Skip to content

TPU v7 Ironwood vs Hopper H100 SXM5

Complete side-by-side comparison of specs, performance, memory, power efficiency, and pricing.

GOOGLE

TPU v7 Ironwood

95

Spec Wins

NVIDIA

Hopper H100 SXM5

73

Detailed Specifications

SpecTPU v7 IronwoodHopper H100 SXM5
ArchitectureIronwood (TPU v7) Hopper
Memory192GB HBM3e 80GB HBM3
Memory Bandwidth7,400 GB/s 3,350 GB/s
FP16 TFLOPS2,307 1,979
FP8 TFLOPS4,614 3,958
BF16 TFLOPS2,307 1,979
INT8 TOPS4,614 3,958
TDP1000W 700W
InterconnectICI (1,200 GB/s) (1200 GB/s) NVLink 4.0 (900 GB/s) (900 GB/s)
Perf Score95 73
EcosystemJAX CUDA
Est. PriceCloud Only $25,000

TPU v7 Ironwood — Best For

Frontier TrainingLarge-Scale InferenceJAX

Hopper H100 SXM5 — Best For

LLM TrainingHPC

Who Should Choose Each GPU?

Choose TPU v7 Ironwood if you…

  • Need more VRAM (192GB vs 80GB) for large model inference
  • Prioritize raw FP8 throughput (4,614 vs 3,958 TFLOPS)
  • Running Frontier Training workloads
  • Running Large-Scale Inference workloads
  • Running JAX workloads

Choose Hopper H100 SXM5 if you…

  • Need maximum CUDA/TensorRT/vLLM ecosystem compatibility
  • Have power-constrained data centers (700W vs 1000W TDP)
  • Running LLM Training workloads
  • Running HPC workloads

Verdict

The TPU v7 Ironwood and Hopper H100 SXM5 target different priorities. The TPU v7 Ironwood's 192GB of HBM3e gives it a clear edge for large-model inference where fitting the full model in VRAM eliminates quantization overhead. For training throughput, the TPU v7 Ironwood's 4,614 FP8 TFLOPS outpaces the Hopper H100 SXM5's 3,958 TFLOPS. Teams already invested in the NVIDIA/CUDA ecosystem will have less friction with the Hopper H100 SXM5, while teams open to JAX can benefit from the TPU v7 Ironwood's advantages. Use our TCO Calculator to model the full 3-year cost difference for your specific utilization and power costs.

TPU v7 Ironwood vs Hopper H100 SXM5: Common Questions

Which is faster, TPU v7 Ironwood or Hopper H100 SXM5?+

In FP8 throughput, the TPU v7 Ironwood leads with 4,614 TFLOPS vs 3,958 TFLOPS. For LLM inference, memory capacity and bandwidth often matter more than raw TFLOPS — the TPU v7 Ironwood has more VRAM (192GB).

Is TPU v7 Ironwood or Hopper H100 SXM5 better for LLM training?+

For LLM training at scale, the TPU v7 Ironwood has higher raw throughput. However, the choice also depends on your software stack: Hopper H100 SXM5 offers CUDA compatibility with the widest framework support (PyTorch, JAX, TensorRT).

What is the price difference between TPU v7 Ironwood and Hopper H100 SXM5?+

Pricing for these GPUs varies by vendor and availability. Check our Buy page for current reseller pricing and cloud rental costs.

Which GPU is more power efficient, TPU v7 Ironwood or Hopper H100 SXM5?+

The Hopper H100 SXM5 has a lower TDP (700W vs 1000W). Performance-per-watt depends on your workload — for FP8 inference, divide TFLOPS by TDP: TPU v7 Ironwood = 4.6 TFLOPS/W vs Hopper H100 SXM5 = 5.7 TFLOPS/W.

Ask AI Advisor