixed-function lighting.

The sw-03 viewset was created from traces of the graphics workload generated by the Solidworks 2009 SP2 application from Dassault Systemes.

The proe-05 viewset was created from traces of the graphics workload generated by the Pro/ENGINEER Wildfire™ 5.0 application from PTC. Model sizes range from 7- to 13-million vertices.

The maya-03 viewset was created from traces of the graphics workload generated by the SPECapc for Maya 2009 benchmark.

The ensight-04 viewset represents engineering and scientific visualization workloads created from traces of CEI's EnSight 8.2 application.

The catia-03 viewset was created from traces of the graphics workload generated by the CATIA™ V5 R19 and CATIA V6 R2009 applications from Dassault Systemes.

The lightwave-01 viewset was created from traces of the graphics workloads generated by the SPECapc for Lightwave 9.6 benchmark.

[siopao1984]

3D Benchmarks and GPGPU compute

3DMark 11 Graphics (Performance Preset)





Cinebench R11.5 OpenGL rendering





ComputeMark DX11 Compute Shader Test





LuxMark 2.0 (Room) OpenCL GPU only

[siopao1984]

Gaming Performance (DX9 and DX11)



Battlefield 3







F1 2011







Left 4 Dead 2







Alien vs Predator (DX11)





Street Fighter 4 (DX9)

[siopao1984]

Conclusion



In order to give a satisfactory conclusion, let us take a look at the models in the Trinity mobile APU family again:




The top of the line and (advertised as) quad-core 35W 'Trinity' A10-4600M SKU that we've just gone through in the last few pages struggled in most of our tests to keep up with mid-range 17W Sandy Bridge / Ivy Bridge dual-cores, sometimes yielding even to the middle aged 2010 35W Core i5 Arrandale. 3D graphics performance was hardly a relavation either, with the integrated 384 Radeon Cores barely matching that of a 2-year old Mobility Radeon 5650M (again really not top drawer, think desktop 5570) and Ivy Bridge's HD 4000 coming uncomfortably close in some titles like Battlefield 3 and Left 4 Dead 2. We've seen that Bulldozer/Piledriver benefits from compiler optimizations to take advantage of new instruction sets (XOP, FMA, AVX) and AMD heavily touts the "emergent experience" that comes with a handful of OpenCL accelerated applications - but the fact remains that single threaded applications and raw memory/FP performance is abysmal even with the aggressive turbo schemes, and any non-proprietary GPGPU efforts should benefit AMD's competitors anyway.

When attempting to survive a nuclear blast, do not hide in a refrigerator.

At press time AMD claims that "entire A-Series systems will start at prices less than the cost of just the high end Ivy Bridge i7 CPU" (lulz really, some i7s have real quad-cores you know) in the range of USD$699. We can safely presume a regular mechanical hard drive and bog standard screen/brick form factor/battery life for that price. Take all the benchmarks of the A10-4600M in mind and then imagine the lesser SKUs with half the CPU modules (17W A6-4455M Ultrathins against i5/i7 IB UltraBooks) and fewer Radeon Cores/Clocks... is the Piledriver powered mobile Trinity really good enough for mainstream consumers in the year 2012 or has AMD jumped the shark? Let us know in the comments section below!