NVIDIA B300

B300, or Blackwell Ultra, is NVIDIA’s follow-on to its B200 GPU. It has similar power consumption to B200 but is new silicon.

Each GPU will have¹

• 2 reticle-sized GPUs
• 288 GB HBM3e
  • 279 GB usable for GB300²
  • 270 GB usable for B300²
• Essentially no FP64 performance
• Upgrade to PCIe Gen6
• Upgrade to 1400 W (from 1200W) for GB300²
• Upgrade to 12-high HBM3e stacks (from 8-high)

Specifically, the “full implementation” (not taking into account deactivated components) has:³

• 8 GPCs
  • 160 SMs (20 per GPC)
  • 20,480 CUDA cores (128 per SM)
  • 640 tensor cores (4 per SM)
  • ? GHz (low precision), ? GHz (high precision)
  • 2:4 structured sparsity
• 288 GB HBM3e (8 stacks, 12-high)
  • 8 TB/s (max)
• 2x 900 GB/s NVLink 5 (D2D)
• 2x 256 GB/s PCIe Gen6 (H2D)
• 1400 W maximum

It is clearly a part optimized for inferencing.

Performance

GB300 NVL72

The GB300 variant of B300 runs at up to 1,400 W.²

Data Type	VFMA	Matrix	Sparse
FP64	1.3	1.3
FP32	80
TF32		1250	2500
FP16		2500	5000
BF16		2500	5000
FP8		5000	10000
FP6		5000	10000
FP4		15000	20000
INT8		165	330

HGX B300

The HGX variant of B300 has similar performance to B200:²

Data Type	VFMA	Matrix	Sparse
FP64	1.2	1.2
FP32	75
TF32		1100	2200
FP16		2250	4500
BF16		2250	4500
FP8		4500	9000
FP6		4500	9000
FP4		18000	14000
INT8		154	307

The biggest differences are:

• Dense FP4 performance is up (14 PF on B300 vs. 9.0 PF on B200)
• FP64 performance is dramatically lower (1.2 TF on B300 vs. 37 TF on B200)

GB300 NVL72

GB300 NVL72 is advertised to have:²

• 1.080 EF dense FP4 for inference
• 360 PF FP8 for training
• 20 TB HBM at 576 TB/s
• 37 TB “Fast Memory”
• 18 NVLink Switches at 130 TB/s
• 130 trillion transistors¹
• 2,592 Grace CPU cores
• 72 ConnectX-8 NICs
• 18 BlueField DPUs

GB300 NVL72 maintains compatibility with the GB200 NVL72 Oberon rack but has a few notable differences:

• It is a single carrier board with two Grace CPUs and four Blackwell GPUs instead of two identical boards with 1C:2G each.
• Mellanox ConnectX-8 NICs are included on the board now
• CPUs, GPUs, and NICs are all socketed and individually replaceable now
• The server is designed to be 100% liquid-cooled

Here is a photo of the GB300 NVL72 superchip board I took at GTC25:

From top to bottom are the NVLink connectors (orange), four B300 GPUs, two Grace CPUs, and four ConnectX-8 NICs (copper).

Here is HPE’s implementation of the GB300 NVL72 server sled:

A liquid manifold replaces the row of fans that was in the GB200 NVL72 server platform.

It uses the same rack power and liquid infrastructure, fits within the same rack power density, and maintains 72 GPUs per NVLink domain.⁴ It debuted at GTC25.

The photo below shows the DGX GB300 (left) and DGX GB200 (right). They are indistinguishable.

A single GB300 NVL72 rack is rumored to cost between $3.7 and $4 million.⁵

Footnotes

1. Jensen Huang’s Keynote at GTC25. ↩ ↩²

2. NVIDIA Blackwell Ultra Datasheet ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

3. https://developer.nvidia.com/blog/inside-nvidia-blackwell-ultra-the-chip-powering-the-ai-factory-era/ ↩

4. Nvidia’s Jensen Huang, Ian Buck, and Charlie Boyle on the future of data center rack density ↩

5. Apple becomes NVIDIA customer, orders estimated $1 billion of new GB300 NVL72 AI servers ↩