Nvidia introduces the RTX A6000 and A40. The former is an Ampere video card for workstations and the A40 is a variant for data centers. The GPUs have 48GB GDDR6 with error correcting code. Nvidia seems to be dropping the Quadro naming convention. The RTX A6000 and A40 are based on the GA102 GPU, as are the GeForce RTX 3090 and 3080. The new cards for workstations and servers have the GPU fully enabled, which is good for 10,752 Cuda cores. The GPUs for gamers have a partially disabled GPU, with 10,496 and 8704 cores for the RTX 3090 and 3080 respectively. NVidia uses GDDR6 memory on the RTX A6000 and A40, while the GeForce variants have Gddr6x. However, this faster memory is not yet available in chips with a capacity of 16Gb / s and because Nvidia wants to provide the workstation cards with 48GB, regular GDDR6 has been chosen. The A40 is almost the same as the RTX A6000. The A40 model is made for data centers and has no active cooling. The card must be placed in a server housing where a lot of airflow is already present. Nvidia provides the A40 with DisplayPort outputs; this was not the case with earlier Tesla GPUs for data centers. The new GPUs will be released at the end of this year, prices has not been announced by Nvidia. It is striking that Nvidia does not mention the Quadro name in its announcement. The RTX A6000 is the successor to Quadro workstation cards such as the RTX 8000 and RTX 6000. Earlier this year, Nvidia stopped using the Tesla naming convention, which was previously linked to GPUs for data centers. The manufacturer would no longer use that name because of the car brand of the same name.
| Nvidia video cards for workstations and data centers | ||||||
| RTX A6000 | A40 | RTX 8000 | GV100 | |||
| Cuda cores | 10752 | 10752 | 4608 | 5120 | ||
| Tensor scores | 336 | 336 | 576 | 640 | ||
| Boostclock | ? | ? | 1770MHz | ~ 1450MHz | ||
| Memory speed | 16Gbps GDDR6 | 14.5Gbps GDDR6 | 14Gbps GDDR6 | 1.7Gbps HBM2 | ||
| Memory bus | 384-bit | 384-bit | 384-bit | 4096-bit | ||
| Vram | 48GB | 48GB | 48GB | 32GB | ||
| Ecc | Partly (dram) | Partly (dram) | Partly (dram) | Fully | ||
| Half Precision | ? | ? | 32.6 TFLOPS | 29.6 TFLOPS | ||
| Single Precision | ? | ? | 16.3 TFLOPS | 14.8 TFLOPS | ||
| Tensor Performance | ? | ? | 130.5 TFLOPS | 118.5 TFLOPs (FP16) |
||
| TDP | 300W | 300W | 295W | 250W | ||
| Cooling | Active | Passive | Active | Active | ||
| NVLink | 1x NVLink3 112.5GB / sec |
1x NVLink3 112.5GB / sec |
1x NVLInk2 50GB / sec |
2x NVLInk2 100GB / sec |
||
| GPU | GA102 | GA102 | TU102 | GV100 | ||
| Architecture | Ampere | Ampere | Turing | Volta | ||
| Manufacturing Process | Samsung 8nm | Samsung 8nm | TSMC 12nm FFN | TSMC 12nm FFN | ||
| Launch Price | ? | ? | $ 10,000 | $ 9,000 | ||
| Launch Date | 12/2020 | Q1 2021 | Q4 2018 | March 2018 | ||
Table compiled by AnandTech.