NVIDIA RTX 5090 Rumors: Blackwell Architecture 70% faster?

Recent disclosures suggest developments within Nvidia's upcoming video card generation. A Chiphell user, Panzerlied, has provided insights into the flagship based on the Blackwell architecture, which is reportedly superior to the RTX 4090. Concurrently, Kopite7kimi unveiled new details about the Blackwell series, NVIDIA's upcoming GPU lineup. 

Within this framework, GB1XX signifies high-performance computing (HPC), whereas GB2XX represents gaming GPUs. Instead of furnishing exact numbers, Panzerlied has depicted percentage improvements for various elements of the Blackwell lineup. Drawing from these, a comparison between NVIDIA RTX 5090 and RTX 4090 indicates:

Panzerlied stated that these figures are relative to the RTX 4090 specifications, not AD102. Using the RTX 4090's 21 Gbps memory as a baseline, an anticipated 32 Gbps (52.4% growth) might imply a move towards GDDR7 technology for the RTX 50 series. Speculations also hint at a successor to the AD102 featuring a 512-bit memory bus, though it's uncertain if this applies to the RTX 5090 or might be specific to a future TITAN/RTX workstation or a 5090Ti model.

Should these assertions hold, a 15% frequency growth could correspond to a 2.9 GHz boost clock, possibly achieving 3.0 GHz or more for some workloads. A cache increase of 78% might indicate the GB202 GPU housing 128MB of L2 cache.

The mentioned 1.7x improvement is still ambiguous in terms of specific workloads, but this likely aggregates from the described core growth, memory bandwidth enhancement, cache growth, and clock speed hike. An approximate specification comparison is as follows:

GeForce GPURTX 4090RTX 5090
Streaming Multiprocessors128192
Cuda cores16,38424,576
Clock speed (boost)2.5 GHz2.9 GHz
L2 cache72 MB128 MB
Memory speed21Gbps24 Gbps (G6X) / 32 Gbps (G7)
Memory bus384 bit512-bit (G6X) / 384-bit (G7)
Bandwidth1,008 GB/s1,536 GB/s

To achieve the speculated 1,536 GB/s bandwidth, either GDDR6X memory at 24 Gbps with a 512-bit memory bus or GDDR7 at 32 Gbps on a 384-bit bus might be employed. VRAM capacity is yet to be confirmed but might range between 24 or 32 GB. The power specifications are also undetermined, with the RTX 4090 using 24 GB of GDDR6X at 450 watts as a reference. It's essential to treat these speculations with caution, given their origin and the 2025 expected product line release. However, some information aligns with disclosures from the leaker, Kopite7kimi.

Sources: Chiphell , Videocardz

Printed from: https://www.guru3d.com/story/nvidia-rtx-5090-rumors-blackwell-architecture-faster/