CHIP DESIGNER Nvidia believes that ARM based processors won't be ready for high performance server applications for at least two years.
Nvidia is itself an ARM chip licensee that has seen significant design wins with its Tegra 3 system-on-chip (SoC) processor, however the firm doesn't see ARM based servers being able to do heavy lifting in server tasks for two years. Sumit Gupta, GM of Nvidia's Tesla Accelerated Computing business unit said that even with GPGPUs, ARM based servers are not yet able to provide the computing power needed to drive high performance servers.
Gupta said, "Performance of these ARM cores is still not where it needs to be for servers. It is getting there; the new ARM64 [processor] is going to get it part of the way." However he did say that eventually ARM SoCs could hit X86-like performance levels. "One day I think ARM will at least get to similar performance levels as X86 performance. The belief is that over the next one or two years these ARM SoCs will be good enough for cloud applications and web serving. I think it will take some more time to be good enough for accelerated computing."
As for Nvidia using its Tegra chips to push work to the firm's GPGPUs, a scenario that would make the firm's accountants very happy, Gupta said he was surprised at the level of interest from developers and questioned the need for powerful CPUs. "We did a small development kit called Karma that has a Tegra 3 and a Nvidia GPU, [and] I was shocked by the number of those kits that have been sold. The interest in this ARM plus GPU is far larger than even I expected. If the GPU can do dynamic parallelism, it becomes more independent than how powerful CPUs do you need? I believe the first thing that will happen is that people will start using lower performing [Intel] Xeons [...] then at some point when these Atom based processors become available they might use that, and when ARM64 is available they'll use that."
Both Dell and HP have been pushing ARM based servers in the past year, and judging by Gupta's comments it seems that the first generation of such servers will be used primarily for front-end work such as web presentation or as part of a Mapreduce cluster dealing with small individual jobs. Nvidia is sitting pretty with its Tegra SoC and its maturing GPGPU technology, but judging by Gupta's comments it could be a few years before the firm gets to make use of these to finally stick two fingers up at Intel for not giving it an X86 license all those years ago. µ