The Engines
A modern GPU is more than its raw FLOPS. Blackwell bundles several dedicated engines, each built to remove a specific bottleneck in reasoning AI — from keeping FP4 accurate, to securing private data, to decompressing databases and predicting hardware faults.
Transformer Engine
Makes ultra-low FP4 precision accurate enough to use.
Attention Acceleration
2× faster attention for long reasoning sequences.
Secure AI
Extends confidential computing from CPU to GPU.
Decompression Engine
Accelerates databases and data analytics.
RAS Engine
Predicts faults to keep huge clusters healthy.
5th-gen NVLink →
The interconnect — its own module, next.
#1 · The 2nd-gen Transformer Engine
This is the system that turns FP4 from a risky idea into a usable speedup. The 2nd-generation Transformer Engine combines custom Blackwell Tensor Cores with NVIDIA’s Dynamo, TensorRT-LLM and NeMo software. Using dynamic range management and fine-grained micro-tensor scaling, it keeps accuracy high while running in 4-bit. The effect on Blackwell’s FP4 path:
- 2× the FP4 compute performance,
- 2× the parameter bandwidth to HBM memory, and
- 2× the model size that fits per GPU.
For Mixture-of-Experts models, Dynamo and TensorRT-LLM add 4-bit quantization, expert-parallelism mapping, and disaggregation — democratizing real-time MoE inference with less hardware, energy and cost. For training, NeMo and Megatron-Core bring new expert-parallelism techniques.
#2 · Attention layer acceleration
Reasoning means working over long input sequences, and the “attention” step is where transformers spend much of that effort. The Blackwell Ultra GPU adds new instructions that give a 2× speedup on attention compute versus a standard Blackwell GPU — cutting latency so reasoning models respond faster and cost less to run.
#3 · Secure AI & confidential computing
Enterprises often can’t put private or regulated data on a model without strong guarantees. Blackwell extends the Trusted Execution Environment — an encrypted, hardware-protected zone — from the CPU all the way to the GPU. It’s the industry’s first TEE-I/O capable GPU, with inline protection over NVLink.
Blackwell confidential computing delivers nearly identical throughput to running unencrypted — so even the largest models can be protected without a performance penalty.
#4 · The Decompression Engine
Data analytics has traditionally leaned on CPUs and been slow. Blackwell adds a dedicated Decompression Engine that unpacks compressed data at up to 800 GB/s, supporting LZ4, Snappy and Deflate. Paired with 8 TB/s of HBM3e and the Grace CPU’s NVLink-C2C link, it accelerates the whole database query pipeline.
#5 · The RAS Engine
When a single job spans thousands of GPUs, one failing part can stall everything. The dedicated RAS Engine (Reliability, Availability, Serviceability) uses AI-powered predictive management to continuously monitor thousands of data points across hardware and software, catching potential faults early to minimize downtime.