"NVIDIA's RTX 5090 Blackwell Die-Shot Leak: Inside the Monster Powering Tomorrow's
Gaming Revolution
A very blurry image has created waves in the tech world: the first die-shot of NVIDIA's GB202, the
silicon heart of the forthcoming GeForce RTX 5090, has leaked. Shared by hardware analyst
*Kurnal* and ASUS China's Tony Yu, this annotated glimpse of NVIDIA's "Blackwell" architecture
shows design so ambitious that it redefines what is possible for gaming, AI, and creative workloads.
Let's dissect the future.
The Blackwell Architecture: A Blueprint for Dominance
Named after mathematician David Blackwell, NVIDIA's Blackwell isn't just an evolution—it's a
revolution. The GB202 die, which is smaller than NVIDIA's data center-focused GH100 at 814
mm², packs a jaw-dropping 24,576 CUDA cores in its full configuration. But here's the kicker: the
RTX 5090 will "only" enable 21,760 cores. Why? Because the *full* GB202 die is reserved for
workstation GPUs, leaving gamers with a slightly trimmed—but still monstrous—version.
-Inside the Beast: A Tour of the GB202 Die:- The die-shot shows an almost obsessively laid-out
structure: Central Cache Hub : 128 MB L2 sits at the center (96 MB enabled for the RTX 5090). It
is flanked by memory controllers. That's a big cache, functioning like a traffic controller,
significantly reducing latency, and feeding hungry cores in the GPU.
- Memory Muscle:- Eight 64-bit memory controllers drive a 512-bit GDDR7 interface, implying
over 1.5 terabytes per second bandwidth—something like nearly two times the rate of the GDDR6X
on the RTX 4090. Therefore, the RTX 5090 will trounce 4K and 8K gaming, AI, and content-
creation workloads.
- GPCs Galore:- Twelve Graphics Processing Clusters (GPCs) surround the cache, each holding
eight Texture Processing Clusters (TPCs) and 16 Streaming Multiprocessors (SMs). The modularity
makes it possible for NVIDIA to scale performance for a specific market.
------------------------------------------------------------------------------------------------------------------------
SM Design: Where Magic Happens
Every SM (Streaming Multiprocessor) is an engineering marvel:
- Quad-Slice Power Four slices of a die, this design contains 128 KB L1 cache as well as four
TMUs, where every slice is comprised of its resources-dedicated files, register warps schedulers,
and AI ray tracing specific units.
- Next-Gen Cores: 5th-gen Tensor cores (768 total) and 4th-gen RT cores (192 total) deliver *50-
100%* AI and ray tracing performance gains over Ada Lovelace. Imagine *Cyberpunk 2077* with
path tracing at 4K/120 FPS—this hardware could make it a reality.
- Efficiency Tweaks: The L0 instruction cache and enhanced schedulers reduce bottlenecks and
ensure each core can stay fed on data.
------------------------------------------------------------------------------------------------------------------------
Media Mastery and Display Tech
Running through the die is a vertical strip carrying NVENC (encoding) and NVDEC (decoding)
units. The RTX 5090 powers three NVENC encoders and two NVDEC decoders; this means that it
is able to support:
- 8K/120 AV1 encoding for streamers and creators.
- Simultaneous multi-stream decoding for AI video workflows.
At the base of the die, the PCIe 5.0 x16 interface and display controllers hint at support for HDMI
2.1a and Display Port 2.1, ready to drive 8K monitors or VR headsets with zero compromise.
------------------------------------------------------------------------------------------------------------------------
Why Workstations Get the Full Die
The full GB202 configuration—24,576 CUDA cores, 128 MB L2 cache—is reserved for NVIDIA's
RTX workstation GPUs (think RTX 6000 Ada's successor). This split means professionals receive
the unbeaten compute density for tasks like 3D rendering and AI training, but the gamers get a
generational improvement.
Blackwell vs. Ada Lovelace: What's Changed?
- CUDA Cores: RTX 5090's 21,760 cores v. RTX 4090's 16,384 - *33% increase*.
- Memory: GDDR7's 32 Gbps+ speeds vs. GDDR6X's 21 Gbps - *50% more bandwidth*.
- RT/Tensor Cores: 4th-gen RT cores improve ray traversal, while 5th-gen Tensor cores could
unlock DLSS 4.0(DLSS is suite of neural rendering technologies that uses AI to boost FPS, reduce
latency, improve image quality) with even smarter upscaling.
The Bottom Line: A New Era for Gamers and Creators
It is not just about higher numbers, but smarter design with NVIDIA's Blackwell architecture. The
die layout of the RTX 5090 shows a GPU that would dominate both rasterization and ray tracing as
well as AI, and with GDDR7, a huge L2 cache, and re-engineered SMs, it may be the first GPU to
offer true 8K gaming without compromise.
For creators, the fully enabled GB202 in workstation cards will redefine rendering and simulation.
But gamers aren’t left behind: the RTX 5090 is poised to be the ultimate 4K/1440p beast, with AI-
driven features we’ve yet to imagine.
*One question remains*: How fast will it sell out?
-Thanks to Kurnal and ASUS China's Tony Yu for the leak.
