[ad_1]
Marvell’s announcement of the network adapter – or in the current industry spokesperson Data Processing Unit (DPU) – Octeon 10 comes with a horde of catchphrases: 5-nanometer technology from TSMC, ARMv9 architecture with Neoverse N2 computing cores, DDR5 controller and PCI Express 5.0 interface. The manufacturer plans to deliver the first cards this year.
The core the Octeon-10 series form differently configured systems-on-chip (SoCs). The first model, the CN106XX, uses 24 Neoverse N2 CPU cores, hexachannel RAM (DDR5-5200 with ECC) and an integrated 200 gigabit switch that can be divided into four 50G ports. The CN106XXS with terabit switch and the DPU400 with 36 CPU cores and 12 DDR5-5200 channels will follow in 2022.
Lots of cache
Marvell draws on the full range of ARM’s processor modules: Each Neoverse N2 core receives the maximum 128 KByte level 1 cache and 1 MByte L2 cache, and there are up to 72 MByte L3 intermediate buffers (2 MByte per core, the all share). The CPU clock frequency is 2.5 GHz. The ARMv9 architecture includes Scalable Vector Extensions 2 (SVE2), which in the case of the Octeon 10 series are to vectorize network packets and thus parallelize them (Vector Packet Processing, VPP).
(Image: Marvell)
The monolithic chips also include crypto processors for encrypting and decrypting data and self-designed AI accelerators consisting of SRAM tiles and multiply accumulate units (MACs). Thanks to the integration into the SoC, the latency should be low. Depending on the model, the AI accelerators have a computing power of between 20 and 100 TOPS – enough, according to Marvel, to run trained networks that are supposed to detect malware in network traffic.
Network adapters such as those from the Octeon 10 series are designed to relieve (main) processors. Marvell wants to cover a broad field: from high-end cards for data centers to 10-watt models for 5G stations.
(Image: Marvell)
High CPU performance
The first work from the Octeon 10 series, the CN106XX, should achieve more than 800 points in the SPECint 2006 benchmark, the upcoming top model DPU400 more than 1200. That is the performance level of mid-range server processors from 2017, such as Intel’s Skylake -SP and AMD’s Naples (Epyc 7001) – only as a co- and not the main CPU.
Intel plans to launch comparable network adapters in the form of SmartNICs in 2022. Nvidia competes with the Bluefield DPU family and Xilinx – now an AMD subsidiary – with Alveo.
(mma)
[ad_2]
