Blackwell architecture — in production

Thousands of cores.
One parallel mind.

A GPU isn't a faster CPU — it's a different shape of computer, built from thousands of small cores working the same problem at once. That shape is why TECHNION GPUs run graphics, simulation, and AI training at a scale a handful of large cores never could.

208Btransistors on a Blackwell GPU
2reticle-limit dies fused into one GPU
10 TB/sdie-to-die link inside the package
5thgeneration of NVLink
GPUs

Same core design, scaled to the room it's running in.

A gaming GPU, a workstation GPU, and a data-center GPU all share the same streaming-multiprocessor building block — what changes is how many are packed in, and what they're wired to next.

RT
GeForce RTX 5090 — Gaming

For real-time ray tracing

Streaming multiprocessors packed with CUDA, RT, and Tensor cores side by side, so lighting and AI upscaling run in the same clock cycle as the rest of the frame.

CUDA cores21,760
RT cores4th gen
Memory32GB GDDR7
ECC
RTX PRO 6000 — Workstation

For certified professional pipelines

The same core array, validated against creative and engineering applications, with ECC memory to catch bit errors long renders can't afford to hide.

CUDA cores24,064
Memory96GB ECC GDDR7
Driver trackStudio / Enterprise
NV-HBI 10TB/s
B200 — Data center

For AI training at scale

Two reticle-limit dies fused into one GPU over a 10 TB/s link, then wired to dozens more GPUs over NVLink so a training job can spread across a whole room.

Transistors208 billion
NVLink1.8 TB/s per GPU
TDP class1000W
Architecture roadmap

A new architecture, almost every year.

Each generation changes what the core array is built from and how the GPUs talk to each other. This is the sequence, in the order it shipped.

2018

Turing

First RT cores for real-time ray tracing.

2020

Ampere

Second-gen RT cores, sparsity-aware Tensor cores.

2022

Hopper

Data-center-only design; introduced the Transformer Engine.

2024

Blackwell

First dual-die GPU, fused over a 10 TB/s link — current generation.

2026

Rubin

Next architecture, announced for data-center systems.

72GPUs linked in one NVLink domain (NVL72)
1.8 TB/sNVLink bandwidth per GPU, 5th generation
4thgeneration Tensor Cores, per SM
2016→CUDA compute platform, continuously extended

Let's build your next platform on TECHNION.

Talk to our team about GPUs, DGX systems, or CUDA developer access.