Shares of Advanced Micro Devices Inc. rose up nearly 4 percent today after news emerged that Amazon Web Services Inc. is using the company’s silicon to power its latest cloud instances.

The chipmaker also unveiled what are hailed as the world’s first data center processors and graphics cards to feature a leading-edge 7-nanometer architecture.

New price-competitive instances

AWS said that the new cloud instances run on customized central processing units from AMD’s Epyc line but didn’t disclose the model. The virtual machines are organized into two families, M5a and R5a, that cost up to 10 percent less than comparable instances running Intel Corp. silicon.

The M5a series comes in six sizes that all target workloads such as application servers. The smallest configuration provides two virtual CPUs with 8 gibibytes of memory, while the most expensive m5a.24xlarge model packs 96 virtual CPUs and 386 gibibytes of memory. (A gibibyte, an alternate measure of memory storage size, is equal to 1.074 gigabytes.)

AWS’ new R5a machines, meanwhile, target analytics applications and other data-intensive workloads that require even more RAM than what M5a offers. The six available configurations provide the same number of virtual CPUs as their M5a counterparts but with significantly more memory: A single virtual machine can be provisioned with as much as 768 gibibytes of RAM.

“The world’s largest cloud provider offering Epyc is the biggest news of the day — a testament to Epyc’s capabilities,” Patrick Moorhead, president and principal analyst at Moor Insights & Strategy, told SiliconANGLE. “I was as impressed when AWS said it takes no software or script changes yet delivers a 10 percent cost reduction.” But Moorhead added that he still sees Intel continuing to be AWS’ main chip supplier.

Powerful new chips

Meanwhile, at an event in San Francisco, AMD refreshed its data center portfolio with seven-nanometer chips that will present a new challenge to Intel’s dominance of the server market.

On the CPU front, the company previewed the long-anticipated Rome processor series, the latest evolution of the Epyc line. The chips are based on AMD’s newest Zen 2.0 architecture and promise to provide up twice as much performance as its current CPUs.

The speed hike is the result of several major technical advances. Most notable among them is an increase in die density that enables Rome to provide up to 64 cores per processor, compared with the 48 cores offered by Intel’s latest Xeon server CPUs. The new chips are also the first to support the PCIe 4.0 computer networking standard, which makes it possible to feed data into a CPU at up to twice the rate of earlier technologies. 

Even with all the enhancements packed into Rome, the series is still compatible with server sockets designed for the previous-generation Naples family. Moorhead believes that this backwards-compatibility will be conducive to adoption.

“With all these improvements, AMD made Rome “socket-compatible” with Naples, which should accelerate uptake with OEMs, ODMs and ultimately, end customers,” Moorhead said. “As AMD has delivered on its promises the last two years, I have little doubt AMD will deliver on-time at quality.”

Lastly, AMD unveiled a pair of seven-nanometer data center graphics processing units optimized for artificial intelligence software. The top-end Radeon Instinct MI60 has a peak speed of 7.4 teraflops when processing data in the double-precision floating-point format, which is commonly used for high-performance use cases. A single teraflop equals a trillion operations per second.

At today’s event, AMD presented internal tests that showed the MI60 can run the popular Resnet-50 image processing AI 2.8 times faster than its current 14-nanometer GPUs. The chip performed 8.8 times better under another popular benchmark known as DGEMM.

AMD will bring the MI60 to market alongside the M50, a cheaper GPU that has a peak performance of 6.7 teraflops when crunching double-precision floating-point values. Like the new Ryzen processors, the chips support PCIe 4.0, which enables them them to ingest up to a terabyte of data per second.

Photo: AMD