Skip to content
Updated minutes ago
other

Trainium2

other · other · 96 GB HBM2e · 600W TDP

VRAM

96 GB

BF16 TFLOPS

756

Bandwidth

3200 GB/s

From

$1.95/hr

Calculate ROI with this GPU →

Spec Sheet

VRAM96 GB HBM2e
Memory Bandwidth3200 GB/s
BF16 TFLOPS756
FP16 TFLOPS756
FP8 TFLOPS1512
INT8 TOPS1512
TDP600W
InterconnectPCIE
Max per Node16
PCIe Gen5
Tensor CoresNo

Pricing by Provider

ProviderOn-DemandReservedSpotBadge
aws$2.78/hr$1.95/hr-Cheapest

Compatible Models (280)

Training Capabilities

Estimated GPU count for full fine-tuning (AdamW, BF16) and QLoRA

Model SizeFull Fine-TuneQLoRA
7B model2 GPUs1 GPU
13B model3 GPUs1 GPU
70B model14 GPUs1 GPU

Energy Efficiency

Estimated tokens/second per Watt for popular models

Mistral 7B
0.73 t/s/WFP8
Qwen 2.5 7B
0.70 t/s/WFP8
Llama 3.1 8B
0.66 t/s/WFP8
DeepSeek V3
0.14 t/s/WFP8
Llama 3.1 70B
0.08 t/s/WFP8
Qwen 2.5 72B
0.07 t/s/WFP8

Similar GPUs

GPUVRAMBF16 TFLOPSBW (GB/s)From
Cloud AI 10032 GB150134$0.00/hr
Groq LPU230 GB18880000$0.00/hr
H2096 GB1484000$0.99/hr
Gaudi 296 GB4322460$1.65/hr
GH20096 GB9904000$2.99/hr

Methodology Note

Performance estimates for the Trainium2are based on InferenceBench's roofline performance model with CUDA kernel-level optimization including FlashAttention v2 and PagedAttention. Memory calculations account for model weights (96 GB HBM2e available), KV-cache allocation, and activation memory. Throughput predictions use the Trainium2's rated 3200 GB/s memory bandwidth and 756 BF16 TFLOPS compute capacity as roofline ceilings, with empirical correction factors per GPU architecture (other). See our full methodology.

Frequently Asked Questions

How many AI models can run on Trainium2?

The Trainium2 can run 280 AI models from our database within a single node. Compatible models range across various parameter sizes depending on the quantization precision (BF16, FP8, INT4). Smaller models fit on a single GPU while larger models may require multi-GPU setups up to 16x Trainium2.

What is the Trainium2 inference throughput?

The Trainium2 delivers 756 BF16 TFLOPS and 1512 FP8 TFLOPS with 3200 GB/s memory bandwidth. Actual inference throughput (tokens/sec) depends on the model size, precision, and batch size. Use our calculator for model-specific throughput estimates.

How much does Trainium2 cost per hour?

The Trainium2 is available starting from $1.95/hour via aws. Prices vary by provider and pricing tier (on-demand, reserved, spot). Compare pricing across all providers in the table above.