DEPLOYDatabase

Chip comparison

Google TPU v4 vs NVIDIA L40S

Side-by-side specs

Straight from each vendor's datasheet. Winner in bold: higher throughput, memory, or transistor count; lower power draw or price. Sparse (2:4) throughput roughly doubles the FP8/INT8 numbers shown.

FieldGoogle TPU v4NVIDIA L40S
Process nodeTSMC N7TSMC 4N
Transistors (B)76.3
TDP192 W350 W
Memory32 GB HBM248 GB GDDR6
Memory bandwidth1,200 GB/s864 GB/s
FP16 (dense)362 TFLOPS
BF16 (dense)275 TFLOPS362 TFLOPS
FP8 (dense)733 TFLOPS
INT8 (dense)275 TOPS733 TOPS
Form factorOAM (per-chip in v4 pod)PCIe
Announced2021-05-182023-08-08
Released2022-05-01

Robots running Google TPU v4

No robots publicly running it yet.

Robots running NVIDIA L40S

No robots publicly running it yet.

Data centers with NVIDIA L40S

No data centers publicly running it yet.

Common questions

Google TPU v4 vs NVIDIA L40S: which is faster for training?

NVIDIA L40S has 2.67x the dense FP16/BF16 throughput of the other (Google TPU v4: 275 TFLOPS; NVIDIA L40S: 733 TFLOPS). Real-world training speed also depends on memory bandwidth, interconnect, and software stack: see the spec diff for those inputs.

Google TPU v4 vs NVIDIA L40S: which has more memory?

NVIDIA L40S carries more HBM (Google TPU v4: 32 GB HBM2; NVIDIA L40S: 48 GB GDDR6). Memory capacity sets the largest model that fits on one chip without partitioning; memory bandwidth sets inference throughput.

Google TPU v4 vs NVIDIA L40S: which is more power-efficient?

Dense FP16/BF16 TFLOPS per watt: Google TPU v4 1.43 (275 ÷ 192 W); NVIDIA L40S 2.09 (733 ÷ 350 W). NVIDIA L40S wins at the die level; system-level efficiency also depends on cooling and interconnect.

Google TPU v4 vs NVIDIA L40S: which is more widely used?

Google TPU v4: 0 robots and 2 data centers publicly running it. NVIDIA L40S: 0 robots and 0 data centers. Empty here means no operator has said publicly, not zero adoption in the market.

See also: every chip comparison · Google TPU v4 full page · NVIDIA L40S full page.