The Green500 List ranks of the world’s most energy-efficient supercomputers. This year, the AMD FirePro S10000 has taken the SANAM supercomputer to #2.
AMD worked with the University of Frankfurt’s Institute for Advanced Studies (FIAS) to research quantum chromodynamics. Powered by 420 AMD FirePro S10000 dual-GPU server graphics cards, the SANAM supercomputer can sustain 420 TFLOPS, providing a system energy efficiency of over 2.3 GFLOPS per watt and performing 2,351 million calculations per second per watt.
Overall, the performance of machines in the Green500 List has increased at a higher rate compared to power consumption. “That’s why the machines’ efficiencies are going up,” says Feng. “We are more performance for the same amount of power.” For machines based on commodity components—machines built with off-the-shelf components—coprocessors and GPUs are attributing a great deal to the efficiency gains. So much so that they are keeping pace and in the latest list even outpacing purpose built systems like IBM’s Blue Gene/Q.
Consuming a total of 3.6 kilowatts, the folks down in the basement lab at AMD have built the fastest multi-GPU compute server using eight of their flagship compute powerhouse boards, the new AMD FirePro S10000!
This 16 GPU (eight S10000s) Exxact Computing Server provides more than 8 TFLOPS of real world double precision computing performance. While these are early drivers, this still means you are still seeing around 70% efficiency of the theoretical peak double precision floating point performance of 11.84 TFLOPS (47.28 TFLOPS peak single precision performance!).
Here’s a look at the DGEMM result and a glimpse at what’s under the hood of the compute server:
Today both AMD and Nvidia announced their new flagship HPC server GPU solutions. AMD released the FirePro S10000 (pdf spec sheet), a dual-GPU card that combines both compute and 3D visualization on a single card. Nvdia announced their Tesla K20 line which target compute only (you need a Quadro card for the 3D visualization capabilities). Below is a chart comparing the specs and performance of the new cards, as well as comparisons to Nvidia's Tesla K10 and compute-specific M2090 cards
AMD FirePro S10000
Nvidia Tesla K10
Nvidia Tesla K20
Nvidia Tesla K20X
Nvidia Tesla M2090
Core Clock Speed
Peak Single Precision
Peak Double Precision
FirePro S10000 is a single card that delivers both compute and professional 3D graphics, (as opposed to needing a Quadro plus Tesla). The FirePro S10000 can handle 2 compute & 1 graphics operation simultaneously
FirePro S10000 is up to 7.8X as fast as the K10 for peak double precision and 1.3X as fast for peak single precision
FirePro S10000 has up to 1.5X the memory bandwidth as Tesla K10
FirePro S10000 can drive up to 5 high resolutions displays vs Tesla (oops you need a Quadro to drive any displays)
Puget systems tested AMD and Nvidia consumer (Radeon &s GeForce) and professional (FirePro & Quadro) cards for performance in Autodesk Maya 2013.
From the review:
“Our benchmarks clearly show that the AMD FirePro cards are the top performers in Maya 2013… It’s rare to see this large of a performance gap between difference series of cards, but our benchmarks clearly show that the AMD FirePro cards completely trounce all of the competition.”
“Unlike other software like AutoCad or Premiere, you will be giving up a lot of performance if you decide to go with a desktop card from the NVIDIA Geforce or AMD Radeon series. Of course, it you are primarily using software that does better with those cards and are only lightly using Maya 2013, those cards may still be the better choice. But if you are only concerned about Maya 2013 performance, you simply can’t beat the performance of the AMD FirePro W-series cards.”
Legit Reviews tests the new AMD FirePro W5000. both with standard benchmarking tools but more importantly in a real workflow.
From the review:
“Benchmarks only tell you half the story. What really matters when it comes to professional graphics solutions is software support. After all, these solutions are expensive and if they don’t deliver when they should its extremely problematic. Therefore we tested the AMD FirePro W5000 and Radeon HD 7970 in a typical 3D workflow. We started by creating a base mesh in Autodesk Maya 8.5, 2009, and 2012 in order to test compatibility across multiple versions of the software. From there we did some UV unwrapping so texturing could be started. After all that was accomplished, we exported the mesh for use in Autodesk’s Mudbox 2013 where a high res sculpt was done on the base mesh. After which the final mesh had its subdivision level reset to zero and we exported it again. After doing so, the final normal, ambient occlusion and displacement maps are generated and brought into Autodesk Maya 2012. We applied the maps to the base mesh we exported and followed it up with further tweaking until we were satisfied. Through this process we really get to see if the AMD FirePro W5000 offered a better user experience over the consumer oriented Radeon line, and let me tell you it is a night and day difference.”
“The fact the FirePro W5000 from AMD performed so well in old and new applications, as well as its attractive street price of just $448 shipped, it truly is an outstanding value in the professional graphics market.”
3D Artist magazine and Develop3D both review the new, wallet-friendly InterPro IPW-iB Core i7 CAD workstation. The 3D Artist review is from a shootout of 4 workstations under £1,200. The Develop3D is one of their ongoing series of workstation reviews.
From 3D Artist: Out of the four systems that we tested, three had near-identiucal configuration. The InterPro IPW-iB was the only ssystem to use an AMD graphcics card ... and it shows. In most of our tests, the FirePro W5000 GPU enabled the InterPro IPW-iB to storm ahead of the pack. The SpecViewPerf Maya benchmark and CINEBENCH OpenGL scores were almost double those of other workstations.
From Develop3D: The IPW-iB’s secret weapon is the FirePro W5000, one of AMD’s new Southern Islands Graphics Processing Units (GPUs). The FirePro W5000 plays its trump card inside PTC Creo 2.0. The graphics score of 5.85 is head and shoulders above anything currently possible with Nvidia Quadro GPUs. AMD’s close co-operation with PTC on developing the Creo 2.0 graphics engine seems to be playing big dividends here.
AnandTech has posted The AMD FirePro W9000 & W8000 Review: Part 1 covering the two high-end Tahiti GPUs of the new FirePro W-series family. The article covers both the technical details of the card as well as thier perception of how AMD is increasing their market share by embracing 3D performance, compute performance, stability and open standards.
The article has an excellent overview of the GCN architecture in terms of why it matters to the professional market - “an architecture that’s significantly better at compute without sacrificing graphics performance, and this is why the resulting GCN architecture is balanced for both compute and graphics.”
It also discusses other defining features including: Error Correcting Code (ECC) memory support, Partially Resident Texture technology, PowerTune (power throttling based on thermal tolerance), PCI-Express 3.0 support, and Video Codec Engine (VCE) hardware H.264 encoder,.
In part II, AnandTech will focus on performance and benchmarks (hopefully real world rather than just synthetic).
Tom’s Hardware reviews the new AMD FirePro W9000 and W8000 cards testing them on Maya, AutoCAD, some synthetic benchmarks (SPECViewperf 11), OpenCL performance and games (I’m not getting the game testing on pro cards) They compare the these results to those from Nvidia Quadro Fermi boards.
In Maya 2013, both the W8000 and W9000 outperformed the Quadro 6000 by impressive margins (basically showing strong price performance advantages). Same for OpenCL image processing and benchmark (2X to 4X) tests. Same on AutoCAD 3D performance (which is not CPU-bound). Same for DirectX performance (while you generally don’t buy a workstation graphics card to play games, it is notable that this is the first time I have seen pro cards have equal or superior gaming performance over the consumer cards).
The Quadro Kepler line faired better in the synthetic SPECViewperf 11 CATIA test (would be nice to see real application performance tests) as well as OpenCL video processing tests (this one surprised me). The FirePro line outperformed significantly in the SPECViewperf 11 Lightwave and Solidworks sequences, as well as Siemens NX and Ensight with higher levels of AA sampling (again, would be nice to see real work application performance).
More than one commenter noted these tests used a low end CPU with a high end GPU, so some of the tests (such as AutoCAD 2D), were CPU-bottlenecked and didn’t show realistic scaling or performance.
Cut to the chase purchase recommendation: “It’s not difficult to give the FirePro W8000 a thumbs-up based on its price to performance ratio. The W8000 costs a lot less, operates more efficiently (and quietly), and performs almost as well as AMD’s FirePro W9000. Yes, it’s a little strange to give a $1600 graphics card a value-oriented recommendation knowing that most folks have no need for such an expensive piece of hardware. However, for the professional users who can’t compromise on points like validation, performance, or support, $1600 can quickly pay for itself in recovered productivity. If you don’t need the FirePro W9000’s extra performance, AMD’s FirePro W8000 should make a great long-term workhorse in your workstation.
3D World does a quick review the Nvidia Quadro 2000 ($480), Quadro 5000 ($1956) and AMD FirePro V7900 ($764).
Some quick summaries:
Nvidia Quadro 2000
“It doesn’t quite compete on performance, but it’s amazing value for a budget workstation”
Nvidia Quadro 5000
“... ultimate performance for merely high-end prices. It does cost twice as much as the Quadro 4000 however, so it’s best partnered with a similarly premium workstation specification.”
AMD FirePro V7900
“The FirePro V7900 goes head-to-head with Nvidia’s Quadro 4000, yet costs at least £100 ($158) less. It sports 1,280 Stream processors, 2GB of GDDR5, and a bandwidth of 160GB/sec – which is more than any current Nvidia Quadro model. This gives it some impressive performance results, making it a very tempting competitor to Nvidia’s high-end offerings.”
This video demo from SIMULIA 2012, shows a 2.9X acceleration of the Abaqus S4b direct solver benchmark unning on an AMD FirePro v9800 using OpenCL. The benchmark simulates bolting a cylinder head onto an engine block with 5,000,000 DOF. The nonlinearity in this problem arises both from changes in the contact conditions and yielding of the gasket material as the bolts are tightened. The job is accelerated using both the CPU and GPU and then passed to the postprocessor for the final simulation.
A tweetable take-away: A 4-core CPU + FirePro GPU using OpenCL finishes the direct solver 2.9 times faster than running just on a 4-core CPU alone, and 2.2 times faster than running on an 8-core CPU.
Below is a chart showing some raw data comparing multi-core CPUs + FirePro GPU using OpenCL vs multi-core CPU only.
(Note: the GPU-acceleration bump decreases with more CPUs because Abaqus v6.11 splits the job to more CPUs and the GPU ends up doing less work)
FireUser.com is a community resource for CAD, visualization, 3D, video and engineering professionals to learn about the latest acceleration and display technologies and news with a focus on the AMD FirePro workstation graphics line.