I have received a couple of emails off blog about my post yesterday about the Cell processor and its application to scientific computing.
The basic premise is one of scepticism. The hot area of interest up to a couple of years ago was Field Programming Gate Arrays. Nowadays a lot of discussions focus on the advantages of GPUs. However, the majority of chemists have not even heard of these processors and they remain of interest to programmers and hardware hobbyists and experts. For the chemists we spoke to at Bio-IT the terms FPGAs and GPUs went over their heads. Not true for the IT people. When we mentioned the Cell processor then it went over the heads of MOST people. So, the Cell is pretty much an unknown entity to most.
People have been programming onto FPGAs for years but none have gone mainstream in the scientific computing world that I am aware of. A number of researchers are now working with GPUs but have any gone mainstream and will they? The Cell might just be different.
So, a question for the readership. What are your thoughts, comments, opinions on FPGAs vs GPUs vs the Cell processor. Where does each have strengths over the others? What do people think about the future of GPUs in terms of scientific computing? What are your thoughts about the Cell processor?
; void 0;){kind=small}
There are currently 7 responses
[...] Williams over at the ChemConnector Blog has had a couple of people ask him for comments about which way to go and which one is better for a particular application ? We’ve just [...]
I’ve discussed this in detail on my blog (http://www.simbiosys.ca/blog/). It’s a long post so I didn’t paste it in here. Visit this post to see my detailed comments: http://www.simbiosys.ca/blog/2008/05/03/the-fast-and-the-furious-compare-cellbe-gpu-and-fpga/
The single sentence version: The FPGA is great for logic decisions and integer calculations, e.g. to be used for DNA sequence analysis, fingerprint searching, on the other hand GPUs and the Cell are better for floating point calculations, 3D modeling, docking, MM calculations, while QM can only be done on the Cell. There are also a number of subtle advanatges of the Cell over the GPU (see tdetails in my blog post).
ZZ
I’m still wondering what the advantage is of the soft microprocessor with programmable logic over the traditional types, at least for these types of applications.
On a side note, don’t they have a Center of Competence for the Cell processor at Georgia Tech where they develop and promote support for the processor? Are they doing development for biological and chemical operations or are they strictly with raw computing?
As for chemists we should not be surprised that FPGA, GPU, Cell processor go over their heads. (See original blog by Antony) They will not even know if their PC runs on an Intel chip or an AMD. And they should not care. What they care about is the utility of the technology. Can they do better science?
The reason I am impressed with the effort of Zsolt Zsoldos at SimBioSys is because the work is a step in this direction. We can do a better conformational sampling and this leads to better quality results in terms of better geometries. The reason that it is only a step, but not a solution, is because chemists care about better enrichment factor and not better geometry. Better geometry will not directly translate into better enrichment factor. The missdocked molecules will have more realistic ranking, but other than that the error in ranking and enrichment will still be controlled by the free energy error. This error is simply so large that improving the geometry in terms of RMSD will not make any significant difference in free energy estimates. The next challenge is to use this computing capacity offered by the Cell architecture to produce a better scoring function or calculate real free-energy estimates. In order to separate actives from inactives we need to separate confidently 3 orders of magnitudes in IC50. This means we must reach much better than 5 kcal/mol error in the free energy. Current scoring functions are far from this goal.
I hope the eHits (by Simbiosys) development goes in this direction.
See more comments on FPGA vs. GPU vs. Cell on http://www.simbiosys.ca/blog/2008/05/03/the-fast-and-the-furious-compare-cellbe-gpu-and-fpga .
I think that by the introduction by Apple of Snow Leopard in 2009, people will be more clued up due to Grand Central amongst other things. nVidia knows it’s got competition. If GPU as CPU can occur, then I’d imagine it’s going to force IBM, Samsung to keep on working on the chips. Just look at Road Runner - upgrading and modifying the Cell chips, and create some very tasty nodes.
For me, the possibilities from Snow Leopard are rivalling it though - in terms of sheer amount of RAM you can throw at it, and the promises of mainstream performance boosts for more generic multi-core Intel systems.
OH~ it’s like great!
mmm gpu, cell and fpga. gpu processors are very fast due to the comercial amount of funding, the race is on to get faster and faster. The cell processor is a very interesting processor but IBM has made it so hard to get a hold of in the market outside of what they want to sell it to.
The cell processor is just not available enough for it to be productive, to be honest looking at the price of intel processors there not even viable anymore. At a cost of £1100.00/processor its stupid cost for that performance at that price.
FPGA offer such and amazing attractive package in 45nm and 40nm silicon metal gate providing much faster more density and amazing configurations to a developer and scientific persons. Above all reconfigurable and debuged to any spacific solution allowing outside the box development. The keys to the farrari are in the users hands.
With FPGA configurations data protocols are as fast as 100gbs and can be configured as and when needed, power modes can be configured as needed, processor cores added how ever many you want within constraint of the transistor limit of course. Powerfull SPU’s that the cell processor run on can be designed and programed by the user on an FPGA unlike being stuck with n amount n type.
GPU technology will always be fast its the high amount of funding research development teams not to mention the creative engineers who design these processors. They will always design more and more in this field as they get paid high, very high wages. Yes money fuels this area more than others.
Scalability is a very important part in this field but flexability is far more important, something intel, ibm and others seem to have forgoten about. I must add that sun micropsystems have paid attention to this area well mind you so one of them understand.
Leave a reply
Trackback this post | Subscribe to the comments via RSS Feed