This month, NVIDIA officially announced its decision not to buy British processor architecture developer Arm. All efforts of the first companies to create acceptable conditions for the regulators were in vain. At the same time, NVIDIA promises to actively develop a family of its own Arm-compatible processors, and the Server Grace is just the first in a series of such products.
At the quarterly reporting conference, the head and founder of NVIDIA Jensen Huang (Jensen Huang) said stressedthat the company holds a 20-year ARM license to develop processors and therefore promotes ARM-based solutions in many market segments, from robotics and automotive to server and supercomputing markets. In the current quarter, the company has to factor in a $1.36 billion penalty to SoftBank for failing in the deal to buy Arm, but in general, that failure doesn’t diminish NVIDIA’s determination to continue actively buying compatible processors to develop architecture.
NVIDIA’s CEO confirmed that the Grace server processor will launch in the first half of 2023 and will be just the first step in expanding the brand’s portfolio of ARM-compatible CPUs. Jensen Huang explained that he would definitely have numerous successors. As the NVIDIA boss added, in recent years market participants have been able to verify the effectiveness of using ARM architecture in data centers and cloud systems.
The delivery of processors of the NVIDIA Orin generation for active driver assistance systems has already started. In the future, the profile type of activity will be very important for NVIDIA, because only by the end of this decade, according to the company’s forecasts, the number of cars using processors of this brand will reach 10 million units. Starting in the second quarter of the current fiscal year, NVIDIA will receive income from the delivery of processors of this family.
In general, CPUs should become an equal category of NVIDIA products as GPUs and dedicated accelerators (DPUs) for artificial intelligence systems. As part of the platform development, NVIDIA does not refrain from using x86-compatible processors from third parties – the main thing is that the finished solutions contain the best chips on the market, as the company founder explained.
He added that CPUs are not the biggest bottleneck in server systems when it comes to performance scaling. The efficiency of this process is more affected by the bandwidth of the memory interface or the PCI Express bus. The transition to PCI Express 5.0 will begin later this year, as Jensen Huang vaguely promised, but eventually this interface will also reach its bandwidth limits. Support for PCI Express 4.0 will continue not only until the end of next year, but for at least the next two years.