Today’s computer games deliver 3D video sequences in photorealistic quality. To do this in real-time, the hardware industry developed high-end graphical processing units, also called GPU. A GPU has unbelievable processing power. Instead of 2, 4 or 8 processor cores as known from the traditional Intel/AMD CPU the GPU uses an arrays of hundreds of parallel floating point processors to compute images in their internal graphics memory.
What value does this bring to finance people? The answer is simple: When doing analysis, planning, budgeting, forecasting, scenarios or reporting a lot of number crunching happens, especially if you are looking at aggregated and multidimensional OLAP data models as we usually do in Business Intelligence or Corporate Performance Management. Number crunching consumes enormous processing power. The number one complaint about BI and CPM software is slow query performance, as BI and OLAP Analyst Nigel Pendse points outs.
So our Palo researchers had a look at the GPU hardware architecture and discovered that GPUs are the perfect hardware accelerator for in-memory OLAP server like Palo. They expect a performance increase by a factor of 20 (not 20%) at least. This would be a performance breakthrough that has never been seen before in the BI industry. The reason why this works so well is the fact that Palo uses an in-memory technology. Since today’s GPUs have 4 GB of Graphic memory it is possible to load the entire cube directly in the GPU RAM. So there is no bottle neck like disk IO etc. that would decrease the GPU power.
And it gets even better: We just had the GPU Technology Conference in San Jose. There NVDIA announced the Fermi Architecture. This new GPU technology is due in 2010 and will again increase the processing power by the factor of 5 (against today’s Tesla technology).
And by the way: Did I tell you that you can combine GPUs? Here at Jedox we run TESLA hardware with 4 parallel GPUs and 16 GB RAM in one server and it still scales almost linear. So this makes 20 x 4 x 5 = 400. A query that took 40 seconds to calculate on a CPU will be done in 0,1 seconds with GPU. Theoretically of course. Results in practice will be seen on CeBIT 2010.
Wow, Kristain, did you say “here at Jedox we run TESLA…?” This is very interesting. I’m no compiler developer, but somehow I thought Tesla technology required massively parallel and highly threaded code to deliver the punch. Please write, expanding on this!!
If true, it’s huge. A moderately powered PC, good GPU and one (or more) Tesla boxes could literally scale way above a higher end PC or multi core CPU!! In real terms an ordinary PC (US$800), good GPU (US$500) and a Tesla box (US$1,200) would rival any $10,000+ workstation.
On the general GPU front, I agree completely with you. Here’s a small CapEx case: I have three systems – my beloved Sun 40 workstation (omitted here) and two new PC’s. I demanded and received a PC with a high end Intel Core i7 cpu and run of the mill GPU. The thing cost well over US$3,500. Still unimpressed, a kid in IT built a much less expensive (say US$1,000) system at home for me using AMD’s gaming technology with 2 high end GPU’s. For heavily cpu-bound and in-memory applications with a lot of data visualization, the low cost system is always faster and more responsive!!
I encourage my CFO and Analyst friends not to take whatever IT issues “just because.” It will save countless hours and make computing much more enjoyable. Best regards, The Old Quant, CFA
Forget your Headline,
dear Kristian, sorry I have to say this, as a consultant I would say that I am sometimes a little bit too technical for my clients, and I have to admit that I could not read your artice because I did not understand it – BI is definetly not a technical issue.
It is not “the more the better” it is about processes, data reduction and information reliability. All this is not technical.
And this is where so called “small” BI-vendors are extremly weak, because they usually cannot handle big company projects with their workforce.
They stick with small and medium clients ( a fact which is not considered in the nice but useless charts above) where it seems that processes are not crucial.
The “Big” Vendor Software is much more process oriented and is in most of the cases the only really solution for tasks in the “big” companies.
While you are writing about Game-Controllers, they are implementing a 1000 Client planning application somewhere, not promissing it will be the fastest, not promissing that it will be done within 3 month but after implementation the planning process will be streamlined and the data will come out of a single source ( hopefully right). I spoke lately with a couple of CFOs of bigger Companies the where all together stating ” I do not care that much about waiting a couple of minutes for my report, what I do care about much more that the numbers in this report are 100% reliable “.
The correctness of “right” numbers is 99% depending on processes and 1% on IT -Tools.
Regards
Andreas,
A CFO might be ok to wait a few minutes for his report, the greater number of business users having to spend hours instead of minutes for entering data and doing detailed analysis will certainly not agree with their CFO’s speaking.
It is not a game controller, it is a processing card with hundreds of number crunching CPUs working in parallel.
As a consultant I have implemented Palo in a 600m turnover organisation, and am currently looking at potential applications in my Banking client.
There is little doubt that GPU acceleration has many people in IT excited and every IT department should be alerted to this robust development, as it significantly improves any business case by providing significantly higher performance for less than traditional cost.
In larger organisations, the technology choice is left to the IT department, while managers are interested in ease of use, data control, process flows, and reasonable application speed. Getting your ETL and processes right are critical for success, but speed can become a major issue whether you are faced with waiting for an hour, or sometimes overnight, for Budget or Treasury models to recalculate.
Operational staff need tools that will work quickly, if they are to meet that working day 3 reporting deadline. Try telling a manager he will have to wait 33 hours for his reports, instead of 16 minutes, just because each one of those 1000 workstations takes 2 minutes per process instead of 1 second.
The world is full of large scale, process compliant installations, that have been abandoned because the process worked, but the technology didn’t.
The key is to ensure you address the technology issues as well as the process issues.
Well do not understand me wrong , I did not say that speed does not matter at all, but in the context of the whole blog, where it is about small software companies are pushing inventions, large companies are blocking it. And Jedox is understanding itself as a real inventor bringing added value to the Business. I have to say sorry -> Speed is already invented. Everbody delivers the fastest planning, MIS, PM software on the market ( just look at the benchmarks – each vendors leads one).
If you have to wait 32 hours for your data, there is something wrong – in infrastructure or more likely in the design.
I have seen SAP- BI solutions which where relatively small in scope and very slow in usage, but I have also seen SAP-BI Solutions which had a huge amount of data in it giving answers on every request send out within milliseconds.
Same technology different design – most likely, if you look at the majority of the big solutions they are providing a solid business value. Do not forget that most of the consultants for the smaller vendors (like Jedox) do just see a small part of the reality – which means if a project with the “BIG” providers flops the company will look around for a different solution and contact the “small” providers. JUST those will be seen by the small providers – which will now say “every SAP-BI Project is unsuccessful”. AND even wars “it is because of the software used”. Of course they have to say this because they have to sell there expensively invented tools and of course the have to say it because they are not qualified in modifying or redesigning the existing application (SAP).
And since having only a few customers and relying on them – they will do the same design-mistakes because they just look at the data not at the process and then modifying the existing software in the way that it will be able to support the (false) design – by speeding up the DB-Engine – since they have the power and the need to keep the customer. Great invention !
Another point is that the smaller software developers are mostly relying on one single technology / architecture approach – here it is OLAP, the solutions with customers always will be OLAP solutions which will be OK in most of the cases in the beginning of the projects when everything is pretty clear and structured. But growing and changing demands during the times will need to over think the BI – architecture.A vendor relying on payback of the investments in devoloping his only database and having his complete consulting team educated in his own and only software will always do very hard to recommend a switch. On the other side a BIG Vendor having all technologies in his portfolio will generate income with any suggested architecture approach ( I am not saying that those suggestion are always the best).
Closing this already to long statement I will take a metaphor from a different branch
It is not about inventing faster cars anymore it is about transporting peaple and goods in a save, ecologic, economic way.