AMD is engaged on an accelerated processing unit that may outperform its present prime APU that powers the world’s first exascale supercomputer.
At its latest analyst day, the corporate launched a brand new high-end accelerator, the Intuition MI300, an APU that mixes Zen 4 CPUs, the newest technology of GPU expertise, plus AMD’s Infinity Cache and Infinity structure in a single package deal. It can ship eight instances the AI efficiency of AMD’s present high-end ACU, the MI250, and can be obtainable subsequent 12 months.
A pool of high-bandwidth reminiscence on the ACU is shared between the CPU and the GPU permitting them to speak freely with out the efficiency or power overhead of redundant reminiscence copies.
“It’s designed for reminiscence bandwidth utility latency, and delivers substantial energy financial savings over the choice architectures,” stated Forrest Norrod, senior vice chairman of AMD’s information heart options group. “And we’re going to be more than happy to speak extra about this as we get nearer to the introduction.”
Proper now, AMD has main bragging rights in that the world’s quickest supercomputer is powered by a mix of Epyc CPUs and Intuition MI250 accelerators.
Zen 4 replace
Additionally on the analyst day, CTO Mark Papermaster stated the corporate’s Zen 4 microarchitecture will supply 15% improved single-thread efficiency in data-center functions over present merchandise plus the addition of AVX512 directions for HPC functions.
Zen 4 involves market this 12 months and would be the world’s first 5nm CPU with design modifications to allow larger frequency along with the improved efficiency. General, Zen 4 will function a 25% enchancment in efficiency per watt over Zen 3 and an total 35% efficiency enchancment versus Zen 3.
AMD additionally talked about Siena, describing it as a Zen 4-based server processor and the primary Epyc processor optimized for clever edge and communications deployments that require excessive compute densities in a low-cost, power-efficient platform.
Papermaster stated Zen 5, due in 2024, can be a ground-up new microarchitecture designed to handle some architectural points for scaling throughout a broad vary of workloads in addition to additional optimizations for AI workloads. He didn’t supply additional particulars.
Xilinx, Pensando property
CEO Lisa Su stated AMD’s two latest acquisitions—FPGA maker Xilinx and networking vendor Pensando—are anticipated to contribute to the corporate’s prime line and backside line. Xilinx is increasing the marketplace for AMD merchandise and Pensando is broadening its data-center options. “So our purpose is to be probably the most strategic provider to the most important information facilities on the earth,” she stated. “Pensando has actually management expertise on the DPU entrance, and what we see is that the options will now work hand-in-hand with the remainder of our computing applied sciences.”
The Xilinx and Pensando acquisitions carry three networking merchandise to AMD: Solarflare, acquired by Xilinx in 2019, and used for ultra-low latency situations akin to monetary exchanges; Alveo, Xilinx’s FPGA-based adaptive accelerator for customized networking features; and Pensando, used for high-volume community flows and processes.
“I’m happy to say that we now have the broadest set of safety expertise that we’ve ever had, and a set of applied sciences that enable us to safe the server from boot,” stated Norrod. “So [we are] really securing all the node, and now with the Alveo and Pensando expertise, we are able to safe the firewall with encryption on each port.”
AMD’s purpose is to supply an entire set of administration and safety providers extending out to the sting and all through the entire community. “Our expertise permits us to achieve telemetry and permit our prospects to completely monitor what’s occurring within the information heart with out taking a efficiency penalty from doing so,” stated Norrod.
Copyright © 2022 IDG Communications, Inc.
