- A transport manifest has detailed what appears to be like like knowledgeable workstation card
- It might presumably be the successor to Nvidia’s RTX 6000 Ada, the costliest graphics card on the earth
- Primarily based on the RTX5090, it’s anticipated to have a whopping 96GB, twice that of it predecessor
The GeForce RTX 5090, the most recent flagship graphics card for avid gamers and creatives in Nvidia‘s GeForce 50 sequence, was unveiled at CES 2025 and has simply gone on sale – buts hortly earlier than it did, rumors started to swirl of an RTX 5090 Ti mannequin that includes a completely enabled GB202-200-A1 GPU and twin 12V-2×6 energy connectors, theoretically permitting for as much as 1,200 watts of energy.
This hypothesis started following the looks of a prototype picture on the Chinese language trade discussion board Chiphell – reporting on the picture, ComputerBase stated, “With 24,576 shaders, the GB202-200-A1 GPU is alleged to supply 192 lively streaming multiprocessors, which had been beforehand rumored to be the complete growth of the GB202 chip. The reminiscence is alleged to proceed to supply 32GB capability, however with 32Gbps as an alternative of 28Gbps, it’s going to exceed the 2TB/s mark.”
Shortly after the engineering card surfaced on-line, ComputerBase alsospotted transport paperwork on NBD Information itemizing a graphics card with 96GB of GDDR7 reminiscence, marked as “for testing.” It’s a affordable assumption that this unidentified mannequin is definitely knowledgeable workstation card, doubtlessly – let’s say in all probability – the RTX 6000 Blackwell.
Helpful for AI purposes
The GeForce RTX 5090 options 32GB of GDDR7, utilizing sixteen 2GB modules linked by means of a 512-bit reminiscence interface. 48GB can be potential if sixteen 3GB chips had been used as an alternative of 2GB chips.
If two of those 3GB chips had been linked to every 32-bit controller, inserting 16 chips on each the back and front of the graphics card in a “clamshell” configuration, the 96GB talked about within the paperwork – which is twice as a lot because the RTX 6000 Ada, the costliest graphics card on the earth – would develop into a actuality.
The transport data point out these GPUs use a 512-bit reminiscence bus, reinforcing this principle. The interior PCB designation PG153, seen within the paperwork, aligns with recognized Nvidia Blackwell designs and has not but appeared in any current client graphics playing cards.
Nvidia is anticipated to introduce the RTX Blackwell sequence for workstations at its annual GPU Expertise Convention (GTC 2025), so we must always know extra about them come March 2025. And sure, in the event you’re pondering 96GB of GDDR7 reminiscence is overkill for gaming or artistic functions I’d agree with you. It’s a good quantity for AI duties although, so we are able to count on to see Nvidia announce an AI model of the RTX 6000 Blackwell when it lastly takes the wraps off its next-gen product.