Google Cloud TPU v5p Pod
Pod / clusterShippingv5p2024
Google's high-performance TPU pod for large-scale training. 8,960 TPU v5p chips in a single pod connected via ICI 3.0 fabric. Powers Gemini 1.5 and internal Google AI training. Only available on Google Cloud. The largest single training domain outside of xAI's Colossus.
8960
GPUs per system
8100 FP8 PFLOPS
Total HBM
860 TB
Host memory
4480 TB
Interconnect
ICI 3.0
4600 TB/s
Networking
6400 Gbps
Storage
10000 TB Persistent SSD
Form factor
Multi-rack pod (64+ racks)
Weight
TBD
Rack units
TBD
Performance
Manufacturer datasheet values · aggregate system compute
| FP4 PFLOPS | TBD |
| FP8 PFLOPS | 8,100 |
| FP16 PFLOPS | 4,050 |
| BF16 PFLOPS | 4,050 |
| Training effective PFLOPS | 3,200 |
Power and cooling
Thermal envelope · cooling requirements · efficiency
Rack power
2500 kW
Per GPU
250 W
Cooling
liquid
PUE estimate
1.1
Power draw relative to tracked systems2500 kW / 2500 kW max
3.24 FP8 PFLOPS per kW · average across all systems is 4.81
TCO analysis
Hardware amortized over 3 years · power at $0.05/kWh
List price
TBD
Lease monthly
$12,000,000/mo
Cost per GPU per month
$1,339
TCO per PFLOPS per year
$17,926
PFLOPS per kW
3.24
98% above the average TCO of $9,046/PFLOPS/year across all tracked systems
Available from
GC
Google CloudKnown deployments
Disclosed in press releases, SEC filings, and conference talks
Quantity
Gemini training
Source
Google research blog
Quantity
Claude training (GCP)
Source
Anthropic partnership announcementSources
Every data point on this page is reproducible
Other AI systems
Compare across the system landscape
DGX GB300 NVL7272x NVIDIA B300 · 1440 PFLOPS
Full rackAnnounced
DGX GB200 NVL7272x NVIDIA B200 · 720 PFLOPS
Full rackShipping
TPU v6e Pod256x Google TPU v6e · 230 PFLOPS
Pod / clusterShipping
Maia 100 Rack32x Microsoft Maia 100 · 96 PFLOPS
Full rackRamping
DGX B2008x NVIDIA B200 · 80 PFLOPS
Server nodeShipping
MI325X Platform8x AMD MI325X · 48 PFLOPS
Server nodeShipping