B200 vs H200 comes down to memory, number formats and power. An eight-GPU HGX B200 holds 1.4 TB of GPU memory against 1.1 TB on an HGX H200 [1, 8]. Each B200 reads its memory at 7.7 TB/s against 4.8 TB/s on an H200, and its tensor cores run FP4, which Hopper GPUs lack [2, 5, 6]. NVIDIA's own B200 build, DGX B200, draws up to 14.3 kW in 10 rack units, against 10.2 kW in 8U for a standard DGX H200 [16, 17, 18].
HGX B300 adds memory and dense FP4 throughput and keeps the B200's memory bandwidth, FP8 rate and NVLink speed [1, 3, 4]. It gives up FP64 and INT8 throughput: NVIDIA lists 10 TFLOPS of FP64 for an HGX B300 against 296 TFLOPS for an HGX B200 [1].
Specifications below come from NVIDIA unless another maker is named. Every FP4, FP8 and INT8 rate is marked dense or sparse, and MLPerf® results test the datasheet gaps.
HGX H200, B200 and B300 specs
NVIDIA's HGX page now covers only Blackwell and Rubin boards [1]. H200 and H100 figures come from their product pages, which mark H200 figures preliminary [6, 7].
| Per GPU | H100 SXM | H200 SXM | B200 in HGX B200 | B300 in HGX B300 |
|---|---|---|---|---|
| GPU memory | 80 GB | 141 GB | 180 GB | 270 GB or 288 GB |
| Memory bandwidth | 3.35 TB/s | 4.8 TB/s | 7.7 TB/s | 7.7 TB/s |
| FP4 tensor, dense | none | none | 9 PFLOPS | 14 PFLOPS |
| FP8 tensor, dense | about 2 PFLOPS | about 2 PFLOPS | 4.5 PFLOPS | 4.5 PFLOPS |
| INT8 tensor, dense | about 2 POPS | about 2 POPS | 4.5 POPS | 0.15 POPS |
| FP64 tensor core | 67 TFLOPS | 67 TFLOPS | 37 TFLOPS | 1.2 TFLOPS |
| NVLink | 900 GB/s | 900 GB/s | 1.8 TB/s | 1.8 TB/s |
| Maximum power | 700 W | 700 W | 1,000 W | 1,100 W |
| Memory, eight GPUs | 640 GB | 1,128 GB | 1.4 TB | 2.1 TB or 2.3 TB |
Sources: [1, 2, 3, 4, 5, 6, 7, 14, 18]. H100 and H200 dense rates are half the sparse rates NVIDIA lists (derived).
NVIDIA lists TF32 and lower-precision tensor rates with sparsity, which assumes weights pruned so that two of every four values are zero [1, 6, 10]. For FP8, INT8, FP16, BF16 and TF32, dense is half the sparse figure, and it is the peak for a model without that pattern [1, 10]. NVIDIA lists FP4 both ways, and HGX B300's dense FP4 is three quarters of its sparse rate, 108 against 144 PFLOPS [1].
B200 vs H200: what changes
Per GPU, B200 in an HGX B200 has 180 GB of memory against 141 GB, and about 1.6 times the bandwidth (derived) [2, 6]. Dense FP8 rises from about 2 to 4.5 PFLOPS per GPU, about 2.3 times (derived) [2, 5]. Fifth-generation NVLink doubles GPU-to-GPU bandwidth to 1.8 TB/s [4].
FP4 is new. NVIDIA says its NVFP4 format cuts memory footprint about 1.8 times against FP8, with 1% or less accuracy loss on the reasoning model it tested [9]. How much model and context a given memory holds is worked through in how much GPU memory LLM inference needs.
Two figures move the other way. B200's FP64 rate, 37 TFLOPS, sits below H200's 67 TFLOPS on tensor cores, and power rises from 700 W to 1,000 W per GPU [2, 6].
B300 vs B200: what changes and what stays
HGX B300 differs from HGX B200 in five ways:
- GPU memory rises from 180 GB to 270 GB per GPU [2, 3].
- Dense FP4 rises from 72 to 108 PFLOPS per system [1].
- Attention throughput doubles by NVIDIA's specification [1, 5].
- Networking bandwidth doubles from 0.8 to 1.6 TB/s per system, and the host link moves to PCIe Gen6 [1, 3].
- INT8 falls from 72 to 3 POPS sparse, and FP64 from 296 to 10 TFLOPS [1].
Memory bandwidth (7.7 TB/s per GPU), FP8 (36 PFLOPS dense per system), FP16 and BF16, and NVLink (1.8 TB/s per GPU) stay level [1, 2, 3]. On paper, a workload limited by memory bandwidth or FP8 math gains nothing from B300. One limited by memory capacity, FP4 math or attention gains.
HGX B300 memory: 2.1 TB or 2.3 TB
NVIDIA's HGX page, Blackwell Ultra datasheet and DGX B300 datasheet give 2.1 TB, from 270 GB per GPU [1, 3, 13]. Its DGX B300 user guide gives "8 x 288 GB = 2.3 TB total" [14], and Supermicro and Dell list 288 GB GPUs for their HGX B300 servers [19, 21]. NVIDIA's blog notes that HBM capacity "varies by SKU" [5]. Every B300 system in MLPerf® Inference v6.1, NVIDIA's DGX B300 included, lists a 270 GB part [34]. Between a 270 GB and a 288 GB B300, an eight-GPU system differs by 144 GB (derived).
B200 shows the same split: 180 GB per GPU in the HGX B200 datasheet, against "up to 192 GB" in NVIDIA's technical brief [2, 4].
B300 FP64 performance and HPC work
At 1.2 TFLOPS per GPU, B300's native FP64 rate is about 3% of B200's and under 2% of H200's tensor-core rate (derived) [2, 3, 6]. NVIDIA's HPC-Benchmarks container notes that B300 "has low native FP64 throughput" and that HPL performance on it "is expected to be low" [22]. NVIDIA's August 2025 line card rates HGX B200 and HGX H200 for HPC and scientific work, and leaves HGX B300 unrated [11].
Emulation recovers part of this for matrix multiplication. cuBLAS emulates FP64 matrix products with methods based on the Ozaki schemes and by default keeps accuracy "equal or better than FP64" [23]. Its first release ran on INT8 tensor cores [24], which B300 cut from 4.5 to 0.15 POPS dense per GPU [4]. In a 2026 preprint, Uchino, Ozaki and Imamura built an FP8-based version that reached 61 to 65 TFLOPS on large matrices on a B200, and expect similar results on B300 [25]. On small matrices, emulation ran slower than native FP64 [25].
DGX B300 power consumption, size and cooling
NVIDIA calls HGX and DGX B300 systems "Standardized 8 GPU Blackwell Ultra configurations" [5]. What a DGX adds around the board is set out in DGX, HGX and MGX explained.
| DGX H200 | DGX B200 | DGX B300 | |
|---|---|---|---|
| Power | 10.2 kW max, standard build | 14.3 kW max | About 14 kW to 15.1 kW |
| Height | 8U | 10U | 10U |
| Weight | 130.45 kg | 142.4 kg | 168 kg with power supplies |
| Operating temperature | 5 to 30 °C | 10 to 35 °C | 10 to 35 °C |
| Cluster network | 8 x ConnectX-7, 400 Gb/s | 8 x ConnectX-7, 400 Gb/s | 8 x ConnectX-8, 800 Gb/s |
Sources: [12, 13, 14, 15, 16, 17, 18].
NVIDIA states DGX B300 power three ways: about 14 kW on its product page, 14.5 kW with busbar power or 15.1 kW with power supplies in its February 2026 datasheet, and 14.5 kW with a 15 kW system maximum in its user guide [12, 13, 14].
Supermicro also builds HGX B300 into an 8U air-cooled server and a 4U liquid-cooled one [19, 20]. Rack-level limits are covered in rack power and liquid cooling, and the parts behind a server's price in what an HGX B200 or B300 server price includes.
H200 vs B200 for inference: MLPerf® results
MLPerf® Inference's closed division holds the model fixed so hardware can be compared [26, 28]. Pairs from the same submitter in the same round:
| Submitter, round | Systems | Benchmark | Server | Offline |
|---|---|---|---|---|
| Nebius, v5.1 | 8x B200 over 8x H200 | 70B question answering | 2.99x | 2.91x |
| CoreWeave, v6.1 | 8x B300 over 8x B200 | 70B question answering | 1.12x | 1.12x |
| Nebius, v6.1 | 8x B300 over 8x B200 | 671B reasoning | 1.14x | 1.22x |
| CoreWeave, v6.1 | 8x B300 over 8x B200 | 117B mixture of experts | 1.23x | 1.23x |
Ratios of published tokens per second (derived) [34, 35].
From H200 to B200 the measured gain is about three times, with B200 running FP4 weights and H200 FP8 [35]. From B200 to B300 the measured gain is 1.1 to 1.2 times, against 1.5 times for dense FP4 and 2 times for attention on the datasheet [1, 34].
MLPerf® notes: MLPerf® Inference: Datacenter v5.1 and v6.1, Closed division, Available category, Server and Offline scenarios, benchmarks llama2-70b-99.9, deepseek-r1 and gpt-oss-120b. Result IDs 5.1-0073, 5.1-0075, 6.1-0021, 6.1-0022, 6.1-0077 and 6.1-0078, retrieved 24 October 2026 from the MLCommons results repositories [34, 35]. Ratios are this article's arithmetic and are not MLPerf® metrics. The MLPerf name and logo are registered and unregistered trademarks of MLCommons Association in the United States and other countries. All rights reserved. Unauthorized use strictly prohibited. See www.mlcommons.org for more information.
Rillor supplies HGX H100, H200, B200 and B300 systems, new or secondary, through its GPU systems service. Each quote states the exact build, its condition and the tests run before delivery.
Read the full report
The full report adds eight-GPU totals for every precision, DGX power supply details, the FP64 emulation research and the MLPerf® results behind each ratio.