inference/research
verifiedhigh · 0.90updated 2026-10-03

NVIDIA GeForce RTX 3090

nvidia ampere (GA102) 2020

Fields

id
accelerators/nvidia-rtx-3090
Vendor
nvidia
Architecture
ampere (GA102)
Release year
2020
Process (nm)
8
Form factors
pcie
VRAM (GB)
24
Memory type
gddr6x
Memory bus (bit)
384
Memory bandwidth (GB/s)
936
Bandwidth basis
GDDR6X 19.5 Gbps data rate x 384-bit / 8 = 936 GB/s (NVIDIA GB202/GB203 comparison table in the RTX Blackwell whitepaper prints 'Memory Clock (Data Rate) 19.5 Gbps' and 'Memory Bandwidth 936 GB/sec' for the 3090 column; the 3090 Ti column in the same table is 21 Gbps / 1008 GB/sec)
Peak FLOPS
precisionfp32tflops35.6denseyesvendor_claimyessource_idconsumer-nvidia-rtx-blackwell-gpu-architecture-whitepaperbasisvendor-whitepaperbasis_detailTranscribed from the vendor's own fp32 for NVIDIA GeForce RTX 3090 (source of record: 'NVIDIA RTX Blackwell GPU Architecture whitepaper (GeForce RTX 50 series)', kind: whitepaper). Recorded dense:true as printed.precisionfp16tflops71.2denseyesvendor_claimyessource_idconsumer-nvidia-rtx-blackwell-gpu-architecture-whitepaperbasisvendor-whitepaperbasis_detailTranscribed from the vendor's own fp16 for NVIDIA GeForce RTX 3090 (source of record: 'NVIDIA RTX Blackwell GPU Architecture whitepaper (GeForce RTX 50 series)', kind: whitepaper). Recorded dense:true as printed.
TDP (W)
350
Unified memory
no
Consumer part
yes
notes
Consumer card, Ampere GA102, Samsung 8N, 10496 CUDA cores, 350 W TGP. Bandwidth: recorded as 936 GB/s (19.5 Gbps part). Note that the RTX 3090 Ti in the SAME NVIDIA whitepaper table is 21 Gbps / 1008 GB/sec. The two 24 GB cards are not interchangeable on bandwidth; recommendations that quote '24 GB RTX 3090' from a Ti benchmark silently overstate this card. Tensor precision: FP16 tensor with FP32 accumulate is 71.2 TFLOPS dense (the Ada whitepaper's GA102 column shows 'Peak FP16 Tensor TFLOPS with FP16 Accumulate 142.3/284.6', i.e. double with FP16 accumulate, and half again in sparse). LLM inference kernels normally accumulate in FP32, so 71.2 is the practically relevant figure. FP8/INT8/INT4 tensor throughput for GA102 was NOT verified against a primary document in this pass and is deliberately omitted rather than estimated. For LLM inference: 24 GB is a hard ceiling on resident weights. Compare against Apple unified-memory parts on both capacity AND addressability - see the Apple records, where a 128 GB unified pool is not 5x a 24 GB VRAM pool. INTERCONNECT FILLED 2026-10-04: interconnect was empty; it now reads ['pcie-gen4'], from the vendor's own spec table (source the GeForce RTX 3090 family product page's 'Technology Support' row). SCALE-UP IS ABSENT AND THAT IS THE POINT: this part has no NVLink and no other GPU-to-GPU fabric, so the host link is the ONLY interconnect a multi-GPU deployment has. Do not infer a fabric from the presence of a PCIe value, and do not read the Gen number as a multi-GPU capability.

Notes

Consumer card, Ampere GA102, Samsung 8N, 10496 CUDA cores, 350 W TGP.

Bandwidth: recorded as 936 GB/s (19.5 Gbps part). Note that the RTX 3090 Ti in the SAME NVIDIA whitepaper table is 21 Gbps / 1008 GB/sec. The two 24 GB cards are not interchangeable on bandwidth; recommendations that quote '24 GB RTX 3090' from a Ti benchmark silently overstate this card.

Tensor precision: FP16 tensor with FP32 accumulate is 71.2 TFLOPS dense (the Ada whitepaper's GA102 column shows 'Peak FP16 Tensor TFLOPS with FP16 Accumulate 142.3/284.6', i.e. double with FP16 accumulate, and half again in sparse). LLM inference kernels normally accumulate in FP32, so 71.2 is the practically relevant figure. FP8/INT8/INT4 tensor throughput for GA102 was NOT verified against a primary document in this pass and is deliberately omitted rather than estimated.

For LLM inference: 24 GB is a hard ceiling on resident weights. Compare against Apple unified-memory parts on both capacity AND addressability - see the Apple records, where a 128 GB unified pool is not 5x a 24 GB VRAM pool. INTERCONNECT FILLED 2026-10-04: interconnect was empty; it now reads ['pcie-gen4'], from the vendor's own spec table (source the GeForce RTX 3090 family product page's 'Technology Support' row). SCALE-UP IS ABSENT AND THAT IS THE POINT: this part has no NVLink and no other GPU-to-GPU fabric, so the host link is the ONLY interconnect a multi-GPU deployment has. Do not infer a fabric from the presence of a PCIe value, and do not read the Gen number as a multi-GPU capability.

Cross-references

Interconnect (1)