|
Immediately, we’re saying the final availability of Amazon Elastic Compute Cloud (Amazon EC2) P6-B300 cases, our next-generation GPU platform accelerated by NVIDIA Blackwell Extremely GPUs. These cases ship 2 instances extra networking bandwidth, and 1.5 instances extra GPU reminiscence in comparison with earlier era cases, making a balanced platform for large-scale AI purposes.
With these enhancements, P6-B300 cases are perfect for coaching and serving large-scale AI fashions, notably these using refined strategies akin to Combination of Consultants (MoE) and multimodal processing. For organizations working with trillion-parameter fashions and requiring distributed coaching throughout 1000’s of GPUs, these cases present the right stability of compute, reminiscence, and networking capabilities.
Enhancements made in comparison with predecessors
The P6-B300 cases ship 6.4Tbps Elastic Material Adapter (EFA) networking bandwidth, supporting environment friendly communication throughout massive GPU clusters. These cases characteristic 2.1TB of GPU reminiscence, permitting massive fashions to reside inside a single NVIDIA NVLink area, which considerably reduces mannequin sharding and communication overhead. When mixed with EFA networking and the superior virtualization and safety capabilities of AWS Nitro System, these cases present unprecedented pace, scale, and safety for AI workloads.
The specs for the EC2 P6-B300 cases are as follows.
| Occasion dimension | VCPUs | System reminiscence | GPUs | GPU reminiscence | GPU-GPU interconnect | EFA community bandwidth | ENA bandwidth | EBS bandwidth | Native storage |
| P6-B300.48xlarge | 192 | 4TB | 8x NVIDIA B300 GPU | 2144GB HBM3e | 1800 GB/s | 6.4 Tbps | 300 Gbps | 100 Gbps | 8x 3.84TB |
Good to know
When it comes to persistent storage, AI workloads primarily use a mixture of excessive efficiency persistent storage choices akin to Amazon FSx for Lustre, Amazon S3 Categorical One Zone, and Amazon Elastic Block Retailer (Amazon EBS), relying on value efficiency concerns. For illustration, the devoted 300Gbps Elastic Community Adapter (ENA) networking on P6-B300 allows high-throughput sizzling storage entry with S3 Categorical One Zone, supporting large-scale coaching workloads. Should you’re utilizing FSx for Lustre, now you can use EFA with NVIDIA GPUDirect Storage (GDS) to attain as much as 1.2Tbps of throughput to the Lustre file system on the P6-B300 cases to rapidly load your fashions.
Obtainable now
The P6-B300 cases at the moment are out there by Amazon EC2 Capability Blocks for ML and Financial savings Plans within the US West (Oregon) AWS Area.
For on-demand reservation of P6-B300 cases, please attain out to your account supervisor. As traditional with Amazon EC2, you pay just for what you employ. For extra data, discuss with Amazon EC2 Pricing. Take a look at the complete assortment of accelerated computing cases that can assist you begin migrating your purposes.
To be taught extra, go to our Amazon EC2 P6-B300 cases web page. Ship suggestions to AWS re:Put up for EC2 or by your traditional AWS Assist contacts.
– Veliswa
