Just two weeks have passed since Nvidia lifted the veil for the Geforce RTX 3000 series graphics card, with the architecture Ampere under the hood. The top card in the lineup is the Geforce RTX 3090, which is equipped with a staggering 10,496 CUDA cores and 24 GB of GDDR6X memory. The graphics circuit that forms the basis of the top card is GA102, which in its full capacity houses 84 calculation units (SM) and 10,752 CUDA cores.
With the consumer card Geforce RTX 3090, Nvidia has shut down two SM units and their associated ray tracing-block RT Core and Tensor cores. A report from Videocardz reveals that Nvidia is working on new models in the Quadro series of graphics cards for workstations, where the models utilize the full capacity of GA102, which means 10,752 CUDA cores.
A major difference from the Geforce RTX 3090 is that Nvidia with the Ampere-cast Quadro RTX does not use GDDR6X as the memory type. Instead, the same GDDR6 is used as in the product family’s current top card Quadro RTX 8000, which also offers the same maximum memory of 48 GB GDDR6. The memory works over a 384-bit memory controller and the speed 16,000 MHz, which gives a total bandwidth of 768 GB / s.
It can be compared to the Geforce RTX 3090, whose GDDR6X memory operates at 19,500 MHz for a memory bandwidth of 936 GB / s over the same 384-bit memory interface. However, the choice is logical as the amount of memory is a more important factor than the highest possible bandwidth for the type of data workstations process.
RTX ‘Ampere’ panel | Geforce RTX 3090 | RTX 8000 panel | RTX 6000 panel | RTX 5000 panel | |
---|---|---|---|---|---|
Architecture | GA102 Ampere | GA102 Ampere | TU102 Turing | TU102 Turing | TU104 Turing |
CUDA cores | 10 752 | 10 496 | 4 608 | 4 608 | 3 072 |
Graphics memory | 48 GB GDDR6 | 24 GB GDDR6X | 48 GB GDDR6 | 24 GB GDDR6 | 16 GB GDDR6 |
Memory speed | 16 000 MHz | 19 500 MHz | 14 000 MHz | 13 000 MHz | 14 000 MHz |
Memory bandwidth | 768 GB/s | 936 GB/s | 672 GB/s | 624 GB/s | 448 GB/s |
Guide price | Unknown | 1 499 USD | 10 000 USD | 6 300 USD | 2 300 USD |
An overview of the specifications indicates a real shot up to CUDA cores, where 10,752 at the Ampere newcomer overshadows the Turing-based Quadro RTX 8000 with its 4,608 CUDA cores. The tensor cores are fewer in number, where the Quadro RTX in Ampere design is equipped with 336 Tensor cores and the Turing equivalent has 576 pieces.
As with Ampere for consumers, however, these consist of third-generation technology which, according to Nvidia, is twice as fast as its predecessors. The number of RT Cores in Quadro RTX “Ampere” measures in at 84 while RTX Quadro “Turing” stops at 72. Also ray tracingThe units in Ampere see performance improvements in their second generation, where Nvidia’s data indicates doubled performance compared to the first generation.
Overall, the performance on paper looks to be a big step forward compared to what the Quadro RTX 8000 “Turing” offers. In connection with Nvidia unveiling the Geforce RTX 3000 series, the large amount of graphics memory and CUDA cores in the RTX 3090 model gave rise to speculation that the card will replace the company’s Titan products.
The fact that the Quadro RTX series unlocks the full capacity of the GA102 circuit can be seen as another sign that this is Nvidia’s strategy for the future. However, the leaked information does not indicate when Quadro RTX “Ampere” is expected to reach the market or at what price levels it is about.