490-BGUT NVIDIA A30 24GB HBM2 PCIe 4.0 GPU, 165W Passive, Dual-Slot Full Height, MIG
Dell 490-BGUT is an NVIDIA A30 Ampere GPU with 24GB HBM2 (933 GB/s), PCIe 4.0, 165W passive cooling and Multi-Instance GPU, for AI inference and HPC.
Also known as · NVIDIA A30 · A30 Tensor Core GPU · A30 24GB HBM2 · Ampere A30 PCIe
1-YR ZAYNTEK WARRANTY
Overview
Dell 490-BGUT is the NVIDIA A30 data center GPU in a passively cooled, dual-slot, full-height card drawing 165W. It is based on the Ampere GA100 processor with 3,584 CUDA cores, 224 third-generation Tensor Cores and 24GB of HBM2 memory delivering 933 GB/s.
The A30 is built for mainstream AI inference, training of moderate-size models and double-precision HPC. It provides 10.3 TFLOPS of FP64 Tensor Core throughput and up to 330 TOPS of INT8 (661 TOPS with sparsity). Multi-Instance GPU (MIG) can split one card into up to four isolated 6GB instances, which lets several users or inference services share one GPU with guaranteed resources.
Two A30 cards can be linked with a third-generation NVLink bridge at 200 GB/s. Because it is passive, it relies on the server's airflow for cooling.
Specifications
General
| Dell part number | 490-BGUT |
|---|---|
| GPU | NVIDIA A30 (Ampere architecture) |
| Graphics processor | GA100 |
| CUDA cores | 3,584 |
| Launch date | April 12, 2021 |
| Tensor Cores | 224 third-generation Tensor Cores |
| GPU clocks | 930 MHz base, 1,440 MHz boost |
Memory
| Capacity | 24GB HBM2 |
|---|---|
| Bandwidth | 933 GB/s |
| Memory bus | 3,072-bit |
| Memory clock | 1,215 MHz |
| ECC | Enabled by default on HBM2, with no capacity or bandwidth overhead |
Performance (peak)
| FP64 | 5.2 TFLOPS |
|---|---|
| FP64 Tensor Core | 10.3 TFLOPS |
| FP32 | 10.3 TFLOPS |
| TF32 Tensor Core | 82 TFLOPS (165 with sparsity) |
| BF16 / FP16 Tensor Core | 165 TFLOPS (330 with sparsity) |
| INT8 Tensor Core | 330 TOPS (661 with sparsity) |
| INT4 Tensor Core | 661 TOPS (1,321 with sparsity) |
Features and interconnect
| Host interface | PCIe 4.0 x16 (64 GB/s) |
|---|---|
| NVLink | Third-generation NVLink bridge, 200 GB/s, two GPUs |
| Multi-Instance GPU | Up to 4 x 6GB, 2 x 12GB, or 1 x 24GB instances |
| Media engines | 4 video decoders (NVDEC), 1 JPEG decoder, 1 optical flow accelerator |
| Security | Secure boot and hardware root of trust (CEC1712 secondary root of trust on some boards) |
| Virtualization | SR-IOV (8 VFs); NVIDIA vGPU 13.0+ (Virtual Compute Server); NVIDIA AI Enterprise |
Physical and power
| Maximum power | 165W |
|---|---|
| Cooling | Passive (requires server airflow) |
| Form factor | Dual-slot, full height, full length (10.5 in / 267.7 mm without bracket) |
| Power connector | 1 x CPU 8-pin (EPS-12V) auxiliary connector |
| Airflow | Bidirectional heatsink (left-to-right or right-to-left) |
| Board weight | 1,240 g (without bracket, extenders or bridge) |
| Operating temperature | 0 to 55 °C ambient |
| Category | AI Infrastructure |
|---|---|
| Warranty | 1-year ZaynTek |
§ 04Questions
What is Dell 490-BGUT?
It is a Dell part number for the NVIDIA A30 data center GPU: 24GB HBM2, PCIe 4.0, 165W, passively cooled, dual-slot and full height.
How much memory bandwidth does the NVIDIA A30 have?
933 GB/s from 24GB of HBM2 memory.
Does the A30 support Multi-Instance GPU (MIG)?
Yes. One A30 can run as a single 24GB GPU, two 12GB instances or four 6GB instances.
Is the A30 good for double-precision HPC?
Yes. It delivers 10.3 TFLOPS FP64 using Tensor Cores and 5.2 TFLOPS standard FP64, which is unusual for a 165W card.
Can two A30 GPUs be linked?
Yes. A third-generation NVLink bridge connects two A30 cards at 200 GB/s.
How many Tensor Cores does the NVIDIA A30 have?
224 third-generation Tensor Cores, alongside 3,584 CUDA cores, with a 930 MHz base and 1,440 MHz boost clock.
§ 05Related — AI Infrastructure
View allNOTE · Condition is shown on each listing where an offer states it; ask for it on any quote. Units marked REFURBISHED are fully tested and carry a 1-year ZaynTek warranty.
