Typically when there is little news surrounding Advanced Micro Devices (AMD) I use the time to look more deeply into questions readers ask and visit potential positives and negatives I have not seen touched on previously. As far as I know the only two major events scheduled for AMD prior to the next earnings call is the Citi Global Technology Conference and the unveiling of their upcoming GPUs.
In this article I will address the potential for AMD in HPC/GPU Compute scenarios and touch on an overlooked potential driver for increased gross margins in the 2014/2015 time frame. I will include a brief summary of each section at the end of each section in the *POINT* summary.
AMD In Work Stations and Servers
First the potential size of this market must be quantified in order to speculate on a potential impact to revenue and bottom line for Advanced Micro Devices. To do this, I will use Nvidia's (NVDA) prior quarterly reports along with the assumption that Nvidia owns the majority share of this market. I have used numbers from Nvidia since I am assuming they have the majority share of this market and Nvidia provides more granularity in their financial data, which should increase the accuracy of these calculations.
To begin, in Q2 2013 (last August for Nvidia), Nvidia's Professional Solutions Group (group that tracked Quadro and Tesla sales) posted revenues of $196M. In the same period, Nvidia's Consumer GPU business posted revenues of $668M. Since this time the accounting has changed slightly, and if we subtract the $66M royalty payments to Nvidia by Intel, this brings the consumer GPU business for Q2 2013 to a total of $602M.
In the most recent quarterly report released by Nvidia, they stated last year's Q2 total GPU business was $798M. The GPU business for Nvidia currently consists primarily of Tesla, Quadro, and consumer GPU sales; professional grade sales are not tracked as a separate entity anymore. The $798M is in line with what I have calculated in the above paragraph.
In August of this year, Nvidia reported total GPU sales of $858M, with revenues from desktop GPUs increasing 3.9%, notebook GPUs decreasing 1.7%, the Tesla line 127.5%, and Quadro line increasing 14.3% -- all YoY.
For ease of calculation I will assume the entire $60M increase of GPU revenues YoY comes from the professional grade graphics. Given Tesla plus Quadro revenues last year equaled $196M, and the increases between last year and this year (and using a system of equations), in the most recent quarter reported Tesla revenues were likely around $68M while Quadro Revenues were likely ~$188M.
An interesting fact here is that you can see Nvidia has experienced explosive growth in their professional grade GPUs. In the past 2 earnings calls, AMD has echoed similar sentiment by stating they have had 4 quarters of growth in their professional graphics business. This sector is growing fast enough to allow AMD and Nvidia to both increase revenues via their professional GPU lines.
There are two key things for AMD to be able to continue this momentum going forward. First they need to keep the press on with their open standards implementations via the HSA Foundation and HSAIL going forward. This is key because Nvidia pushes their proprietary CUDA programming model to take advantage of GPU compute, whereas AMD and Intel (INTC) push OpenCL. Nvidia does however support OpenCL as well. Therefore, right now Nvidia has a solid base built around their CUDA architecture, as well as a foothold in academia. Meaning that students that learn to program on Nvidia GPUs in college may be more likely to stay with Nvidia. If the industry shifts to adopt OpenCL, this can chip away at part of Nvidia's advantage.
The second point I am looking for is an improvement in performance/watt from AMD when they update their FirePro line of GPUs. Looking through a few benchmarks it appears Nvidia has the upper hand in this category.
Lastly, there were two headlines that received very little attention that I would like to bring up here. AMD had been collaborating with a company named PGI to work on an optimized API for AMD. This announcement is from July. A couple of weeks later, another headline cropped up stating that Nvidia had purchased PGI. To me, this illustrates two points. AMD is at least competitive enough to warrant attention from Nvidia. But it also demonstrates Nvidia's advantage from their cash position.
*POINT* The entire professional GPU market is most likely between $1.2B to $1.7B per year, if you assume Nvidia owns most of the market share. I base this on my calculation of Nvidia's most recent quarter, and provide a range in which I feel should accompany AMD's contribution. The key here is that this market is exhibiting good growth, especially among the server cards as HPC becomes more popular. In the near term I am looking for AMD to at least maintain their revenues static in the professional sector. Meaning in the upcoming quarter I would like to hear something in the range of "slight decline" to "5th straight record quarter." Longer term I would like to see continued 'HSA' momentum and increased software support from AMD.
Smaller Die Sizes Leading To Increased Gross Margins
Last year, AMD's CTO Mr. Mark Papermaster gave a presentation in which he talked about AMD's 'High Density Cell Libraries.'
During an interview with Rage3D, AMD's Mr. Jim Keller spoke about bridging the performance gap between AMD and Intel on the CPU front.
Whether or not the performance gap gets bridged is yet to be seen, but as an AMD investor the thing that most interests me is the reductions in area and power consumption from the use of these high density cell libraries. When AMD created 'Bulldozer' and 'Pile Driver,' they were laid out by hand. As you can see above, by incorporating both cell libraries and hand tuning, AMD should be able to achieve a balance of performance while minimizing die area.
As you can see above, Bulldozer has the worst transistor density of any 32nm process CPU, while AMD's 'Fusion' Llano CPU has the highest, at close to 2x the density of the 8C Bulldozer. Based on the various timelines of products, Mr. Keller's arrival back at AMD, and various leaks regarding upcoming products and TDPs, my 100% guess is that for the big cores, this could come into play around the end of 2014/beginning of 2015 when AMD unveils the successor to Kaveri.
To show the potential impact of the high density libraries combined with a node shrink, take the Cape Verde GPU for example. Cape Verde GPU is 123 mm^2 and contains 1.5B transistors. 1.5B/123mm^2 = 12.2 million transistors per mm^2. This is 3.2x denser than Bulldozer and almost 2x as dense as Llano, in a .5 node shrink (32nm to 28nm). For comparison, FX-8350 is 1.2B transistors in 315 mm^2 die area. Keep in mind that this is comparing a GPU to a CPU. I don't expect the big x86 cores to obtain the same density as a GPU, but I do feel there is room for improvement here, and this improvement could impact gross margins.
*POINT* As AMD transitions to Kaveri and later Kaveri's successor, the use of high density cell libraries could help AMD increase margins as they are able to squeeze more transistors into a given area. This will also allow them to reduce power consumption. These two things combined should serve to make them more competitive on a performance/watt basis, as well as helping with margins as this extra performance is squeezed into a smaller area.
I feel the potential for AMD lies beyond the 2013 time frame. I would personally like to see them at least not lose ground in professional class GPUs in the near term as we see if HSA gains traction. Further out (2014/2015), I am looking to see if they are able to increase transistor densities through the use of high density libraries, which should assist in both power consumption and gross margins.
Lastly, there isn't much news scheduled to be released between now and the next earnings call. I feel the Citi conference is up in the air as to the importance, given no stock moving news came out of Hot Chips. The only other event that I know of right now is the GPU unveil at the end of September.
Additional disclosure: I actively trade my AMD and INTC positions, and may add or liquidate positions at anytime. I am long AMD in both shares and options.