AWS Graviton5 launches with 192 Arm cores, challenging Intel and AMD in cloud computing

AWS Unveils Graviton5 with 192 Arm Cores and Advanced Architecture

Amazon Web Services has launched AWS Graviton5, its most powerful custom server CPU to date, now generally available across M9g and M9gd instance types. Developed by Annapurna Labs and manufactured using TSMC's 3nm process, the next-generation Graviton5 processor features a chiplet-based architecture comprising four compute dies that collectively deliver 192 Arm cores2

2

. Each core includes 1 MB of dedicated cache, with the chip offering five times larger L3 cache compared to its predecessor, giving each core access to 2.6 times more cache than Graviton44

4

. This general-purpose Arm CPU operates as a fully coherent processor despite its multi-chiplet design, implementing a die-to-die interconnect that delivers 420 GB/s of bandwidth to maintain cache coherency across the package2

2

.

Performance Gains Target AI Inference and Database Workloads

The Graviton5 processor delivers measurable performance improvements across multiple workload categories. Applications run 35% faster, machine learning inference accelerates by 35%, and databases see 30% faster performance compared to Graviton4-based instances1

1

. AWS positions these gains as targeting up to 25% better performance versus previous-generation instances powered by Intel Xeon Cascade Lake and AMD EPYC Genoa processors2

2

. The chip achieves 33% lower inter-core latency, critical for parallelized cloud computing tasks3

3

. Memory bandwidth reaches over 800 GB/s through a 12-channel DDR5 memory subsystem supporting speeds up to DDR5-8800, which AWS claims represents the fastest memory of any processor instances in the cloud2

2

3

3

.

Advanced Connectivity Through PCIe Gen 6 and DDR5 Memory

Connectivity represents a major focus of the Graviton5 design, with the in-house processor integrating a 96-lane PCIe Gen 6 root complex that significantly increases available I/O bandwidth for accelerators, GPUs, networking hardware, and storage devices2

2

. Network bandwidth increases by up to 15% on average across instance sizes, with the largest instances seeing up to twice the network bandwidth of previous generations. Amazon Elastic Block Store bandwidth improves by 20% on average, enabling faster data transfers and improved performance for distributed applications4

4

. The M9gd instances specifically target customers requiring high-speed local SSD storage, offering up to 11.4 TB of capacity with 30% higher IOPS versus previous generations4

4

.

Meta Leads Major Customer Deployments

AWS confirms that Meta has become one of its largest customers, deploying Graviton5 at scale with tens of millions of CPU cores supporting the company's agentic AI initiatives3

3

. Snowflake has committed $6 billion to AWS for computing capacity, while Uber also ranks among major customers leveraging the custom silicon platform1

1

4

4

. After eight years of continuous investment across five generations, Graviton now powers over 350 instance types serving more than 120,000 customers ranging from startups to large enterprises3

3

. AWS is deploying new Amazon EC2 M9g instances across the United States and parts of Europe, with availability remaining strong even for smaller customers1

1

.

Pricing Considerations and Market Positioning

The new instances come with a 9% price increase over Graviton4-based offerings, marking a shift from earlier generations when upgrades delivered both better performance and lower costs1

1

. AWS maintains that customers running large fleets at high CPU utilization will see net cost savings through improved price-performance, though customers requiring fixed node counts face a direct price increase. This pricing pressure reflects broader industry dynamics, with component costs rising as companies compete for manufacturing capacity. Despite marketing emphasis on AI inference capabilities and the "agentic AI era," industry observers note that Graviton remains fundamentally a general-purpose Arm CPU rather than specialized AI hardware like AWS's Trainium accelerators1

1

. The processor's actual strengths lie in web applications, analytics, databases, gaming, and video encoding workloads, with AI inference representing one capability among many3

3

. AWS's continued investment in custom silicon positions the company to optimize hardware specifically for cloud computing workloads while reducing dependence on Intel Xeon and AMD EPYC processors that have traditionally dominated data center infrastructure.