Just back from my third SC conference, SC09, in Portland where I was able to confirm Doug's observation that the word 'Supercomputing' no longer figures on any of the signs, banners and conference material. It's like when Kentucky Fried Chicken went to KFC, but I'm still trying to figure out what what's so bad about the words 'super' or 'computing'. Regardless, two words that were anything but denigrated at the show were 'cloud' and 'GPU', which I'd argue were the two most prevalent themes of this year's show. I'll stay more down to earth, i.e. not in the clouds, with this blog, and focus on the GPU side of things, which is obviously of greater interest to Acceleware, its customers, and hopefully its blog readers.

The first big observation is that NVIDIA's marketing machine continues to amp up the rhetoric. There was a certain swagger in many of the employees words and actions, no doubt founded on the belief that GPU computing is at the dawn of more mainstream adoption. The recent announcement of the Fermi Architecture, NVIDIA's new GPU, does indeed give them reason to believe. ECC memory and pipeline, more device RAM, larger and more flexible caches , 8x the double precision performance, MIMD, full C++ support, and many others are indeed significant advancements, and will benefit many, Acceleware included. The other and potentially more telling advancement is the proliferation of people actually using and supporting GPUs.

12% of the papers in the technical side of the conference were about GPU computing. At least a dozen vendors featured GPUs as part of their hardware or system offerings, including Dell and HP. And let's not forget AMD, who actually have a production GPU with Fermi-like performance today, and who announced a GPU/CPU hybrid that debuted at number five on the recently updated Top 500 list. That is a first for GPU/hybrid computing. Mellanox, (Infiniband) company also announced special support of GPU configurations.

A more significant level of companies supporting applications and application developers have sprung up as well. The Portland Group with their Fortran / C compiler that supports the CUDA architecture. CAPS, a young French company, has also added complier-level support of CUDA and GPUs. TotalView and Allinea with debugging tools round out that group. All of these companies are focused more on the tool side of GPU programming. Their customers are developers that are actually writing (or porting) parallel code for GPUs. Acceleware is effectively a potential customer for all of these companies as we have a very high concentration of parallel GPU programmers.

Closer to the applications side were Accelereyes and WiPro, both of whom I met and spoke with at the show for the first time in person. Accelereyes, who assure me that the choice of name was purely a coincidence, continues to port the over 1200 Matlab functions to use their GPU-enabled data types and extensions. WiPro, the largest IT services company in India, has a fledgling group with a mandate to port applications to run on GPUs. Both Accelereyes and Wipro are more focused on the actual code-porting itself and in that sense are more similar to Acceleware. We spoke at length with both these companies discussing opportunities to work together or partner on various projects which I hope will bear fruit in 2010. As alluded to above, each of the above companies knew Acceleware before the show, and showed a clear respect for our experience and active customer base. It was strange to be seen as an industry veteran after only five years doing this work.

Now despite the abundance of companies active in the GPU space, there still seemed to be a dearth of mainstream applications and marquis customers announcing GPU-enabled codes. I sense however that situation will change in 2010. We met with several large ISVs, companies, and computing consortia that have been waiting and observing the GPU computing evolution. Many of them have now realized that they need to get off the fence, and make a decision about adding GPU support to their portfolio. Acceleware is in an interesting position in this sense. We have hundreds of active end users that have been using GPU acceleration for several years. We have seen the impact that it has had on their workflow and the value it has added to their simulation tools and design process. We have seen the evolution of engineering simulation that acceleration has enabled, and it's amazing. We also know that porting code is non-trivial, and not all applications will see the same speed ups on GPUs, some run better on multi-core CPUs (which has been part of our repertoire since day one), and some will always be serial. One aspect of our company that I'm very proud of is that we offer all our customers and potential customers an unbiased and realistic view of what can be achieved with their codes in combination with our libraries and all or any of the parallel hardware available. Our recognized expertise and experience as parallel developers becomes extremely valuable to customers trying to make a decision on parallelization.

On the topic of parallel hardware, the show was a good opportunity to strengthen our relationship with all of the players: NVIDIA, AMD, Intel, HP, Dell, and others. Understanding the offerings and direction of each of these companies helps us with our planning and ensures that all existing and future customers get the benefit of any new advancements on the hardware front as soon as they are available and with little to no effort. It's a complicated and rapidly evolving world, and we are again uniquely positioned to add value for our customers.

In conclusion, there are many stars aligning for Acceleware in 2010, including a great impetus to unlock the potential of GPUs and parallel hardware for the mainstream user. It will no doubt be a telling year for Acceleware and the industry in general, and we all look forward to the opportunity and challenges ahead.

Comments

Hello Mike,

Thank you for this interesting post. It completes Rob's previous answers to my previous questions.

About the terminology people have been using recently, it also seems that the word GPGPU has become obsolete, although still used by the people who started following this trend in '05/'06 and in some tech papers such as the Inquirer and Fudzilla. In my own opinion, it converges with the "stream processor" term used by both nVidia and AMD. The GPU becomes a co-processor and might in the future be the most important part of supercomputers.

I would like to point out that most of the GDDR5 DRAM features such as ECC have been used for now almost 2 years by AMD, nVidia not bringing anything new to the table on the memory interface. The architecture changes brought by the Fermi architecture are promising but still need to be demonstrated. The nVidia announcements have been made just to remind the public that new products are coming around but so far, nothing has been made public in terms of performance benchmarks. If the changes brought are as radical as what brought the G80 family, I am certain it will reinforce AXE's postion on the massively parallel programming market. Has AXE been able as a privileged nVidia partner to have a preview of the capabilities of Tesla Fermi enabled platforms?

It is great to see that a lot of people are developing Fortran compiler because it seems that a big chunk of the code used by companies has kernels written in such a language. Would it make sense to AXE to absorb or develop a core team with tremendous Fortran compiler experience? The software tool chain would be then complete for AXE and would allow control the programming flow's performance from bottom to the top.

2008 should have been AXE's breakout year but this did not happen because of the economic meltdown and the global recession. Mike is somehow sensing that 2010 could be the turnaround year. What is the growth expected in terms of revenue development? What is the trend in AXE's customer portfolio development (flat or up, if up by what percentage)?

AXE'stock dropped sharply in 2008 when the announcement was made that the company was laying some personnel off. What is expected by AXE in terms of stock price evolution/liquidity (back to pre summer '08 price/more volume)? What is the trend in AXE's personel development, by how much does AXE want to expand in 2010?

Thanks for your answers.

David/John

# Posted By david | 12/19/09 3:08 PM

Hi David/John,

Thanks very much for your comments.

On the topic of ECC, as you state, AMD does has available hardware that supports ECC (not quite in their Firestream line yet, but close enough) for some time now. The difficulty in their programming model as well as lack of double precision support in OpenCL are still roadblocks that have prevented us from moving more aggressively to support AMD hardware. These issues are being addressed by AMD, and when they do, that certainly makes AMD Firestream a significantly more attractive platform to target with our software / solutions. I am excited to see how this plays out in 2010. From NVIDIA, we were hoping to get a Fermi-based Tesla as a nice stocking stuffer, but they are still in extremely limited supply, even at NVIDIA. We are still on NVIDIA Santa’s ‘good’ list though for early access to the hardware.

I can tell you we have no plans to develop a Fortran compiler – our expertise is in building libraries, but I can’t comment further on financials or outlooks for 2010 on the blog. If you have further questions on those topics, I would direct you to investors@acceleware.com.

All the best to you (and your Acceleware investment) in 2010!

# Posted By Mike | 1/7/10 9:31 AM

Mike, out of curiosity, how may people typically attend the CUDA training seminars that you hold?

# Posted By Grant | 1/14/10 10:33 AM

Sorry Mike, I meant how many people usually attend the CUDA traning seminars

# Posted By Grant | 1/17/10 6:06 PM

Class sizes vary depending on the location of the course, but we restrict our classes to 7 students to ensure a high quality learning experience. In some cases, we do onsite courses for companies where the classes are even larger and we may have more than one instructor in the room. Since the exercises are very hands on and work with advanced concepts, we want to encourage a lot of interaction with the instructor.

# Posted By Chris Mason | 1/19/10 1:27 PM

Nice post Mike! I am in my Masters right now, learned lot from last 1.5 year about computing :), and always interested in GPU computing but never get a chance to work on them...Now days in my spare time I work on most of the checking bench mark for BLAS libraries....I am just wondering if there is any discounted GPU course available for students from university..I am really interested in attending one

# Posted By Akshay | 4/8/10 2:02 PM

Hi Akshay,
We definitely have academic pricing for our CUDA courses. Send us an email to services@acceleware.com and we can give you more information!

# Posted By admin | 4/13/10 10:44 AM