A10G Specifications
AWS-specific A10 variant behind g5 instances. 150 W single-slot makes it dense and cheap to run.
Buy it for
Cloud inference and graphics on AWS G5, which is essentially the only place it exists. 24 GB at 150 W makes it cheap to rent per hour.
Look elsewhere if
You are buying hardware. It is an AWS-specific variant not sold through normal channels, so it is a rental decision, not a purchase decision.
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
- 1 × 8-pin EPS (CPU-style — not a PCIe 8-pin)
- Board power
- 150 W
- Suggested PSU
- 500 W or better
- Interface
- PCIe Gen 4
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 A10G does and does not support in hardware.
Blackwell-generation feature — B200 has it
Predates hardware FP8; use INT8 or BF16 instead — H100 has it
Wider exponent than FP16 — fewer overflow issues when training
Drop-in FP32 speedup with no code changes
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 A10G, assuming ~20% of the 24 GB is reserved for KV cache and activations.
| Model | FP16 / BF16 | INT4 / FP4 |
|---|---|---|
| Llama 8B | Fits | Fits |
| Llama 70B | 8× GPUs | 2× GPUs |
| Mixtral 8x22B | 15× GPUs | 4× GPUs |
| Llama 405B | 43× GPUs | 11× 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 (1,840 kWh/year).
At the US average rate, three years of continuous operation costs $662 — roughly 22% of the card's $3,000 launch price, before any cooling or hosting fees.
Availability & market
Compare per-hour against the L4, which is newer, uses less power, and is available from more providers.
Memory
- Capacity
- 24 GB
- Type
- GDDR6
- Bandwidth
- 600 GB/s
- Bus width
- 384-bit
- GB per TFLOPMemory relative to compute — higher favours large models
- 0.192
Compute
- FP64
- —
- FP32
- 31.5 TFLOPS
- TF32 dense
- —
- FP16 dense
- 125 TFLOPS
- FP16 sparseRequires structured sparsity; rarely reached in practice
- 250 TFLOPS
- FP8 dense
- —
- INT8 dense
- —
Silicon
- Codename
- GA102
- Process node
- Samsung 8N
- Transistors
- 28.3B
- Die size
- 628 mm²
- CUDA cores
- 9,216
- Tensor cores
- 288
- SM count
- 72
- Boost clock
- 1,710 MHz
Platform & power
- TDP
- 150 W
- TFLOPS per wattDense FP16 ÷ TDP — drives your power bill at scale
- 0.83
- Form factor
- PCIe single-slot
- Cooling
- passive
- Interconnect
- None
- Interconnect speed
- —
- PCIe generation
- Gen 4
- MIG
- No
- Launch
- 2021-04
- MSRP at launch
- $3,000
A10G — 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 A10G have?
- The A10G has 24 GB of GDDR6 memory with 600 GB/s of bandwidth. In practice plan on roughly 19 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 A10G run?
- On a single A10G, about 9B parameters at FP16, or roughly 38B 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 A10G in a desktop PC?
- Only with added cooling. The A10G 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 A10G use?
- The A10G is rated at 150 W board power. It is fed through 1 × 8-pin EPS (CPU-style — not a PCIe 8-pin). For a single-card system, a 500 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 A10G?
- At the US commercial average of $0.12/kWh, running one A10G continuously costs about $221 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 22% of what the card cost at launch.
- Does the A10G support FP8?
- No. The A10G 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 A10G support FP4?
- No. The A10G 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 A10G support INT8?
- Yes. The A10G supports INT8 acceleration. INT8 remains the most widely supported quantised inference format across frameworks.
- Does the A10G support BF16?
- Yes. The A10G supports bfloat16. Its wider exponent range makes training far less prone to the overflow that plagues FP16, which is why most current training recipes assume it.
- Does the A10G have NVLink?
- No. The A10G has no NVLink. Multi-GPU work goes over PCIe Gen 4, which is an order of magnitude slower and makes tensor parallelism across cards expensive.
- Does the A10G have RT cores?
- Yes. The A10G includes dedicated ray-tracing cores, so it is usable for rendering and visualisation as well as compute.
- Does the A10G have ECC memory?
- Yes. The A10G has ECC, so single-bit memory errors are corrected rather than silently corrupting a long run.
- Is the A10G worth buying?
- Buy it for: Cloud inference and graphics on AWS G5, which is essentially the only place it exists. 24 GB at 150 W makes it cheap to rent per hour. Look elsewhere if: You are buying hardware. It is an AWS-specific variant not sold through normal channels, so it is a rental decision, not a purchase decision.