TPU v5p specs
Google · Shipping now · first available Dec 2023 (estimated). Dense compute, memory, bandwidth and power from the vendor's documents; each precision stays separate.
All accelerators · Compare with TPU v5e
Specifications
| Dense BF16 · FP32 accumulate | 459 TFLOPS |
|---|---|
| Dense FP8 · FP32 accumulate | 459 TFLOPS |
| Dense FP4 · FP32 accumulate | Not published |
| Memory | 102.01 GB · HBM95 GiB as published |
| Memory bandwidth | 2.77 TB/s |
| Maximum chip power | Not published |
| Interconnect | ICI 3D torus · 8,960-chip pods |
| First availability | Dec 2023est — estimated / secondary-source figure |
| Availability | Shipping |
Announced alongside Gemini 1.0 (Dec 2023), which was trained on TPU v4 and v5e.
Dense tensor rates with FP32 accumulation, never sparse or FP16-accumulate headline figures. Memory is in decimal GB. The est marker flags a value its vendor-document review did not confirm.
Rent it
$4.20/chip-hrCheapest on-demand · Google Cloud · observed
1 provider: Google Cloud. Quotes checked through .
TPU v5p rental prices from every provider →Value ratios
| $ per dense BF16 PFLOP-hour | $9.15 |
|---|---|
| Dense BF16 TFLOPS per watt | No power figure |
| Bandwidth bytes per 1,000 BF16 FLOP | 6.03 |
| $ per dense FP8 PFLOP-hour | $9.15 |
| Dense FP8 TFLOPS per watt | No power figure |
| Bandwidth bytes per 1,000 FP8 FLOP | 6.03 |
| $ per GB of memory per hour | $0.041 |
Prices divide the cheapest on-demand list price; theoretical ratios do not measure application speed.
Sources
Specifications reviewed . Per-chip table; FP8 runs at the BF16 rate and capacity is specified in GiB.
Rental price: Google Cloud published pricing ↗ (TPU v5p; per chip-hour).