[11:22 Thu,10.August 2023 by Rudi Schmidts] |
Nvidia has hardly used the current Ada chip generation in its workstation graphics cards so far. Apart from a single top model (the RTX 6000 Ada), the Ada chips have so far only been found in gaming cards, whose high-end model (the RTX4090) strongly resembles the RTX 6000 Ada. From the hardware side, the biggest difference is the memory expansion (48 GB vs. 24 GB), but this is a
But now Nvidia has switched more RTX workstation GPUs from the Ampere generation to current Ada chips, which are also available at a much lower price than the Ada top model. The typical workstation surcharge still remains, though. The Nvidia RTX 5000 Ada with 32 GBytes DDR6 RAM (256 bit bus, 576 GB/s) should cost around 4,000 US dollars, the RTX 4500 Ada with 24 GB RAM (192 bit, 432 GB/s) still 2250 US dollars and the RTX 4000 Ada with 20 GB RAM (160 bit, 360 GB/s) still 1250 US dollars. Even though they have 20 to 33% more memory than their predecessors on paper, Nvidia has cut the memory transfer rate. The RTX 5000, for example, still had 768 GB/s, whereas the Ada variant only has 576 GB. However, the memory speed is often a bottleneck especially for AI models, which makes Nvidia&s design decision a bit incomprehensible. The memory speed is also often quite relevant for GPU effects in Resolve. ;) Nvidia&s new workstation GPUs are now also based on Ada chips. Nvidia also has stronger than usual workstation competition from AMD in this generation. Just recently, AMD introduced  two new entry-level models, the W7600 and W7500. However, the new RTX 5000 Ada competes directly with AMD&s Radeon Pro W7900, but it brings 48 GBytes instead of just 32. And the new RTX 4500 Ada costs as much as AMD&s Radeon Pro W7800 which can offer 32 instead of 24 GB for this. In contrast to AMD, Nvidia also only offers the older Displayport 1.4a interface. AMD has already reached 2.1 here. Nvidia also has little reason for the workstation surcharge apart from the certified drivers. The so-called Studio drivers for many video editing applications have also been running with the gaming cards for a long time.However, NVLink has been a weighty argument for the workstation versions for AI so far, because several GPUs could communicate with each other particularly fast without a detour via the motherboard.This is an important design advantage for large AI models from Nvidia. However, Nvidia has surprisingly omitted the NVLink support from the new Ada workstation cards. Apparently, Nvidia wants to prevent using multiple cards for elaborate AI training with their workstation GPUs. The Nvidia RTX 5000 Ada should be available immediately, the RTX 4000 Ada in September and the RTX 4500 Ada in October.  more infos at bei www.heise.dedeutsche Version dieser Seite: Neue Workstation-Grafikkarten von Nvidia - ausgebremster Speicher und kein NVLink |
 |
|
|