GPUs/NVIDIA B300

B300
coming soon.

Blackwell Ultra with up to 288 GB of HBM3e at 8 TB/s, 1.5× Blackwell's dense FP4 compute and 2× its attention performance, for reasoning-model inference and large-scale training.

Blackwell UltraComing soon
288GB HBM3e, max

Not available to rent yet

Memory288GB max
Bandwidth8,000GB/s
ArchitectureBlackwell Ultra
Form factorSXM
Good for

What the B300
is built for.

Tensor-core precisions: NVFP4, FP6, FP8, BF16, FP16 and TF32.

Best at

  • Reasoning-model inference with long contexts
  • Large MoE models on fewer GPUs
  • FP4 and FP8 training and inference
  • Multi-tenant serving with MIG

Coming soon

The B300 is not on sale yet. Tell us what you would run on it and we will let you know when it launches.

Get notified
Sizing

Will my model fit
on a B300?

Rules of thumb for weights plus a modest KV cache or optimizer state. Longer contexts and bigger batches need more.

WorkloadNeedsOn a B300
Llama 3.1 8B · FP16 inference~18 GBFits
Qwen 32B · FP8 inference~36 GBFits
Llama 70B · FP8 inference~78 GBFits
Llama 70B · FP16 inference~140 GBFits
Llama 8B · full fine-tune (BF16)~70 GBFits
70B · QLoRA fine-tune~48 GBFits
Flux.1 [dev] · BF16~33 GBFits
Llama 405B · FP8 inference~410 GBToo big
Other GPUs

Not the right fit?
Compare the fleet.

These are on sale now, billed by the minute with no contract.

The B300
is coming soon.

Tell us what you would run on it and we'll let you know when it launches.

Get notified
No contractTalk to a human any time

Token Factory is live - free for all users.

GET STARTED FOR FREE
Start Building →
packet.ai/GPUs/NVIDIA B300
Coming soon, join waitlist
NVIDIA Blackwell Ultra288 GB HBM3eSXM

NVIDIA B300 GPU

Blackwell, pushed further.

The NVIDIA B300 is built on Blackwell Ultra: up to 288 GB of HBM3e at 8 TB/s, 1.5× Blackwell's dense FP4 compute and 2× its attention-layer performance, for reasoning-model inference and large-scale training. Coming soon to packet.ai: join the waitlist.

See all pricing
No commitmentEmail when it launches
NVIDIA B300 GPU
288GB
HBM3e memory, max
8TB/s
Memory bandwidth
15PFLOPS
Dense FP4 compute
1.8TB/s
5th-gen NVLink
Specifications

NVIDIA B300 specifications.

SpecificationValueGreat for
GPU architecture
NVIDIA Blackwell Ultra2 dies · 208B transistors
Two dies joined at 10 TB/s, programmed as one CUDA GPU.
GPU memory
Up to 288 GBHBM3e
50% more than Blackwell. Capacity varies by SKU.
Memory bandwidth
8 TB/s
2.4× H100.
FP4 Tensor compute
15 PFLOPSdense NVFP4
1.5× Blackwell, for FP4 inference.
FP8 Tensor compute
5 PFLOPSdense · 10 sparse
FP8 training and inference.
Attention acceleration
2× Blackwell10.7 TeraExponentials/s
Faster softmax for long contexts and reasoning.
NVLink
1.8 TB/s5th generation
GPU-to-GPU across the 8-GPU HGX node.
MIG support
Up to 7 instances
Split a single B300 into isolated GPU instances.
Availability
Coming soonJoin the waitlist
Tell us what you would run on it.
Architecture

Built on NVIDIA Blackwell Ultra, with 50% more memory.

Blackwell Ultra keeps Blackwell's two-die, 208-billion-transistor design and raises HBM3e to up to 288 GB, with 1.5× the dense FP4 compute and 2× the attention-layer performance.

Two dies, one GPU

Two reticle-sized dies joined at 10 TB/s by NV-HBI, 208 billion transistors on TSMC 4NP, programmed as a single CUDA GPU.

Up to 288 GB HBM3e at 8 TB/s

50% more memory than Blackwell and 3.6× H100, so larger models and KV caches stay on one GPU.

2× attention, 1.5× FP4

Blackwell Ultra Tensor Cores double attention-layer throughput and raise dense NVFP4 compute to 15 PFLOPS.

NVLink 5 at 1.8 TB/s

Fifth-generation NVLink connects the eight GPUs of an HGX B300 node, with 14.4 TB/s of total NVLink bandwidth.

Use Cases

What the B300 is built for.

Reasoning-model inference

2× Blackwell's attention-layer performance speeds up the softmax that dominates long reasoning traces and long contexts.

  • Long context windows
  • 2× attention vs Blackwell
  • NVFP4 inference

Large MoE inference

Up to 288 GB per GPU puts very large mixture-of-experts models, DeepSeek-R1 class, on fewer GPUs.

  • 50% more memory than B200
  • Fewer GPUs per model
  • 8 TB/s bandwidth

Large-scale training

FP8 and NVFP4 Tensor Cores with 1.8 TB/s NVLink between the eight GPUs of an HGX node.

  • 5 PFLOPS dense FP8
  • 5th-gen NVLink
  • 8-GPU HGX B300 node

Multi-tenant serving

Split one B300 into up to seven isolated MIG instances to serve many smaller models.

  • Up to 7 MIG instances
  • 34 GB per instance
  • Isolated performance
Pricing

B300 pricing.

B300 is coming soon to packet.ai. Join the waitlist to be notified.

FAQ

NVIDIA B300, answered.

What is the NVIDIA B300?

The B300 is NVIDIA's Blackwell Ultra GPU: two dies with 208B transistors, up to 288 GB HBM3e at 8 TB/s, and 2× the attention-layer performance of Blackwell.

How is B300 different from B200?

Same 8 TB/s bandwidth and 1.8 TB/s NVLink. B300 adds memory (up to 288 GB vs 192 GB), 1.5× the dense FP4 compute and 2× the attention performance. See B200 →

Can I rent a B300 on packet.ai?

Not yet. The B300 is coming soon: join the waitlist and we'll email you when capacity opens.

Does B300 support MIG?

Yes, up to 7 MIG instances of 34 GB each, or 4 × 70 GB, or 2 × 140 GB.

Is the B300 good for FP64 HPC?

NVIDIA rates HGX B300 at 10 TFLOPS FP64 against 296 TFLOPS for HGX B200, so FP64-heavy simulation runs better on B200. See B200 →

Coming soon

B300: Coming soon. Get notified.

Join the waitlist and be first to know when B300 capacity opens on packet.ai.

Talk to a human

No commitment · we'll notify you by email

NVIDIA B300