1 min read

Rumors Surrounding Nvidia’s Blackwell: RTX 5090 Expected to Deliver 1.7x Performance Boost with Multi-Chiplet Design for Compute GPUs

Information about Nvidia’s upcoming RTX 5000 series graphics cards is starting to surface as the RTX 4000 series becomes a thing of the past. These leaks and rumors provide insights into the RTX 5090’s expected performance boost of 1.7x compared to its predecessor. Additionally, there are indications that Nvidia will adopt a multi-chiplet design for high-performance compute GPUs.

A leaker on the Chiphell forum, known as Panzerlied, has shared alleged specifications for the RTX 5090, including a 50% increase in scale (likely referring to cores), a 52% boost in memory bandwidth, a 78% larger L2 cache, a 15% higher frequency, resulting in the reported 1.7x performance improvement. Extrapolating these details to the RTX 4090 successor suggests it may feature approximately 24,000 CUDA cores, a 2.9 GHz boost clock, and 128MB of L2 cache.

The RTX 5090 is rumored to utilize GDDR7 memory with a boosted speed of 32 Gbps, and while the AD102 GPU successor may include a 512-bit memory bus, it might not apply to the RTX 5090, with potential configurations like 512-bit/24 Gbps or 448-bit/28 Gbps, as noted by VideoCardz.

Panzerlied has a history of accurate leaks, and these recent rumors have received confirmation from the respected hardware leaker Kopite7kimi. If these figures hold true, it raises questions about the pricing of the RTX 5090, as Nvidia faced criticism for its Lovelace pricing, making it unlikely that the RTX 5090 would be cheaper than or even priced similarly to the $1,600 RTX 4090.

Previous RTX 5000-series rumors have also hinted at substantial performance improvements compared to Lovelace, though there is no official word on a release date, but many expect them to arrive next year.

In related news, Kopite7kimi provided insights into Nvidia’s next-gen products, suggesting the use of the Blackwell architecture for both consumer and datacenter GPUs, departing from the current Ada Lovelace/Hopper division. Additionally, Nvidia might introduce a multi-chiplet design for its datacenter-class GPUs, following the path of Intel and AMD. Despite this, Nvidia is expected to maintain a monolithic design for consumer products.

Leave a Reply

%d bloggers like this: