RTX A5000 and RTX A4000, but not only: Nvidia Ampere invades the professional world
Nvidia announced a series of new products professional architecture-based Ampere. For the desktop workstation sector, here are the solutions RTX A5000 e RTX A4000, two profoundly different cards following the RTX A6000 unveiled last October.
On board of RTX A5000 we find a GA102 GPU – 8192 CUDA core, 256 Tensor core and 64 RT core – flanked by 24 GB of GDDR6 memory at 16 Gbps on 384-bit bus. The TDP is 230W. The card therefore has 2560 fewer CUDA cores than the RTX A6000 and half the video memory, even if the bandwidth is the same, equal to 768 GB / s. The FP32 computing power is equal to 27.8 TFLOPs, less than the 38.7 TFLOPs of the older sister.
Nvidia RTX A5000
The RTX A4000 instead it is based on a GA104 GPU with 6144 active CUDA cores (192 Tensor core e 48 RT core) e 16 GB of GDDR6 memory on a 256-bit bus, for a bandwidth of 448 GB / s. The TDP is 140W. FP32 computing power settles at 19.2 TFLOPs. Both models have a PCI Express 4.0 x16 connection interface, but only RTX A5000 supports NVLink (for parallel connection with another RTX A5000) and, in addition to GPU passthrough, Single Root Input / Output Virtualization (SR-IOV) functionality.
Nvidia RTX A4000
For the new RTX A5000 Nvidia has opted for a dual-slot cooling system, while the RTX A4000 takes up one slot. They both have four DisplayPort 1.4 and will be available later this month at undisclosed prices.
The US company has also thought about workstation mobile con four new GPUs calls RTX A2000, A3000, A4000 and A5000 arriving in the second quarter. The first model is based on a GA106 GPU with 2560 active CUDA cores, 80 Tensor cores and 20 RT cores, while the A3000 e A4000 they adopt a chip GA104. In the case of the RTX A3000 we have 4096 CUDA cores, while in that of the RTX A4000 we find 5120 CUDA cores. Finally, the RTX A5000 mobile is based on a GA102 GPU with 6144 CUDA cores, therefore a decidedly neutered version with respect to the potential of the chip.
The “old” Turing architecture finds its place in a sea of Amps with T1200 and T600, two graphics chips always dedicated to mobile workstations but low-end, in fact the chips have respectively 1024 and 896 CUDA cores for a power of 3.7 and 2.5 TFLOPs in combination with 4 GB of GDDR6 memory on 128 bit bus ( for a bandwidth of 160 and 192 GB / s respectively). This is Turing GPU without RT and Tensor core, such as those seen in the GeForce GTX 1600 range last year. The TDP is 35-95W for the T1200 depending on the configuration chosen by the OEM, while in the case of the T600 we are talking about 25W.
For the datacenter world, here is Nvidia A30, A16 e A10, which are added to an offer that already includes the A100 and A40 models. Much information is still missing from these proposals, but we know that The A30 is based on a GA100 GPU and offers 24 GB of HBM2 memory on a 3072-bit bus for a bandwidth of 933 GB / s and a TDP of 65W. In the case of A10instead, we have a chip GA102 con 9216 CUDA core, 288 Tensor core and 72 RT core supported by 24 GB of GDDR6 memory. Thanks to the 384-bit bus, the bandwidth reaches 600 GB / s. The TDP, in this case, is equal to 150W.
Last but not least, Nvidia A16. We are faced with a “special” product if you pass the term there because it sees the presence of four GPU Amps on a single PCB. Each chip is accompanied by 16GB of GDDR6 memory, for a total of 64GB. Nvidia A16 proposes itself as a multimedia accelerator (offers four NVENC encoders and eight NVDEC decoders) or a vGPU solution.
All datacenter models support PCIe 4.0 and have no video outputs, but only the A100, A40 and A30 can be combined with the same model via NVLink. In all cases, the cooling is passive and the power supply is entrusted to an eight-pin EPS12V connector. A30 and A10 should be available by the end of the month, while A16 will have to wait until the end of the year.