COMPARE / T4 VS V100
T4 vs V100
Specs from the chip makers' datasheets; prices from the live Compute Cafe panel, refreshed hourly. Every rate is an on-demand published price with a source receipt.
CHEAPEST T4
$0.070/hr
SpotGPUs · median $1.09 · 16 providers
CHEAPEST V100
$0.170/hr
Akash · median $3.70 · 11 providers
CHEAPER RIGHT NOW
T4
live panel, 2026-09-27
| T4 | V100 | |
|---|---|---|
| MAKER | NVIDIA | NVIDIA |
| ARCHITECTURE | Turing | Volta |
| LAUNCHED | 2018 | 2017 |
| MEMORY | 16 GB GDDR6 | 16 or 32 GB HBM2 |
| MEMORY BANDWIDTH | 320 GB/s | 900 GB/s |
| TENSOR COMPUTE | 65 TFLOPS FP16 tensor (130 TOPS INT8) | 125 TFLOPS FP16 tensor (SXM2) |
| POWER DRAW | 70 W | 300 W (SXM2) / 250 W (PCIe) |
| INTERCONNECT | PCIe 3.0 x16 | NVLink 2.0 300 GB/s (SXM2) or PCIe 3.0 |
HOW TO CHOOSE
V100 vs T4
A T4 is far cheaper and newer, but has one-fifth the memory bandwidth and no NVLink. For training, the V100 wins; for light serving, the T4 wins on cost.
Workload fit
T4: small-model inference, video transcoding, fine-tuning. V100: budget training, batch inference, legacy cuda stacks.
T4 VS V100 - FAQ
Which is cheaper to rent, T4 or V100?
By today's published on-demand rates: T4 starts at $0.070/GPU/hr (median $1.09, 16 providers) and V100 starts at $0.170/GPU/hr (median $3.70, 11 providers). The cheapest live listing is the T4 at SpotGPUs. Figures recompute from the panel on every refresh.
T4 or V100 - which should I rent?
T4: The T4 is NVIDIA's low-power inference workhorse: a 70-watt single-slot card that became the default cost floor for model serving. V100: The V100 was the first GPU with tensor cores and dominated deep-learning training from 2017 to 2020. The side-by-side spec table and live price rows above give the full picture.
Where can I rent T4 and V100?
16 providers publish T4 rates and 11 publish V100 rates on Compute Cafe. Every row links to the provider's own published price - committed-use and reserved rates are excluded.
Full pricing: T4 · V100 · all comparisons