The 5-step pipeline.
How energy turns into a deployed, financed, offtake-matched cluster, and what SLYD does at every step.
How SLYD works →Platform
SLYD Cloud
New and recovered GPU systems.
NVIDIA and AMD systems through documented manufacturer and qualified channel supply, with financing and deployment coordinated on the same platform.
Explore GPU hardware →By accelerator
Infrastructure and services
Compute, hardware, and power in one book.
Browse available GPU capacity by accelerator, configuration, region, and price, or bring supply to qualified demand.
Open marketplace →Compute
Bring supply
For buyers
Start with an indicative structure.
Tell us deal size, structure, and offtake. Any range is preliminary and subject to underwriting, diligence, and documentation.
Open Configure →For lenders
GPU market trends and deployment playbooks.
Infrastructure best practices, hardware comparisons, and industry analysis from the SLYD team.
Read the blog →Company
NVIDIA Blackwell B200 and B300 systems
Blackwell is not one product. B200 and B300 arrive inside specific server platforms with different memory, networking, double-precision, and cooling profiles, and the differences between them are decisive more often than the headline compute number.
What are NVIDIA B200 and B300 systems?
NVIDIA B200 and B300 are Blackwell-generation accelerators delivered through specific server and rack platforms. Buyers must compare the exact GPU variant, memory, form factor, interconnect, server configuration, cooling design, power requirement, software support, and commercial terms rather than treating every Blackwell system as interchangeable.
Four different things get called Blackwell
Most specification confusion in Blackwell procurement comes from mixing these up. They are separate products with separate published numbers.
B200 and B300 GPUs
The Blackwell and Blackwell Ultra GPUs themselves. They are not sold as standalone cards for most enterprise deployments.
HGX B200 and HGX B300
The 8-GPU baseboard platforms NVIDIA publishes specifications against, and what server manufacturers build their systems around.
DGX systems
NVIDIA's own complete systems, with NVIDIA's software, support, and reference architecture attached.
GB200 and GB300 NVL72
Rack-scale Grace Blackwell platforms with an integrated CPU and a 72-GPU NVLink domain. Covered on the Grace Blackwell page.
HGX B300 and HGX B200 as NVIDIA publishes them
These are 8-GPU platform totals, reproduced as published. Per-GPU figures are deliberately not derived by division here, because the platform total is what NVIDIA states and what a server is actually quoted against.
| Specification | HGX B300 | HGX B200 |
|---|---|---|
| Form factor | 8x NVIDIA Blackwell Ultra SXM | 8x NVIDIA Blackwell SXM |
| Total memory | 2.1 TB | 1.4 TB |
| FP4 Tensor Core | 144 PFLOPS sparse, 108 PFLOPS dense | 144 PFLOPS sparse, 72 PFLOPS dense |
| FP8 or FP6 Tensor Core (sparse) | 72 PFLOPS | 72 PFLOPS |
| FP16 or BF16 Tensor Core (sparse) | 36 PFLOPS | 36 PFLOPS |
| TF32 Tensor Core (sparse) | 18 PFLOPS | 18 PFLOPS |
| FP32 | 600 TFLOPS | 600 TFLOPS |
| FP64 and FP64 Tensor Core | 10 TFLOPS | 296 TFLOPS |
| NVLink | Fifth generation, NVLink 5 Switch | Fifth generation, NVLink 5 Switch |
| NVLink GPU-to-GPU bandwidth | 1.8 TB/s | 1.8 TB/s |
| Total NVLink bandwidth | 14.4 TB/s | 14.4 TB/s |
| Networking bandwidth | 1.6 TB/s | 0.8 TB/s |
CPU and system memory specifications for these platforms are defined by the manufacturer's server design, not by the HGX platform, so they are confirmed against the specific server model.
Where the choice is actually made
Three of the published differences change real decisions. The rest of the table is identical between the two platforms.
Memory capacity
2.1 TB against 1.4 TB per 8-GPU platform. If the working set, context length, or KV cache is the binding constraint, this is the difference that decides it and no amount of extra compute substitutes for it.
Double precision
296 TFLOPS of FP64 on HGX B200 against 10 TFLOPS on HGX B300. Simulation and technical computing workloads that depend on FP64 should treat this as disqualifying rather than as a tradeoff.
Networking bandwidth
1.6 TB/s against 0.8 TB/s. This shapes how the platform scales out across a cluster and how much switch and optic capacity the fabric design has to carry.
The server around the platform
Rack height, cooling method, CPU choice, memory, storage, and network adapters are all manufacturer decisions, and they change what the system costs to deploy far more than the accelerator choice alone.
Where B200 and B300 appear
Systems published by the manufacturers in their own current documentation, reviewed 18 August 2026. Named to show where the accelerators appear, not to indicate that any system is available.
| Manufacturer | Published system | As described by the manufacturer |
|---|---|---|
| Dell | PowerEdge XE9780 | 10U, 8 x NVIDIA HGX B200 or B300 |
| Dell | PowerEdge XE9780L and XE9680L | 8 x NVIDIA HGX B300 and 8 x NVIDIA HGX B200 respectively |
| HPE | ProLiant Compute XD685 | 8-way GPU server, 5U for direct liquid cooling and 6U for air, with a choice of eight B300 HGX, B200, H200, or AMD Instinct MI355X accelerators |
| HPE | Compute XD690 | Eight NVIDIA Blackwell Ultra (B300 HGX) GPUs |
| Supermicro | Liquid-cooled HGX B300 systems | 4U for 19-inch EIA racks and 2-OU for 21-inch OCP ORV3 racks |
| Lenovo | ThinkSystem SR680a V3 | Published with B200 |
What the server needs around it
Rack power design
Dense 8-GPU servers concentrate load. Rack power, distribution, and redundancy follow from the exact server model and how many go in a rack.
CoolingAir or direct liquid
The same platform ships in both forms with different rack heights and facility requirements. Confirm which one the quote describes.
NetworkScale-out fabric and adapters
NVLink handles the 8 GPUs inside the server. Connecting servers into a cluster is a separate fabric with its own adapters, optics, and switches.
StorageData path and capacity
Sustained throughput during real job phases, not peak specifications, is what determines whether accelerators stay busy.
Confirm before ordering
- Exact GPU variant and whether the quote is HGX, DGX, or rack scale
- GPU count per node and total memory for that configuration
- CPU, system memory, storage, and network adapter configuration
- Cooling method and rack height
- Power draw and connector requirements
- Firmware and software stack compatibility
- Equipment condition and provenance
- Warranty provider, scope, term, and transferability
- Availability confirmation with its date and expiration
- Export, end-use, and geographic restrictions
Qualifying recovered supply is evaluated at the lot level for configuration, condition, provenance, seller authority, availability, and commercial terms. Testing and warranty evidence are disclosed for the specific lot when they exist.
Blackwell server questions
What is the difference between NVIDIA B200 and B300?
B300 is the Blackwell Ultra GPU and B200 is the Blackwell GPU. At the 8-GPU platform level NVIDIA publishes HGX B300 with 2.1 TB of total memory and 1.6 TB/s of networking bandwidth, and HGX B200 with 1.4 TB of total memory and 0.8 TB/s of networking bandwidth. B300 also trades double-precision throughput away, which matters for HPC.
Is HGX the same thing as DGX or a GB200 rack?
No. HGX is the 8-GPU baseboard platform that server manufacturers build systems around, DGX is NVIDIA's own complete system, and GB200 and GB300 NVL72 are rack-scale Grace Blackwell platforms with an integrated CPU and a 72-GPU NVLink domain. Specifications are not interchangeable between them, so a quote has to say which one it describes.
Should I choose B200 or B300 for HPC and simulation?
Check the double-precision requirement first. NVIDIA publishes 296 TFLOPS of FP64 and FP64 Tensor Core performance for HGX B200 against 10 TFLOPS for HGX B300. If a workload depends on FP64 throughput, that difference is decisive regardless of how the low-precision figures compare.
Why are NVIDIA's FP4 figures written as two numbers?
NVIDIA publishes FP4 Tensor Core performance as sparse and then dense. HGX B300 is published as 144 PFLOPS sparse and 108 PFLOPS dense, and HGX B200 as 144 PFLOPS sparse and 72 PFLOPS dense. Quoting only the first number without saying it assumes sparsity overstates what most workloads will see.
Which manufacturers build B200 and B300 servers?
Server manufacturers publish their own implementations. Dell publishes PowerEdge XE9780 with 8 NVIDIA HGX B200 or B300 GPUs and XE9780L with 8 HGX B300, HPE publishes ProLiant Compute XD685 with a choice of eight B300 HGX, B200, H200, or AMD Instinct MI355X accelerators, Supermicro publishes 4U and 2-OU OCP liquid-cooled HGX B300 systems, and Lenovo publishes ThinkSystem SR680a V3 with B200.
Do Blackwell servers need liquid cooling?
It depends on the server, not on the accelerator generation. Manufacturers publish both air-cooled and direct-liquid-cooled implementations, and some publish the same platform in both forms with different rack heights. HPE publishes ProLiant Compute XD685 at 5U for direct liquid cooling and 6U for air, which is a good illustration that the cooling choice changes the physical system.
What does SLYD publish about Blackwell pricing and availability?
Nothing static. Price, condition, availability, lead time, and warranty change by product, supplier, geography, and transaction, so they are confirmed for the specific configuration and quantity requested rather than published as page content.
Sources for the specifications on this page
- NVIDIA, HGX platform page. Reviewed 18 August 2026. nvidia.com/en-us/data-center/hgx
- Dell, PowerEdge AI servers catalog. Reviewed 18 August 2026. dell.com PowerEdge AI servers
- HPE, ProLiant Compute XD685 product page. Reviewed 18 August 2026. hpe.com ProLiant Compute XD685
- Supermicro, liquid-cooled NVIDIA HGX B300 announcement. Reviewed 18 August 2026. supermicro.com press release
- Lenovo Press, AI servers documentation. Reviewed 18 August 2026. lenovopress.lenovo.com/servers/ai-servers
Specifications belong to the manufacturers named and are reproduced with their own caveats. They are not SLYD guarantees of performance, configuration, or availability.
Scope a Blackwell server requirement
Share the workload, node count, site, cooling constraint, and timeline. The right configuration falls out of those, not out of the accelerator name.
Page updated: August 18, 2026