NVIDIA
Coming SoonNVIDIA B300
288GB HBM3e Blackwell Ultra — coming to TATC bare metal
NVIDIA BLACKWELL Series
Specifications (preview)
The B300 anchors NVIDIA's Blackwell Ultra generation — 288GB of HBM3e per card with eighth-generation Tensor Cores. TATC is provisioning B300 nodes for customers who need the maximum HBM capacity on the CUDA stack; detailed specifications are finalising with NVIDIA.
Figures marked TBC / est. are finalising with NVIDIA and will be confirmed at public announcement. Reserve now to lock launch pricing.
Architecture
Blackwell Ultra
VRAM per card
288GB HBM3e
Memory bandwidth
8 TB/s
FP16 / BF16 tensor
TBC (8-GPU)
FP8 tensor
TBC (8-GPU)
FP4 tensor
TBC (8-GPU, FP4)
TDP
~1400W
Interconnect
NVLink 5 (planned)
Why it matters
- High-capacity 288GB HBM3e unified memory per accelerator
- Next-generation NVLink 5 scale-out (specs TBC)
- Full CUDA ecosystem with TensorRT-LLM inference
Built for
- Frontier LLM training (100B+ parameters)
- Ultra-long-context inference
- Large MoE serving at scale
Software stack
CUDA 12.9+
PyTorch
Triton
vLLM
TensorRT-LLM
Also on the roadmap
AMD MI35X Series — AMD Instinct MI355
Lock your place in line
Tell us your target model size and context length — we'll match you to the first node that fits.