Skip to content

Amazon EC2 P6-B300 Instances Launch in APJ (Seoul and Hyderabad region)

Amazon EC2 P6-B300 instances, a next-generation GPU instance powered by eight NVIDIA Blackwell Ultra (B300) GPUs, have officially launched in Seoul region (ap-northeast-2) and Hyderabad region (ap-south-2) in August 2026. Compared with the previous-generation P5en (H200), P6-B300 delivers 1.9× the GPU memory (2.1TB), 1.6× the memory bandwidth (7.7 TB/s), and 2× the NVLink and EFA bandwidth, while adding FP4 precision — substantially improving the performance and token economics of AI training and inference workloads.
AWS is the only major CSP to offer a B300 8-GPU configuration, and these APJ regions are among the first anywhere to provide P6-B300 outside US regions. APJ customers can now enjoy the benefits of the latest Blackwell GPUs from in-region data centers, meeting data residency requirements while minimizing latency for local users.

Amazon EC2 P6-B300 — now available in APJ, in the Seoul (Aug 20, 2026) and Hyderabad (Aug 28, 2026) regions. 1.9× GPU memory, 1.6× memory bandwidth, 2× GPU interconnect, 2× inter-node networking, new FP4 precision

1. P6-B300 Key Highlights

Four key highlights of the P6-B300 — Blackwell Ultra architecture, 1.9× GPU memory and 1.6× bandwidth, 2nd-gen Transformer Engine + FP4, 2× interconnect

2. Key Specifications — Improvements vs the Previous Generation (P5en)

Specification P5en (H200) P6-B300 (B300) Improvement
Release '24.12 '25.11 —
GPU 8 × NVIDIA H200 8 × NVIDIA B300 (Blackwell Ultra) —
Architecture Hopper (single die, 80B) Blackwell Ultra (dual die, 208B) 2.6× transistors
GPU memory 1,128 GB HBM3e (141GB/GPU) 2,144 GB HBM3e (268GB/GPU) 1.9×
Memory bandwidth 4.8 TB/s 7.7 TB/s 1.6×
FP8/FP6 16 PFLOPS 36 PFLOPS 2.25×
FP4 Not supported 108 PFLOPS (without sparsity) / 144 PFLOPS (with sparsity) New support
FP16/BF16 8 PFLOPS 18 PFLOPS 2.25×
NVLink NVLink 4 — 900 GB/s NVLink 5 — 1,800 GB/s 2×
Networking 3,200 Gbps (EFAv3) 6,400 Gbps (EFAv4) + 300 Gbps dedicated ENA 2× EFA, new dedicated ENA
CPU Intel Sapphire Rapids Intel Emerald Rapids New CPU
System memory 2 TB 4 TB 2×
Nitro System Nitro v5 Nitro v6 Latest Nitro

3. Ideal Use Cases

  • Large-scale inference serving

    • Online inference optimization for 70B+ models (DeepSeek, Kimi series, and more)
    • 100K+ tok/s throughput on a single GPU
    • 268GB holds the entire KV cache, serving 128K+ contexts natively
    • ~50% fewer GPUs · 40–60% ↓ latency

  • Multi-node large-scale training

    • 400B+ parameter models
    • Distributed training across thousands of GPUs with 6,400 Gbps EFAv4 + NVLink 5
    • 2.1TB+ of GPU memory per instance serves a 400B model with ample batch and KV cache
    • 2× communication bandwidth → higher scaling efficiency

4. Key Benefits for Training

The P6-B300's larger memory and higher bandwidth can reduce the number of GPUs needed for a given model, which in turn cuts infrastructure complexity — sharding strategy, distributed-training overhead, debugging, and more. FP8-native compute raises throughput 2–4× versus P5en, shortening training time and improving time-to-market.

Phase 1 — Built-in performance gains

  • 4× higher compute: the dual-die design and 5th-gen Tensor Cores cut iteration time by ~50% at the same model and batch size.
  • 2× GPU-to-GPU bandwidth: NVLink 5 delivers 1.8 TB/s of bidirectional bandwidth per GPU, relieving the network and storage bottlenecks inherent to tensor and pipeline parallelism.
  • Enhanced AWS infrastructure: EFAv4 (6,400 Gbps), dedicated ENA (300 Gbps), and Nitro v6 (400 Gbps per NIC with RDMA support) keep data I/O from competing with training communication, minimizing GPU idle time.

Phase 2 — Gains from additional optimization

  • GPU savings through memory optimization: 1.9× memory per GPU (H200 141GB → B300 268GB) lets you train the same model with half the GPUs — for a 70B model, multi-node FSDP can collapse to single-node DDP.
  • FP8/FP4 quantization: hardware-native MXFP8 holds BF16-equivalent accuracy while delivering 2× throughput. NVFP4's 4× memory efficiency enables larger micro-batches, reducing gradient accumulation and raising throughput.
  • Data-pipeline optimization (nvCOMP): a dedicated CPU→GPU hardware compression/decompression offload cuts I/O bottlenecks and checkpoint size by 20–30% each, speeding up experiment iteration and lowering FSx cost.

5. Key Benefits for Inference

  • 1.6× memory bandwidth (4.8 → 7.7 TB/s): since decode performance scales with memory bandwidth, this lowers time-per-output-token (TPOT) and improves real-time responsiveness.
  • 1.9× memory capacity (H200 141GB → B300 268GB/GPU): ~50% fewer GPUs per model and a lower TP degree remove GPU-to-GPU communication overhead, while a larger KV cache supports more concurrent sessions or longer contexts.
  • NVFP4: ~4× memory savings versus FP16 and ~1.8× versus FP8, with accuracy on par with FP8 — and up to 50% less communication time in large prefill phases.
  • NVLink 5 (900 → 1,800 GB/s): 2× the bandwidth for Expert Parallel (EP) all-to-all communication in MoE inference, lowering multi-GPU serving latency for large MoE models — and also easing TP AllReduce bottlenecks in dense models.

6. Why AWS

  • Expertise

    • First to offer cloud GPU services, with the largest number of customer references worldwide
    • A dedicated GPU team based in Korea
    • Best practices for AI infrastructure design, operation, and optimization
    • Over 15 years of collaboration with NVIDIA and the deepest joint software optimization

  • Performance

    • Nitro System: dedicated hardware and a lightweight hypervisor remove virtualization overhead for bare-metal-class performance
    • 3-tier storage configuration removes I/O bottlenecks and optimizes cost (Instance Store + FSx for Lustre + S3)
    • Preconfigured Deep Learning AMIs/Containers shorten time-to-market and reduce operational burden

  • Reliability

    • 99.99% uptime SLA
    • Nitro security chip installed on the host motherboard, with complete isolation between VMs
    • Lowest total downtime among major APAC clouds (per Frost & Sullivan, less than one-third of competitors)

  • Flexibility & Ecosystem

    • The only major CSP to support a B300 8-GPU config (including in the Seoul region)
    • The broadest instance portfolio, so you can choose the optimal chip and size for each workload
    • A range of purchase options: On-Demand, Capacity Blocks, Spot, and more
    • Native integration with 200+ AWS services (S3, FSx, EKS, CloudWatch, and more)
    • Cost optimization through a GPU + Bedrock hybrid
    • Optimized for the NVIDIA stack — DCGM Exporter, MIG, DRA, Run:ai, NIM, and more
    • A rich open-source and partner ecosystem