Google TPU v3 32GB vs Google TPU v5e 16GB

The TPU v5e delivers 60% more BF16 throughput than the TPU v3 (197 vs 123 TFLOPS dense). The TPU v3 counters with 16 GB more memory (32 GB vs 16 GB).

TPU v3: TPU, 2018 TPU v5e: TPU, 2023
BF16 dense lead
+60%
TPU v5e: 197 vs 123 TFLOPS
VRAM
32 vs 16 GB
16 GB more for the TPU v3
Bandwidth
900 GB/s vs 800 GB/s
TPU v3 moves data faster

Specifications

TPU v3
TPU v5e
Architecture
TPUTPU
Launch Year
20182023
Form Factor
VRAM
32 GB16 GB
Memory Bandwidth
900 GB/s800 GB/s
TDP
Process Node

Performance (TFLOPS)

TPU v3
TPU v5e
BF16
123.0 TFlops 197.0 TFlops
INT8
393.0 TFlops

FLOPS by Precision

What actually differs

The TPU v5e is the newer part: TPU, launched in 2023, against the TPU v3's TPU from 2018. Newer architectures typically add lower-precision formats and better throughput per watt, so check the precision rows your workload actually uses.

Memory is 32 GB against 16 GB, fed at 900 GB/s versus 800 GB/s. More memory per GPU means larger models fit before you have to shard across cards, which often matters more than raw TFLOPS for inference.

For LLM training and inference, weight the FP8 and FP16 rows and memory capacity most heavily. For scientific and HPC workloads, the FP64 row is the one to read.

Frequently asked questions

Is the TPU v3 faster than the TPU v5e?

At BF16 precision the TPU v5e reaches 197 TFLOPS dense against 123 TFLOPS for the TPU v3. The performance table on this page lists every published precision for both GPUs.

Which has more memory, the TPU v3 or the TPU v5e?

The TPU v3 carries 32 GB of VRAM versus 16 GB for the TPU v5e. Memory bandwidth is 900 GB/s for the TPU v3 and 800 GB/s for the TPU v5e.

Get Comparison Updates

New GPUs added weekly. Be the first to see how they compare.