While the second quarter earnings call of NVIDIA last week was largely without any big news or surprises with the future of the company, as always there are some interesting things, which we’ll like to cover here.
Regarding 28nm manufacturing at TSMC, Jen-Hsung Huang pointed out that the company is far better prepared for it than they were at a similar point regarding 40nm manufacturing. Due to the success with the prior nodes, he claimed the whole industry had underestimated the difficulties of process technology advancement. As a consequence NVIDIA now dedicated an entire organization to oversee advanced nodes.
Currently, the company is said to have working 28nm test silicon and about to start volume production. The CEO appears to be rather confident about it. The first 28nm product codenamed Kepler is said to hit the market in early 2012.
Kal-El is not only supposed to be a top performer, but also very power efficient. Huang highlighted that Kal-El is more power efficient than Tegra 2 under any scenario according to internal measurements. NVIDIA uses a technology they call Variable SMP (symmetric multiprocessing), that allows the chip to deliver high performance when needed and otherwise conserve power. We’d like to note that it could very well be, that under peak load it uses more power, but since it can process the workloads faster, it could go back to sleep mode more quickly. How it plays out in practice remains to be seen, once first products based on the chip can be reviewed.
Another interesting thing regarding the Tegra products is that NVIDIA now publicized that they are working on a SoC with an integrated modem. When they announced the acquisition of Icera, they were a bit cautious regarding such integration, but now it’s part of the roadmap. NVIDIA also hopes to make some money off LTE MultiMode modems, which are supposed to launch in the first quarter of 2012. NVIDIA is quite proud that their 3G/4G modems passed 100% of the applicable tests.
When it comes to Tesla products, the CEO explains, that it is still a small, but growing business. At this point it is largely dependent on demanding software being ported to CUDA and/or OpenCL, which is a complicated and expensive effort.
Make no mistake: this server runs on NVIDIA silicon only, courtesy of four Tesla M2050 GPU Computing Cards
Towards their competitor AMD Huang notes (without specifically mentioning them), that the chip alone is not enough to be successful in this business. The software stack and proper developer support are paramount to be successful in this arena, something that NVIDIA did with their CUDA API and dedicated developer support. NVIDIA considers itself to be the only company in this segment. While the company has a huge head start, this advantage may start to crumble once OpenCL gets wider industry acceptance.