GNR (8i) on AWS

Notes on AWS EC2 8i instances — the Granite Rapids (Intel Xeon 6) generation — focused on NUMA/SNC3 topology, plus key features, specifications versus prior generations, and the resulting strengths.

See also: GNR (C4) on GCP — the same processor generation on Google Cloud — and Cross-Cloud Comparison for the AWS-versus-GCP node-width differences.

Contents

NUMA Topology

8i instances expose the Intel Xeon 6 SNC3 feature for best performance. Each socket is presented to the OS as 3 NUMA nodes, each with 32 cores and memory closer to those cores, giving the best memory latencies for NUMA-aware applications.

SNC3 architecture

Each GNR socket contains 3 compute dies and 12 memory channels. With SNC3 enabled, each compute die is presented as its own NUMA node:

8i socket: 3 compute dies, each 32 cores / 64 threads, each its own NUMA node

Per-socket comparison

  Cores/socket vCPUs/socket NUMA nodes/socket
6i 32 64 1
7i 48 96 1
8i 96 192 3

Sockets by VM size

VM size 6i sockets 7i sockets 8i sockets
4xl 1 1 1
8xl 1 1 1
12xl 1 1 1
16xl 1 1 1
24xl 2 1 1
32xl 2 NA 1
48xl NA 2 1
96xl NA NA 2

Note — 8i reaches 48xl within a single socket, where 7i needs two.

NUMA nodes by VM size

VM size 6i NUMA 7i NUMA 8i NUMA
4xl 1 1 1
8xl 1 1 1
12xl 1 1 1
16xl 1 1 1
24xl 2 1 2
32xl 2 NA 2
48xl NA 2 3
96xl NA NA 6

8i NUMA node and compute die layout by VM size

VM size vCPUs Sockets NUMA nodes vCPUs/node Cores/node Compute die coverage
4xl 16 1 1 16 8 Partial compute die
8xl 32 1 1 32 16 Partial compute die
12xl 48 1 1 48 24 Partial compute die
16xl 64 1 1 64 32 1 full compute die
24xl 96 1 2 48 24 2 partial compute dies
32xl 128 1 2 64 32 2 full compute dies
48xl 192 1 3 64 32 Single socket (3 compute dies)
96xl 384 2 6 64 32 Full system (6 compute dies)

Note — the vCPUs/node and Cores/node columns are derived (vCPUs ÷ NUMA nodes, halved for Hyper-Threading).

Practical guidance

Intel Architecture Instance Types on AWS

AWS offers Intel-based instances across every major family. Instance families are grouped by workload profile:

Family Purpose
General Purpose (M, T) Balance of compute, memory, and networking for diverse workloads.
Compute Optimized (C) Compute-bound applications that benefit from high performance processors.
Memory Optimized (R, X, U, z1d) Fast performance for workloads processing large in-memory data sets.
Storage Optimized (I, D, H) High sequential read/write access to very large data sets on local storage.
Accelerated Compute (P, G, F, DL1) Hardware accelerators / co-processors for specialized functions.
HPC Optimized (HPC6id) Large, complex simulations and deep learning workloads.

Processor generation mapping

AWS naming Intel generation Codename
8i (M8i, C8i, R8i, X8i, plus -flex and d variants) Intel Xeon 6 Granite Rapids
I7i, I7ie 5th Gen Intel Xeon Scalable Emerald Rapids
7i (M7i, C7i, R7i, R7iz, U7i, U7in, U7inh, plus -flex variants) 4th Gen Intel Xeon Scalable Sapphire Rapids
6i (M6i, C6i, R6i, X2idn, X2iedn, I4i) 3rd Gen Intel Xeon Scalable Ice Lake
5-series, later parts (M5n, R5n, X2iezn) 2nd Gen Intel Xeon Scalable Cascade Lake
5-series, early parts (M5, C5, R5, z1d, I3en) Intel Xeon Scalable Skylake
M4, C4, R4, X1, D2 Intel Xeon v3 / v4 Haswell / Broadwell, varies by size

Note — the AWS generation digit does not track the Intel generation. I7i is Emerald Rapids, one Intel generation newer than the Sapphire Rapids M7i/C7i/R7i, because the storage-optimized line numbers independently of the compute lines. No AWS 7i compute family is Emerald Rapids at all: those went Sapphire Rapids (7i) straight to Granite Rapids (8i). Confirm the processor for a specific instance type in the AWS specification tables rather than inferring it from the name.

Note — the 8i row is Intel Xeon 6, not “6th Gen Intel Xeon Scalable”; Intel dropped the “Nth Gen … Scalable” scheme for this generation, and AWS follows the Xeon 6 branding.

AWS instance suffix decoder

Suffix Meaning
d Instance store volumes
n Network and EBS optimized
e Extra storage or memory
z High performance
-flex Flex variant (see the AWS references below)

Reference: https://aws.amazon.com/ec2/instance-types/ and https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instance-types.html

The per-instance-type processor is listed in the AWS specification tables, which are the source for the mapping above: general purpose, compute optimized, memory optimized, storage optimized.

AWS GNR Instances & Key Features

vCPU-to-memory ratio by family

Instance families Memory per vCPU (GB)
C8i, C8i-flex, C8id 2
M8i, M8i-flex, M8id 4
R8i, R8i-flex, R8id 8
X8i 16

Instance sizes

Size vCPUs
large 2
xlarge 4
2xlarge 8
4xlarge 16
8xlarge 32
12xlarge 48
16xlarge 64
24xlarge 96
32xlarge 128
48xlarge 192
96xlarge 384
metal-48xl 192
metal-96xl 384

Specifications: 6i vs 7i vs 8i (per socket)

  6i 7i 8i
Micro architecture Ice Lake Sapphire Rapids Granite Rapids
Cores per socket 32 48 96
vCPUs per socket 64 96 192
Max frequency 3.5 GHz 3.8 GHz 3.9 GHz
All-core frequency 3.5 GHz 3.2 GHz 3.9 GHz
Memory DDR4-3200 DDR5-4800 DDR5-7200
L2 cache per core 1.3 MB 2 MB 2 MB
L3 cache 54 MB 210 MB 480 MB
Memory channels 8 8 12
Sub-NUMA Cluster (SNC) No No Yes (SNC3)
TDP 300 W 385 W 600 W
Accelerators — AMX, QAT, IAA/DSA, DLB AMX, QAT, IAA/DSA, DLB

Differentiation: Strengths