Short answer: NVIDIA B300 Blackwell Ultra is marketed as a 288GB HBM3E accelerator, but NVIDIA’s Blackwell Ultra datasheet lists the HGX B300 GPU at 270GB HBM3E and 7.7TB/s, while NVIDIA’s enterprise architecture table describes the tier as 288GB and up to 8TB/s. For purchase and model-fit planning, verify the exact platform and use the lower usable figure until the supplier proves otherwise. At 270GB, the listed Qwen3.5-122B BF16 files fit only narrowly, and Qwen3.5-397B 4-bit files fit with limited but meaningful file margin.
B300 is not a 288GB desktop GPU. It is a Blackwell Ultra SXM-class datacenter accelerator sold in qualified HGX/DGX systems. The complete system, power, cooling, firmware, CUDA stack, networking and support matter more than a bare-module listing.
B300 capacity: 288GB marketed, 270GB usable in HGX documentation
| NVIDIA source / context | Memory | Bandwidth | Planning use |
|---|---|---|---|
| Enterprise HGX architecture table | 288GB HBM3E per GPU | Up to 8TB/s | Product-class headline |
| Blackwell Ultra datasheet, HGX B300 | 270GB HBM3E per GPU | 7.7TB/s | Conservative usable-device planning |
| DGX B300 system | 8 × 288GB = 2.3TB total | Platform specification | Aggregate system headline, not one pool |
This page reports both official figures instead of silently treating them as interchangeable. Confirm what the intended system exposes to CUDA and the runtime before committing to a boundary deployment.
B300 specifications that matter for local AI
| Architecture | NVIDIA Blackwell Ultra |
|---|---|
| Memory | 288GB HBM3E product tier; 270GB listed for individual HGX B300 GPU in NVIDIA’s datasheet |
| Memory bandwidth | Up to 8TB/s product tier; 7.7TB/s in the HGX B300 datasheet |
| Interconnect | Fifth-generation NVLink, up to 1.8TB/s per GPU |
| Host interface | PCIe Gen 6, 256GB/s |
| Maximum TDP | Configurable up to 1,100W for HGX B300 |
| Multi-instance GPU | Up to seven MIG instances |
| Eight-GPU HGX/DGX capacity | 2.3TB marketed aggregate |
What fits on one B300?
Qwen3.5-122B-A10B using 270GB for planning
| Artifact | Files | Margin in 270GB | Capacity result |
|---|---|---|---|
| GPTQ 4-bit | 78.864GB | 191.136GB | Fits by files |
| Q4 plus projector | 77.446GB | 192.554GB | Fits by files |
| Official FP8 | 127.163GB | 142.837GB | Fits by files |
| Q8 plus projector | 130.781GB | 139.219GB | Fits by files |
| Official BF16 | 250.173GB | 19.827GB | Fits by files; tight runtime boundary |
The BF16 row is not a deployment guarantee. Its 19.827GB remainder must absorb CUDA context, cache, temporary buffers, multimodal processing and implementation overhead. Measure peak device use at the intended context and concurrency.
Qwen3.5-397B-A17B using 270GB per GPU
| Artifact | Files | One B300? | Arithmetic minimum |
|---|---|---|---|
| GPTQ 4-bit | 235.708GB | Yes; 34.292GB file margin | 1 |
| Q4 plus projector | 245.012GB | Yes; 24.988GB file margin | 1 |
| Official FP8 | 406.152GB | No | 2 × 270GB |
| Q8 plus projector | 422.426GB | No | 2 × 270GB |
| Official BF16 | 806.796GB | No | 3 × 270GB is only 810GB and not practical |
Three 270GB devices exceed the BF16 file total by only 3.204GB before any runtime allocation, so four devices are the credible starting capacity tier. These are still separate memory spaces; the framework must shard the model correctly.
B300 vs B200 vs H200
| GPU | Architecture | Headline memory | Headline bandwidth |
|---|---|---|---|
| H200 SXM | Hopper | 141GB HBM3E | 4.8TB/s |
| B200 SXM | Blackwell | 180GB HBM3E | Up to 8TB/s |
| B300 SXM | Blackwell Ultra | 288GB HBM3E tier; 270GB in HGX datasheet | Up to 8TB/s; 7.7TB/s in HGX datasheet |
B300 changes the single-device boundary for very large 4-bit models and some 250GB-class BF16 checkpoints. It does not turn an HGX system into a workstation, and it does not remove the need to validate framework, CUDA, driver, container and kernel support for Blackwell Ultra.
B300 vs AMD’s 288GB MI350 series
B300 and MI350X/MI355X occupy the same headline 288GB capacity class, but the platforms differ. B300 uses the CUDA ecosystem and fifth-generation NVLink; AMD uses ROCm, CDNA 4 and Infinity Fabric. MI350X is a 1000W air-cooled option, MI355X is a 1400W direct-liquid-cooled option, and HGX B300 is listed up to 1100W. Compare exact framework support, usable memory, complete-system price, cooling, topology and measured workload results—not nominal capacity alone.
Validation checklist
- Obtain the exact system SKU and confirm whether NVIDIA reports 270GB usable per HGX B300 GPU.
- Record driver, CUDA, framework, container, kernels and model revision.
- Record per-device placement, topology and any CPU-memory or storage offload.
- Measure peak device memory, context, batch/concurrency, time to first token, prompt and generation throughput, throttling and wall power.
- Compare complete supported systems, not bare SXM modules.
Plan the next step
- Build a reproducible benchmark record
- Compare H200 and B200
- Review H200 form factors
- Compare AMD’s 288GB MI350 tier
- See Qwen3.5-122B capacity tiers
- See Qwen3.5-397B capacity tiers
- Compare all large-model hardware paths
- Browse the hardware guide
- Check the live hardware market
Primary sources
- NVIDIA HGX AI Factory components and H200/B200/B300 comparison
- NVIDIA Blackwell Ultra datasheet
- NVIDIA DGX B300 user guide
Capacity calculations use decimal GB. Product specifications and software support can change; confirm the exact system’s current documentation before purchase.
For the direct same-capacity alternative, compare NVIDIA B300 with AMD MI355X across usable memory, software, power and model fit.
Choosing between Blackwell generations? Compare NVIDIA B200 and B300 for local AI using conservative usable-memory math, exact Qwen3.5 fit, software validation and complete-system requirements.
Comparing the previous Blackwell tier? Review NVIDIA B200 180GB for local AI for exact capacity boundaries and platform checks.
Choosing among four NVIDIA datacenter tiers? Use the H100/H200/B200/B300 decision hub for one capacity and platform comparison.