T4 Specifications
70 W and single-slot, so it fits anywhere. Still common for light inference and VDI, but Turing is two generations behind.
Buy it for
Cheap 24 GB for light inference and VDI, in servers with no spare power budget.
Look elsewhere if
You are running modern models — Turing has no BF16 and no FP8, so current training recipes need rework.
Will it fit your system?
Passively cooled — needs server airflow
This card has no fans. It relies on a chassis forcing air through its heatsink. In a normal desktop it will throttle within minutes and can overheat outright. It needs a rack server with high-static-pressure fans, or an aftermarket shroud and blower.
- Form factor
- PCIe single-slot
- Cooling
- passive (no fans)
- Power connectors
- None — slot-powered
- Board power
- 70 W
- Suggested PSU
- 400 W or better
- Interface
- PCIe Gen 3
PSU figure assumes one card plus a mainstream CPU and drives, with headroom for transient spikes — modern cards can draw well above their rated TDP for milliseconds, which is what trips an undersized supply.
Supported formats & features
What the T4 does and does not support in hardware.
Blackwell-generation feature — B200 has it
Predates hardware FP8; use INT8 or BF16 instead — H100 has it
FP16 only, so training recipes may need loss scaling
Not available on this architecture
Quantised inference supported
INT4 tensor operations supported
Precision must be managed manually
Multi-GPU traffic goes over PCIe — a real limit for distributed training
Hardware ray tracing for rendering workloads
Corrects single-bit errors — expected for long training runs
What fits in memory
Largest model that fits on one T4, assuming ~20% of the 16 GB is reserved for KV cache and activations.
| Model | INT4 / FP4 |
|---|---|
| Llama 8B | Fits |
| Llama 70B | 3× GPUs |
| Mixtral 8x22B | 6× GPUs |
| Llama 405B | 16× GPUs |
Weights only. Long contexts, large batch sizes and optimiser state during training all need substantially more — full fine-tuning typically needs 3–4× the weight memory.
Workload character
Bandwidth and compute are matched closely enough that neither dominates across mixed training and inference.
Cost per unit of compute
Breakeven counts hardware cost against rental only. Power, cooling, hosting and depreciation push the real crossover substantially later — see colocation pricing for what hosting actually costs.
Cost to run
Electricity for one card at 100% duty, including cooling and PSU overhead (858 kWh/year).
At the US average rate, three years of continuous operation costs $309 — roughly 15% of the card's $2,000 launch price, before any cooling or hosting fees.
Where it sits in the line
A newer part existing does not make this one a bad buy — it makes it cheaper. What matters is price per unit of the throughput you actually need.
Availability & market
Very cheap used and still common in cloud fleets, but two generations past its useful life for new AI work.
Discontinued. Expect secondary-market stock only, with no manufacturer warranty and limited driver support ahead.
Memory
- Capacity
- 16 GB
- Type
- GDDR6
- Bandwidth
- 320 GB/s
- Bus width
- 256-bit
- GB per TFLOPMemory relative to compute — higher favours large models
- 0.246
Compute
- FP64
- —
- FP32
- 8.1 TFLOPS
- TF32 dense
- —
- FP16 dense
- 65 TFLOPS
- FP16 sparseRequires structured sparsity; rarely reached in practice
- —
- FP8 dense
- —
- INT8 dense
- 130 TOPS
Silicon
- Codename
- TU104
- Process node
- TSMC 12nm FFN
- Transistors
- 13.6B
- Die size
- 545 mm²
- CUDA cores
- 2,560
- Tensor cores
- 320
- SM count
- 40
- Boost clock
- 1,590 MHz
Platform & power
- TDP
- 70 W
- TFLOPS per wattDense FP16 ÷ TDP — drives your power bill at scale
- 0.93
- Form factor
- PCIe single-slot
- Cooling
- passive
- Interconnect
- None
- Interconnect speed
- —
- PCIe generation
- Gen 3
- MIG
- No
- Launch
- 2018-09
- MSRP at launch
- $2,000
T4 — common questions
Answered from the specifications above. Questions this card's data cannot answer with confidence are left out rather than guessed.
- How much VRAM does the T4 have?
- The T4 has 16 GB of GDDR6 memory with 320 GB/s of bandwidth. In practice plan on roughly 12 GB for weights: KV cache, activations and fragmentation take the rest, and long contexts or large batches take considerably more.
- What size model can the T4 run?
- On a single T4, about 6B parameters at FP16, or roughly 25B quantised to FP4. Anything larger has to be split across several GPUs, which is slow on this card because it has no NVLink. These are weight-only figures — a long context window can consume tens of gigabytes on its own.
- Can you put the T4 in a desktop PC?
- Only with added cooling. The T4 is a passively cooled card — it has no fans at all and depends on a rack chassis forcing air through its heatsink. Dropped into an ordinary desktop it will throttle within minutes and can overheat outright. It needs either a rack server with high-static-pressure fans or an aftermarket shroud and blower.
- How much power does the T4 use?
- The T4 is rated at 70 W board power. It is fed through None — slot-powered. For a single-card system, a 400 W supply or better leaves headroom for the CPU, drives and the transient spikes that trip undersized units.
- How much does it cost to run the T4?
- At the US commercial average of $0.12/kWh, running one T4 continuously costs about $103 a year in electricity. That figure includes cooling and power-supply overhead at a factor of 1.4 over board power, and over three years comes to roughly 15% of what the card cost at launch.
- Does the T4 support FP8?
- No. The T4 has no hardware FP8 path. FP8-quantised checkpoints will still load, but the values are upcast before they reach the tensor cores, so you get the memory saving without the speedup.
- Does the T4 support FP4?
- No. The T4 predates hardware FP4. You can run 4-bit quantised weights on it and the memory saving is real, but they are dequantised before compute, so there is no arithmetic speedup.
- Does the T4 support INT8?
- Yes. The T4 supports INT8 acceleration at 130 dense TOPS. INT8 remains the most widely supported quantised inference format across frameworks.
- Does the T4 support BF16?
- No. The T4 predates bfloat16 support and can only use FP16 for half-precision work. FP16 training on it usually needs loss scaling to stay numerically stable.
- Does the T4 have NVLink?
- No. The T4 has no NVLink. Multi-GPU work goes over PCIe Gen 3, which is an order of magnitude slower and makes tensor parallelism across cards expensive.
- Does the T4 have RT cores?
- Yes. The T4 includes dedicated ray-tracing cores, so it is usable for rendering and visualisation as well as compute.
- Does the T4 have ECC memory?
- Yes. The T4 has ECC, so single-bit memory errors are corrected rather than silently corrupting a long run.
- What replaced the T4?
- The L4 (2023) is its successor, with 24 GB of GDDR6 at 300 GB/s. It is about 1.9x faster in dense FP16. On the used market that gap is what sets a fair price for the T4.
- Is the T4 worth buying?
- Buy it for: Cheap 24 GB for light inference and VDI, in servers with no spare power budget. Look elsewhere if: You are running modern models — Turing has no BF16 and no FP8, so current training recipes need rework.