Nvidia's Massive RTX 5090 GPU: A Closer Look at the GB202 Die Shot
Discover the grandeur of Nvidia's RTX 5090 GPU through an extensive exploration of its GB202 die shot, revealing intricate architectural details.
A Detailed Look at Nvidia’s RTX 5090: The GB202 Die Shot
Count the transistors with me. One, two, skip a few, ninety-two point two billion.
With the RTX 50 series launch at CES earlier this month and subsequent reviews, especially in-depth ones detailing specifications, it’s clear that the Blackwell chip powering the GeForce RTX 5090 is impressively vast. After a new die shot was revealed, we can now observe all its shaders and cache.
Creating a comprehensive die shot isn’t straightforward. It takes multiple attempts, often resulting in cracked chips, to achieve clarity. And when the chip in question is Nvidia’s GB202, costing over $2,000, the stakes are even higher.
Collaboration with Asus
Enter Tony Yu, the General Manager of Asus China, who shared a high-resolution image of the GB202 die that another user, Kurnal, managed to capture. This image highlights all essential components, labeled for clarity.
While the die image doesn’t disclose major surprises, as Nvidia uses a similar design layout for many years, it does illustrate the interesting engineering choices necessary to fit everything into its physical dimensions.
For example, a comparison shows that all the logic blocks for NVENC video encoders and decoders have shifted from the bottom to the center of the chip. This redesign was essential as the GB202 keeps three encoders and two decoders, compared to two and one for the previous architecture.
Design Insights
Another visible aspect is the substantial amount of L2 cache at the die’s center. Nvidia’s simplistic cache design contrasts with AMD’s multi-level hierarchy, comprising major L1 caches per Streaming Multiprocessor (SM) and a significant L2 cache.
Despite being the largest chip Nvidia has included in a gaming graphics card by transistor and shader count, the GB202 measures 750 mm², making it 23% larger than its predecessor, the AD102.
Conclusion
Whether future GPUs will reach this size remains uncertain, but if this marks the end of extraordinarily large chips, then enthusiasts at least have this stunning die shot as a keepsake.
