Amazon Web Services has introduced a new CPU customized for high-performance computing (HPC) and the next generation of its Nitro smart networking chip, plus instances that take full advantage of the hardware.
The Arm-based CPU is called the Graviton3E and has been optimized for floating point math, key in HPC, the company announced at AWS re:Invent conference. Amazon said Hpc7g instances powered by the new Graviton3E chips offer up to double the floating point and vector performance compared to the current generation of instances.
The vast datasets that accompany HPC need to be moved around, so Amazon also introduced the fifth generation of its Nitro smartNICs, offering up to twice the network bandwidth and up to 50% higher packet processing-per-second performance compared to current generation networking-optimized instances.
itro v5 has about twice the transistors as the previous generation Nitro chip, which gives it about twice the computational power, said Peter DeSantis, senior vice president of AWS utility computing during his keynote. It also has 50% more memory bandwidth and a PCIe adapter that provides about twice the bandwidth and supports a 60% higher packet-per-second rate with a 30% reduction in packet latency.
Accompanying the new chip is a new Elastic Compute Cloud (EC2) instance, the networking optimized C7G. The C7G uses the Graviton 3 processor and is designed to deliver optimized networking performance for the most network-intensive workloads. It offers 200Gbps. of throughput and up to 50% higher packet-processing performance over the previous network optimized instance.
Replacing TCP
To accelerate data movement for HPC workloads, AWS created the elastic fabric adapter (EFA) for scalable, high-speed inter-node communication. As part of EFA is another AWS creation, Scalable Reliable Datagram (SRD), an alternative to the widely used TCP Ethernet protocol.
DeSantis said AWS internal networking relies on multipath routing, but creaky old TCP uses a single path and isn’t too good at noting if performance is compromised. TCP also transmits packets in order, which can cause latency.
SRD makes use of multipath routing so doesn’t transmit packets to arrive order, but it reorders the packets at the receiving end. DeSantis said it will retransmit dropped packets “in microseconds, not milliseconds” and speed up networks hosted on the AWS cloud.
And now all of AWS will gain use of SRD, because AWS has built a new version of the network driver offered with EC2 instances called ENA Express to offer native SRD support. So SRD is being rolled out site-wide.
