Reserve B300 Capacity

Tell us about your deployment. A solutions architect will be in touch shortly

How many nodes do you need? (1 node = 8 GPUs)

What is your workload?

Timeline

Environment

<< Previous
Next >>

Available Now

Learn More

AMD Instinct™
MI325x

AMD Instinct™
MI325x

The most capable AI accelerator available today. Deployed on Bentaus-owned energy infrastructure. Reserve capacity now.

The most capable AI accelerator available today. Deployed on Bentaus-owned energy infrastructure. Reserve capacity now.

The most capable AI accelerator available today. Deployed on Bentaus-owned energy infrastructure. Reserve capacity now.

256GB

HBM3E PER GPU

1.8×

Performance

25%

More efficient

Order Now

Order Now

Reserve Capacity

Reserve Capacity

8x GPU NODE

Bandwith

6 TB/s

HBM3E

256 GB

Bentaus is an energy-first NeoCloud that owns and operates its own full stack. Our MI325X deployments are engineered for best-in-class performance, efficiency, and reliability.

Bentaus is an energy-first NeoCloud that owns and operates its own full stack. Our MI325X deployments are engineered for best-in-class performance, efficiency, and reliability.

A dark abstract backdrop with soft diagonal ridges, symbolizing secure layers, encrypted data, and a protected connection.
Mobile VPN app

WHY BENTAUS

Why Enterprise Teams Choose Us Over Today's Hyperscalers

Deploy dedicated MI325X capacity without the delays, higher cost, or limits of traditional hyperscale solutions.

Order Now

Order Now

WHY BENTAUS

Why Enterprise Teams Choose Us Over Today's Hyperscalers

Deploy dedicated MI325X capacity without the delays, higher cost, or limits of traditional hyperscale solutions.

Order Now

Order Now

WHY BENTAUS

Why Enterprise Teams Choose Us Over Today's Hyperscalers

Deploy dedicated MI325X capacity without the delays, higher cost, or limits of traditional hyperscale solutions.

Order Now

Order Now

Lower Cost Than Hyperscalers

Enterprise MI325X capacity without the hyperscaler premium. Bentaus gives clients transparent pricing, and better cost control as GPU demand scales.                                           

Lower Cost Than Hyperscalers

Enterprise MI325X capacity without the hyperscaler premium. Bentaus gives clients transparent pricing, and better cost control as GPU demand scales.                                           

Best in Class Performance

Powered by AMD CDNA 3 architecture, MI325X delivers more HBM3e capacity and bandwidth per GPU than any competing shipping accelerator, for enterprise training, inference, and agentic workloads.

Best in Class Performance

Powered by AMD CDNA 3 architecture, MI325X delivers more HBM3e capacity and bandwidth per GPU than any competing shipping accelerator, for enterprise training, inference, and agentic workloads.

Best in Class Performance

Powered by AMD CDNA 3 architecture, MI325X delivers more HBM3e capacity and bandwidth per GPU than any competing shipping accelerator, for enterprise training, inference, and agentic workloads.

Faster Deployments

From instant deployments to full clusters in weeks, not months. Our Bentaus designs are ready for today's datacenters and future proofed for tomorrow.

Faster Deployments

From instant deployments to full clusters in weeks, not months. Our Bentaus designs are ready for today's datacenters and future proofed for tomorrow.

Faster Deployments

From instant deployments to full clusters in weeks, not months. Our Bentaus designs are ready for today's datacenters and future proofed for tomorrow.

Super Efficient

Bentaus optimizes MI325X deployments for stronger watt-per-token performance, helping clients generate more AI output with less wasted power.

Super Efficient

Bentaus optimizes MI325X deployments for stronger watt-per-token performance, helping clients generate more AI output with less wasted power.

Choose Your Pricing Model

On-Demand

Flexible Access

$6.50 - $8.50

On-Demand Access · MI325X · Per Hour

AMD Instinct MI325X

Ideal for burst inference workloads

Elastic capacity on demand

Managed infrastructure stack

Order Now

Order Now

Reserve Capacity

Reserve Capacity

Reserved

Reserved Term · Priority Access · Locked Rate

$4.25 - $6.50

/GPU hour

Priority access, guaranteed capacity

Lower hourly rate vs. shared cloud

Custom term lengths available

Dedicated solutions architect

Order Now

Order Now

Reserve Capacity

Reserve Capacity

AI Factory

Dedicated Cluster


Custom Pricing

Design your own AI factory

Full cluster isolation

Monitoring, support, and SLAs

Additional feature

Contact Sales

Contact Sales

Reserve Capacity

Reserve Capacity

AMD Instinct™ MI325X Specifications

01

GPU Architecture

AMD CDNA 3

02

VRAM

256 GB HBM3e

03

Memory Bandwidth

Up to 6.0 TB/s

04

Form Factor

8-GPU UBB-8

05

GPU-to-GPU BW

448 GB/s

06

FP64 Performance

86 TFLOPS

07

FP32 Performance

121 TFLOPS

08

BF16 Performance

847 TFLOPS

09

FP8 Tensor Performance

1,548 TFLOPS

10

Scale-Out Bandwidth

50 GB/s

11

Network

400 GbE / PCIe Gen5

12

Max GPU Power

Up to 1,000W

COMPARISON

MI325X vs MI300X vs H200

Side-by-side specs per node. See why the MI325X leads on memory where it matters most.

Spec (per node)

MI325X

Top choice

MI300X

H200

Architecture

CDNA 3

CDNA 3

Hopper

Architecture

CDNA 3

CDNA 3

Hopper

GPUs / node

8× MI325X

8× MI300X

8× H100

GPUs / node

8× MI325X

8× MI300X

8× H100

Memory / GPU

256GB HBM3e

192GB HBM3e

80GB HBM3

Memory / GPU

256GB HBM3e

192GB HBM3e

80GB HBM3

Total HBM / node

2TB HBM3e

1.5TB HBM3e

640GB HBM3

Total HBM / node

2TB HBM3e

1.5TB HBM3e

640GB HBM3

Peak HBM Bandwidth

6.0 TB/s

5.3 TB/s

No

Peak HBM Bandwidth

6.0 TB/s

5.3 TB/s

No

FP4 Compute

6.0 TB/s

5.3 TB/s

No

BF16 Throughput

847 TFLOPS

668 TFLOPS

16 PFLOPS

BF16 Throughput

847 TFLOPS

668 TFLOPS

16 PFLOPS

FP8 Throughput

847 TFLOPS

668 TFLOPS

16 PFLOPS

FP8 Throughput

1,548 TFLOPS

1,240 TFLOPS

NVLink 4

FP8 Throughput

1,548 TFLOPS

1,240 TFLOPS

NVLink 4

NVLink Fabric

1,548 TFLOPS

1,240 TFLOPS

NVLink 4

Networking

400 GbE / PCIe Gen5

400 GbE / PCIe Gen5

400G InfiniBand

Networking

400 GbE / PCIe Gen5

400 GbE / PCIe Gen5

400G InfiniBand

Order Now

Order Now

Reserve Capacity

Reserve Capacity

WORKLOADS

What the MI325X is Built For

From frontier model training to real-time inference and always-on agentic AI applications.

  • from accelerate import Accelerator
    from transformers import AutoModelForCausalLM
    import torch

    accelerator = Accelerator {mixed_precision = "fp8"}
     model = AutoModelForCausalLM.from_pretrained
     "meta-llama/Llama-3-405b",
     torch_dtype=torch.float8_e4m3fn)
     trainer.train(dataset="1T_tokens")

  • from accelerate import Accelerator
    from transformers import AutoModelForCausalLM
    import torch

    accelerator = Accelerator {mixed_precision = "fp8"}
     model = AutoModelForCausalLM.from_pretrained
     "meta-llama/Llama-3-405b",
     torch_dtype=torch.float8_e4m3fn)
     trainer.train(dataset="1T_tokens")

TRAINING

Fine-Tuning & Training at Scale

Fine-Tuning & Training at Scale

Full fine-tuning and mid-scale pre-training runs that benefit from more memory per GPU. 256 GB of HBM3e enables bigger batch sizes and fewer GPUs per model shard — MLPerf benchmarks show MI325X training Llama 2 70B up to 1.33× faster than MI300X.

Full fine-tuning and mid-scale pre-training runs that benefit from more memory per GPU. 256 GB of HBM3e enables bigger batch sizes and fewer GPUs per model shard — MLPerf benchmarks show MI325X training Llama 2 70B up to 1.33× faster than MI300X.

Frontier Model Training

Large-Scale Fine-Tuning

RLHF

INFERENCE

High-Throughput Inference

High-Throughput Inference

Serve large models with more headroom per GPU. MLPerf benchmarks show up to 1.25× higher inference throughput than MI300X, with FP8 and BF16 support across production serving and offline batch workloads.

Serve large models with more headroom per GPU. MLPerf benchmarks show up to 1.25× higher inference throughput than MI300X, with FP8 and BF16 support across production serving and offline batch workloads.

Production APIs

70B–400B models

Real-time serving

Inference Ready

8 x B300 - FP4 - 18ms avg - zero queue

|

Inference Ready

8 x B300 - FP4 - 18ms avg - zero queue

|

Agent Runtime

Autonomously coordinating reasoning, memory, and tool execution in real time.

Context Analysis

Reasoning Engine

Tool Execution

All Tasks

  • Distributed training run

    Running 8 x B300 - Started 2h ao

  • Inference serving - production

    94% active - 2.4M tok/s - 0 queue

  • Model checkpoint sync

    completed - 4 min ago

  • Agent Pipeline #4

    Queued - starts in 3 min

All Tasks

  • Distributed training run

    Running 8 x B300 - Started 2h ao

  • Inference serving - production

    94% active - 2.4M tok/s - 0 queue

  • Model checkpoint sync

    completed - 4 min ago

  • Agent Pipeline #4

    Queued - starts in 3 min

AGENTIC AI

Multi-Agent Systems, Production-Ready

Multi-Agent Systems, Production-Ready

Autonomous agents and multi-step reasoning pipelines need memory that holds context and a software stack that doesn't fight you. 256 GB of HBM3e keeps more working context on-GPU across agent steps, and AMD's ROCm stack ships with out-of-the-box support for the major ML frameworks — no custom porting to get an agent pipeline running.

Autonomous agents and multi-step reasoning pipelines need memory that holds context and a software stack that doesn't fight you. 256 GB of HBM3e keeps more working context on-GPU across agent steps, and AMD's ROCm stack ships with out-of-the-box support for the major ML frameworks — no custom porting to get an agent pipeline running.

Autonomous AI Agents

Multi-Step Reasoning

Use tools at scale

AI FACTORIES

Enterprise AI Factories

Enterprise AI Factories

Every AI factory is different. Bentaus designs and delivers AMD Instinct infrastructure tailored to your workloads, from sovereign AI to advanced simulation, backed by hardware already running in production across Tier 1 clouds and 10+ neo-cloud providers, with a full ROCm software stack and expert support from design through deployment.

Every AI factory is different. Bentaus designs and delivers AMD Instinct infrastructure tailored to your workloads, from sovereign AI to advanced simulation, backed by hardware already running in production across Tier 1 clouds and 10+ neo-cloud providers, with a full ROCm software stack and expert support from design through deployment.

AI Factory Design

Sovereign AI

Healthcare AI

GPU Compute

Memory Fabric

AI Workloads

Deploy Your MI325X Solution Today!

MI325X deployments are among the most versatile AI infrastructure solutions moving fast in the market. Our solution architects are ready to help you design, configure, and deploy now.

Order Now

Order Now

Available Now

NVIDIA B300s faster than you expected.

Order Now

Order Now

Deploy Your MI325X Solution Today!

MI325X deployments are among the most versatile AI infrastructure solutions moving fast in the market. Our solution architects are ready to help you design, configure, and deploy now.

Order Now

Order Now