NVIDIA B200 (HGX B200)
NVIDIA: specifications, comparisons and alternatives.
The NVIDIA B200 is a Blackwell GPU for AI training and inference, supplied in the HGX B200 platform with 8 SXM GPUs. According to the data, each GPU has 180 GB of HBM3e, up to 8 TB/s of memory bandwidth and a configurable power of up to 1,000 W. NVLink 5 connects the GPUs at 1.8 TB/s.
NVIDIA
- 180 GB
- Memory
- HBM3e
- Memory type
- 8 TB/s
- Memory bandwidth
- 1.000 W
- Power (board power or TDP)
You can see the current price and availability with the partner as soon as a link has been set up.
How are the points calculated?
Clouwino points are an arithmetic summary of the specifications the manufacturer states, on a scale of 0 to 100. It is not a benchmark and not a judgement on quality or suitability: a product with fewer points may suit your situation better. Products are compared with other products of the same type. For each property we count how far a product sits between the lowest and highest value in the group:
- FP32 compute (for: consumer), carries heavy weight
- Memory (for: consumer), carries heavy weight
- Memory bandwidth (for: consumer), carries heavy weight
- Compute units (for: consumer), carries light weight
- Boost clock (for: consumer), carries light weight
- Memory (for: professional and data centre), carries medium weight
- Memory bandwidth (for: professional and data centre), carries medium weight
- FP16/BF16 (for: professional and data centre), carries heavy weight
- FP8 (for: professional and data centre), carries medium weight
- FP64 (for: professional and data centre), carries medium weight
If too much data is missing, we do not show points. Points are relative: when a new product is added, they can shift. Read more at how we compare.
Specifications of NVIDIA B200 (HGX B200)
All specifications the manufacturer lists, per item. Want to see another product next to it? Pick one below.
| Specificatie | NVIDIA B200 (HGX B200) |
|---|---|
| Overview | |
| Clouwino points | Not stated by the manufacturer |
| General | |
| Architecture | Blackwell |
| Manufacturing process | TSMC 4NP |
| Form | SXM (HGX B200, 8x Blackwell SXM) |
| Launch | 2024-03 |
| Compute | |
| FP32 compute | Not stated by the manufacturer |
| FP64 | Not stated by the manufacturer |
| FP16/BF16 | Not stated by the manufacturer |
| FP8 | Not stated by the manufacturer |
| FP4 | Not stated by the manufacturer |
| Performance notes | On its product pages NVIDIA only publishes figures for the 8-GPU HGX/DGX B200 system, not per-GPU figures; conversion is not permitted, so all compute performance fields are null. System values HGX B200 (8 GPUs): FP4 144 PFLOPS sparse / 72 PFLOPS dense; FP8 72 PFLOPS sparse (dense = half); FP16/BF16 36 PFLOPS sparse; TF32 18 PFLOPS sparse; INT8 72 POPS sparse; FP32 600 TFLOPS; FP64/FP64 Tensor Core 296 TFLOPS. |
| Memory | |
| Memory | 180 GB |
| Memory type | HBM3e |
| Memory bandwidth | 8 TB/s |
| Power and connections | |
| Power (board power or TDP) | 1.000 W |
| GPU interconnect | NVLink 5: 1.8 TB/s GPU-to-GPU |
| Multiple instances | Not stated by the manufacturer |
Note: Memory (180 GB), bandwidth (up to 8 TB/s) and 'configurable up to 1 kW per GPU' come from NVIDIA reference architecture documentation (docs.nvidia.com), not from a datasheet table; the DGX B200 page states 1,440 GB and 64 TB/s for 8 GPUs. TDP is a maximum (configurable). Process node: NVIDIA states a custom TSMC 4NP process for Blackwell GPUs (press release 18 March 2024). Launch date = announcement 18 March 2024 (GTC).
Alternatives to NVIDIA B200 (HGX B200)
Products with similar specifications from the same overview. This is not a ranking and not buying advice.
- NVIDIA H100 SXM80 GB · HBM3 · 3,35 TB/sCompare with NVIDIA B200 (HGX B200) →
- NVIDIA H200 SXM141 GB · HBM3e · 4,8 TB/sCompare with NVIDIA B200 (HGX B200) →
- NVIDIA Rubin (per GPU)288 GB · HBM4 · 22 TB/sCompare with NVIDIA B200 (HGX B200) →
- NVIDIA L40S48 GB · GDDR6 (ECC) · 0,86 TB/sCompare with NVIDIA B200 (HGX B200) →
- NVIDIA RTX PRO 6000 Blackwell Server Edition96 GB · GDDR7 (ECC) · 1,6 TB/sCompare with NVIDIA B200 (HGX B200) →
- AMD Instinct MI300X192 GB · HBM3 · 5,3 TB/sCompare with NVIDIA B200 (HGX B200) →
B200 and the HGX platform
The B200 is the Blackwell generation for data centres, made on a custom TSMC 4NP process. NVIDIA announced Blackwell on 18 March 2024. You mainly see the B200 as part of the HGX B200 board with eight Blackwell SXM GPUs, and as a DGX B200 system.
That matters for reading the figures: NVIDIA publishes the compute performance for the complete system of 8 GPUs, not per GPU. That is why we do not give per-GPU values for FP64 through FP4 here.
It also means you need to read B200 information from providers carefully: is a figure given for one GPU or for the whole board?
System figures and what they mean
Per GPU, the reference documentation gives 180 GB of memory and up to 8 TB/s of bandwidth. The DGX B200 page gives 1,440 GB and 64 TB/s for the 8 GPUs combined. Want to know what this form of memory involves? Then read about HBM.
The compute figures apply to the HGX B200 system with 8 GPUs. For FP4, NVIDIA lists 144 PFLOPS with sparsity and 72 PFLOPS dense. FP8 is 72 PFLOPS with sparsity, FP16/BF16 36 PFLOPS with sparsity. FP32 is 600 TFLOPS and FP64 296 TFLOPS for the system. Do not divide such numbers by eight yourself to get GPU values; the manufacturer does not do that either.
With every figure, also look at the label: sparse and dense are two different numbers for the same precision, and you can only compare like-for-like labels.
What to look out for with an HGX B200
With a power draw that can rise to 1,000 W per GPU, cooling is the first point of attention. A system with eight of these GPUs needs a data centre that is prepared for it.
- NVLink 5 at 1.8 TB/s between GPUs is intended for exchanging data within the system, see NVLink, PCIe and GPU interconnects.
- Check whether a provider supplies a complete 8-GPU system or part of one.
- Do not compare system figures with per-GPU figures of other models.
Finally, for the complete system the DGX page gives a total of memory and bandwidth, which is a different unit from the value per GPU.
You can see the current price and availability at the partner via the button on the page.
Frequently asked questions
Why are there no per-GPU figures for the B200?
On its product pages, NVIDIA only publishes compute performance for the 8-GPU HGX or DGX B200 system. Because converting to a single GPU is not done by the manufacturer, we leave those fields empty.
How much power does a B200 use?
According to the NVIDIA reference documentation, the power is configurable up to 1,000 W per GPU. That is a maximum, so actual consumption depends on the setting and the load.
What FP4 value does NVIDIA give for the HGX B200 system?
For the system with 8 GPUs, NVIDIA lists 144 PFLOPS FP4 with sparsity and 72 PFLOPS dense. So it is a system value and not a figure per GPU.
Last updated on
