A range of cards with a maximum of 48 GB of video memory... But what game could possibly need that much capacity?
No, of course not, no video game - not even the heaviest of the moment - needs that much video memory, not even when the graphics settings are set to maximum 8K detail. By the way, this announcement from NVIDIA is not about new GeForce, nor have you missed the presentation of a new architecture from the firm with the chameleon. These RTX 4000, RTX 4500 and RTX 5000 are actually a range of cards designed by NVIDIA for workstations. Notice how the American company is careful not to use the term GeForce.
© NVIDIA
These three new boards follow on from the announcement a few weeks ago when NVIDIA launched the RTX 6000, an absolutely monstrous board featuring an AD102 GPU - Ada Lovelace architecture - with 18,176 CUDA cores, 568 Tensor cores and, consequently, 48 GB of G6 ECC memory at 20 Gbps on a 384-bit interface bus for an impressive bandwidth of 960 GB/s. Such a monster is obviously worth its weight in peanuts, with a price tag of around $6,800. Obviously, such a monster couldn't be aimed at the general public, but even in the professional world of workstations, it's not easy to come up with that kind of money.
Videocardz
To ease the pain, NVIDIA has come up with three new boards, all designed for workstations. The RTX 5000 lowers the cursor a little, but its AD102 GPU retains 12,800 CUDA cores and 400 Tensor cores, while the memory subsystem consists of 32 GB of G6 ECC at 18 Gbps on a 256-bit interface bus for a bandwidth of 578 GB/s. Admittedly, this is still a pretty baby, and that's reflected in the price: NVIDIA is still asking $4,000 for it. The RTX 4500 takes things down another notch, this time with an AD104 GPU, 7,680 CUDA cores and 240 Tensor cores. The memory subsystem makes do with 24 GB in G6 ECC at 18 Gbps and a 192-bit interface bus for a bandwidth of 432 GB/s. The price is $2,250.
© NVIDIA
But if you really want to get down to earth on price ($1,250), you'll have to turn to the RTX 4000. Here, the AD104 GPU has just 6,144 CUDA cores and 192 Tensor cores, with "only" 20 GB of G6 ECC at 18 Gbps on a 160-bit bus, for a bandwidth of 360 GB/s. For the same price, NVIDIA offers an "SFF" version of this card. More compact, it fits into slimmer cases, but has to make a few concessions: lower operating frequencies, slightly slower memory and reduced bandwidth (320 GB/s). On the other hand, this allows for a lower TDP (70 Watts), so it can be used in more cramped spaces.
