B300 beats B200 on memory (288 GB against 180 GB per GPU), dense FP4 throughput (15 against 10 PFLOPS per GPU) and attention speed (2x, per NVIDIA), and it matches B200 on dense FP8, FP16 and BF16 compute, NVLink bandwidth and memory bandwidth. It gives up most FP64 and INT8 throughput to get there. If your bottleneck is memory capacity or FP4 inference, pick B300; if it is not, B200 does the same FP8 and BF16 work.
For each chip in depth, see our B200 guide and B300 guide. All of this sits inside our datacenter GPU guide.
TL;DR
- B300 wins: memory per GPU (288 GB vs 180 GB), dense FP4 (1.5x), attention (2x, NVIDIA's claim), and network bandwidth per node (1.6 TB/s vs 0.8 TB/s).
- They tie: dense FP8, FP16/BF16 and TF32 tensor throughput, 8 TB/s memory bandwidth, and 1.8 TB/s NVLink per GPU.
- B200 wins: FP64 and INT8 throughput, and a lower power ceiling (up to 1,000 W per GPU against up to 1,400 W).
- Verdict: B300 for long-context and large-model FP4 inference. B200 for FP8 or BF16 training and inference where the model fits in 180 GB, and for anything that needs FP64.
Side-by-side specs
Per-GPU figures are from NVIDIA's Blackwell Ultra technical blog and its HGX reference architecture. Board figures are from NVIDIA's HGX page, where both are eight-GPU baseboards.
| Spec | B200 | B300 | Difference |
|---|---|---|---|
| Memory per GPU | 180 GB HBM3e | 288 GB HBM3e | 1.6x |
| Memory bandwidth per GPU | Up to 8 TB/s | 8 TB/s | Same |
| NVFP4 dense per GPU | 10 PFLOPS | 15 PFLOPS | 1.5x |
| NVFP4 sparse per GPU | 20 PFLOPS | 20 PFLOPS | Same |
| FP8 dense per GPU | 5 PFLOPS | 5 PFLOPS | Same |
| Attention performance | 1x | 2x | NVIDIA claim |
| NVLink per GPU | 1.8 TB/s | 1.8 TB/s | Same |
| Max power per GPU | Up to 1,000 W | Up to 1,400 W | +400 W |
| Board FP4, sparse / dense | 144 / 72 PFLOPS | 144 / 108 PFLOPS | 1.5x dense |
| Board FP8 / FP6 (sparse) | 72 PFLOPS | 72 PFLOPS | Same |
| Board FP16 / BF16 (sparse) | 36 PFLOPS | 36 PFLOPS | Same |
| Board FP64 | 296 TFLOPS | 10 TFLOPS | B200 far higher |
| Board INT8 | 72 POPS | 3 POPS | B200 far higher |
| Board total memory | 1.4 TB | 2.1 TB | 1.5x |
| Board network bandwidth | 0.8 TB/s | 1.6 TB/s | 2x |
The 8-GPU board rows are NVIDIA's published HGX figures. The per-GPU rows come from a different NVIDIA document, so they do not divide exactly: for example, 288 GB times eight is about 2.3 TB, while NVIDIA lists 2.1 TB for the board.
Where B300 pulls ahead
Memory capacity. 288 GB per GPU means a model replica fits on fewer GPUs, and a long context leaves more room for KV cache. Both matter most for reasoning models and large mixture-of-experts models, where capacity, not compute, often sets how many GPUs you rent.
FP4 throughput. B300 delivers 15 PFLOPS dense NVFP4 per GPU against 10 for Blackwell, NVIDIA says. NVIDIA's DGX B300 page states the same 1.5x for the full system. To use it you must actually run FP4, which means a model quantized to NVFP4 and an engine with FP4 kernels.
Attention. NVIDIA doubled special function unit throughput for the exponential used in softmax, from 5 to 10.7 tera-exponentials per second, and claims up to 2x faster attention-layer compute. This is a vendor claim for the attention layer alone, not an end-to-end 2x.
Networking. The DGX B300 page lists eight ConnectX-8 adapters at up to 800 Gb/s, against eight ConnectX-7 at up to 400 Gb/s on DGX B200. If you scale across nodes, that is a real difference.
Where B200 holds its ground
Everything that is not FP4. NVIDIA's HGX table shows identical FP8/FP6, FP16/BF16 and TF32 tensor throughput on the two boards. A BF16 training run or an FP8 inference service gets no raw compute gain from B300. The gain, if any, comes from the extra memory.
FP64 and INT8. B300's board FP64 is listed at 10 TFLOPS against 296 for B200, and INT8 at 3 POPS against 72. Anything that depends on double precision, such as scientific simulation, or on INT8 quantization is better served by B200. NVIDIA does not explain the reduction on the page we read; the pattern is consistent with Blackwell Ultra being tuned for low-precision AI, but that is our reading, not NVIDIA's statement.
Power. B200 is configurable up to 1,000 W per GPU. Blackwell Ultra is listed at up to 1,400 W. The DGX systems land close together at about 14.3 kW for DGX B200 and about 14 kW for DGX B300, per NVIDIA, so at the system level power is not a deciding factor, but check the vendor's configuration of any HGX server.
Your workload decides. NVIDIA publishes no head-to-head benchmark of eight-GPU B300 against B200 on the pages we read, and we do not publish one either. Measure your own model on both before committing.
Published performance, labelled
NVIDIA states that DGX B300 "boosts dense FP4 performance by 1.5x and attention performance by 2x over DGX B200" on its DGX B300 page, with no footnote.
For measured results, MLPerf Inference v5.1 (September 9, 2025) reported DeepSeek-R1 per-GPU throughput of 4,024 tokens per second offline on GB200 NVL72 and 5,842 on GB300 NVL72, about 45 percent higher, with the server scenario at 2,327 and 2,907, about 25 percent higher. These are rack-scale systems that use Blackwell and Blackwell Ultra GPUs, not eight-GPU HGX boards, and the per-GPU figures are NVIDIA's division of system throughput by GPU count. Treat them as the best published evidence of the gap on one reasoning model with NVFP4, not a forecast for your workload. Note also the gap there (up to 45 percent) is smaller than the 1.5x FP4 claim, and the MLPerf run mixes memory, attention and software effects.
Which one to pick
| If your workload is | Pick | Why |
|---|---|---|
| Long-context or reasoning inference in NVFP4 | B300 | Memory, FP4 and attention all help |
| Very large MoE model that barely fits on B200 | B300 | 288 GB cuts GPUs per replica |
| FP8 or BF16 training | B200 | Same compute, lower power ceiling |
| Fine-tuning a model under 100B parameters | B200 | Fits in 180 GB with room |
| Anything needing FP64 | B200 | B300 FP64 is far lower |
| Multi-node jobs limited by network | B300 | 2x node network bandwidth |
Quantization is the swing factor. If you cannot or will not run FP4, most of B300's headline advantage disappears and only memory capacity is left. The quantization glossary entry covers the accuracy trade-offs, and FP4 explains the format.
Cost: price per useful unit
Do not compare hourly prices alone. Compare cost per token (or per training step) using your measured throughput: divide each GPU's live hourly price by its tokens per hour. B300 only wins if its extra memory or FP4 throughput raises tokens per hour by more than the price gap, or lets you use fewer GPUs per replica. Use the live prices below, plus /pricing and the GPU index, to run that sum for your own workload.
Rent today
Aquanode manages and optimizes GPUs for training and inference workloads, and you can rent the GPUs in the box below on demand. The box shows live prices, or "None right now" when there is no offer.
Specs and availability live on the B300 page and the B200 page.
What's next
Both chips are succeeded by Rubin. NVIDIA's Rubin page describes the Vera Rubin platform as in full production but gives no quantified comparison to Blackwell and no availability window. See the Rubin guide. For rack-scale versions of these two chips, read GB300 NVL72 vs GB200 NVL72.
FAQ
Is B300 faster than B200?
Only for some work. NVIDIA lists 1.5x dense FP4 and 2x attention for B300. Its HGX table lists the same dense FP8, FP16/BF16 and TF32 throughput for both.
Does B300 have more memory than B200?
Yes: 288 GB of HBM3e per GPU against 180 GB in the HGX and DGX B200 configuration. Memory bandwidth is the same, at 8 TB/s per NVIDIA.
Why is B300 FP64 so low?
NVIDIA's HGX table lists 10 TFLOPS of FP64 for the B300 board against 296 TFLOPS for B200. NVIDIA does not give a reason on that page, so check your FP64 needs before choosing B300.
Should I upgrade from B200 to B300?
If you run FP4 inference that is memory or attention bound, it is worth testing. If you train in BF16 or FP8 and your model fits, the compute is the same, so measure before paying more.
Do B300 and B200 use the same NVLink?
Yes. Both use fifth-generation NVLink at 1.8 TB/s per GPU and 14.4 TB/s across the eight-GPU baseboard.
Sources
- NVIDIA HGX platform (B300 and B200 spec table): https://www.nvidia.com/en-us/data-center/hgx/
- NVIDIA DGX B300: https://www.nvidia.com/en-us/data-center/dgx-b300/
- NVIDIA DGX B200: https://www.nvidia.com/en-us/data-center/dgx-b200/
- NVIDIA, Inside NVIDIA Blackwell Ultra: https://developer.nvidia.com/blog/inside-nvidia-blackwell-ultra-the-chip-powering-the-ai-factory-era/
- NVIDIA HGX AI Factory reference architecture (B200 memory and bandwidth): https://docs.nvidia.com/enterprise-reference-architectures/hgx-ai-factory-h100-h200-b200/latest/components.html
- NVIDIA, Blackwell Ultra sets new inference records in MLPerf debut (MLPerf Inference v5.1): https://developer.nvidia.com/blog/nvidia-blackwell-ultra-sets-new-inference-records-in-mlperf-debut/
- NVIDIA Rubin platform page: https://www.nvidia.com/en-us/data-center/technologies/rubin/
