As a part of a collection of SIGGRAPH-related bulletins, NVIDIA this morning is asserting a brand new addition to their RTX line {of professional} video playing cards, the RTX A2000. Following NVIDIA’s transitional top-to-bottom product rollouts, the most recent member of NVIDIA’s professional GPU lineup is the smallest and least expensive card but, thanks largely to its use of NVIDIA’s GA106 GPU. Paired with 6GB of VRAM and aimed partially at small type issue workstations, NVIDIA expects the cardboard to go on sale in October.
Though a relative light-weight in comparison with the way more highly effective RTX A6000, the A2000 is an fascinating (and in some methods a lot delayed) improvement in NVIDIA’s skilled visualization product stack. Despite the truth that NVIDIA has been transport ray tracing-capable {hardware} for nearly three years now, the corporate has by no means beforehand launched a mid-range ProViz card with that sort performance – the previous-generation Quadro RTX Stack stopped on the RTX 4000, and under that, all the things has remained Pascal. And although that is in the end primarily based on the truth that NVIDIA solely put RT {hardware} of their higher-end Turing GPUs, the tip end result is identical: that is NVIDIA’s first RT-capable mid-range professional card.
| NVIDIA Professional Visualization Card Specification Comparison |
||||||
| A6000 | A5000 | A4000 | A2000 | |||
| CUDA Cores | 10752 | 8192 | 6144 | 3328 | ||
| Tensor Cores | 336 | 256 | 192 | 104 | ||
| Boost Clock | 1800MHz | 1695MHz | 1560MHz | 1200MHz | ||
| Memory Clock | 16Gbps GDDR6 | 16Gbps GDDR6 | 14Gbps GDDR6 | 12Gbps GDDR6 |
||
| Memory Bus Width | 384-bit | 384-bit | 256-bit | 192-bit | ||
| VRAM | 48GB | 24GB | 16GB | 6GB | ||
| ECC | Partial (DRAM) |
Partial (DRAM) |
Partial (DRAM) |
Partial (DRAM) |
||
| Single Precision | 38.7 TFLOPS | 27.eight TFLOPS | 19.2 TFLOPS | eight TFLOPS | ||
| Tensor Performance | 309.7 TFLOPS | 222.2 TFLOPS | 153.four TFLOPS | 63.9 TFLOPS | ||
| TDP | 300W | 230W | 140W | 70W | ||
| Cooling | Active | Active | Active | Active | ||
| NVLink | 1x NVLink3 112.5GB/sec |
1x NVLink3 112.5GB/sec |
N/A | N/A | ||
| GPU | GA102 | GA102 | GA104 | GA106 | ||
| Architecture | Ampere | Ampere | Ampere | Ampere | ||
| Manufacturing Process | Samsung 8nm | Samsung 8nm | Samsung 8nm | Samsung 8nm | ||
| Launch Price | N/A | N/A | N/A | $450 | ||
| Launch Date | 12/2020 | 04/2021 | 04/2021 | 10/2021 | ||
Fittingly, it’s additionally the primary card launched by NVIDIA because the Quadro P period that’s obtainable in a low-profile design. Thanks to NVIDIA considerably turning down the GA106 GPU’s clockspeeds, the RTX A2000 has a TDP of simply 70 watts. This retains the warmth generated by the cardboard low sufficient that it may be cooled by a dual-slot low-profile cooler – making it appropriate to be used in true small type issue workstations – albeit at the price of capping the cardboard’s total efficiency.
Digging down into the specs, what we discover is a GA106-based video card with a number of SMs disabled and a comparatively low increase clockspeed of simply 1200MHz. On paper, this offers the cardboard roughly eight TFLOPS of single precision (FP32) graphics and compute efficiency. Or, tapping into the cardboard’s 104 tensor cores, it could wring out nearly 64 TFLOPS of efficiency with tensor operations.
This is backed by 6GB of GDDR6 reminiscence, sitting on prime of a 192-bit reminiscence bus. Like the GPU clockspeed, NVIDIA is protecting the reminiscence clockspeed on the low facet as nicely (presumably once more for energy causes), leading to a reminiscence clockspeed of 12Gbps. Thankfully, the broader reminiscence bus offsets this considerably, placing the ultimate reminiscence bandwidth at a good 448GB/second. Meanwhile, as this can be a ProViz card, like the remainder of the RTX A sequence the…
