NVIDIA Quadro RTX 8000 (Turing, 48 GB)
nvidia turing (TU102) 2018
Fields
id
accelerators/nvidia-quadro-rtx-8000Vendor
nvidia
Architecture
turing (TU102)
Release year
2018
Process (nm)
12
Form factors
pcieworkstation
VRAM (GB)
48
Memory type
gddr6
Memory bus (bit)
384
Memory bandwidth (GB/s)
672
Bandwidth basis
Datasheet prints 'Memory Bandwidth 672 GB/s', 'Memory Interface 384-bit', 'GPU Memory 48 GB GDDR6'. Same bandwidth as the Quadro RTX 6000; only capacity doubles.
Peak FLOPS
precisionfp32tflops16.3denseyesvendor_claimyessource_idnv-quadro-rtx-8000-datasheetbasisvendor-spec-sheetbasis_detailDatasheet 'Single-Precision Performance 16.3 TFLOPS'. Same TU102 config as the Quadro RTX 6000 (4608 CUDA cores), so the number is identical and dense:true.unitTFLOPS
TDP (W)
295
Unified memory
no
Consumer part
no
notes
USED-MARKET 48 GB CARD. Same 4608 CUDA / 576 Tensor / 72 RT core TU102 as the Quadro RTX 6000 with memory doubled to 48 GB and board power raised to 295 W (total graphics power 260 W). NVLink HB bridge quoted at 100 GB/s bidirectional for 96 GB combined, application support required. Bandwidth is the trap: 672 GB/s is well under the RTX A6000 (768 GB/s) and RTX 6000 Ada (960 GB/s), so a cheap 48 GB listing is not an A6000. ECC, active dual-slot, PCIe 3.0 x16.
Notes
USED-MARKET 48 GB CARD. Same 4608 CUDA / 576 Tensor / 72 RT core TU102 as the Quadro RTX 6000 with memory doubled to 48 GB and board power raised to 295 W (total graphics power 260 W). NVLink HB bridge quoted at 100 GB/s bidirectional for 96 GB combined, application support required. Bandwidth is the trap: 672 GB/s is well under the RTX A6000 (768 GB/s) and RTX 6000 Ada (960 GB/s), so a cheap 48 GB listing is not an A6000. ECC, active dual-slot, PCIe 3.0 x16.