The 20 petaflops is NVIDIA's peak FP4 rating with sparsity, a ceiling for the GPU's Tensor Cores rather than a measure of everyday speed. In independent testing, the NVIDIA DGX Station GB300 delivered about 6.1 petaflops of real NVFP4 matrix maths, which is still a great deal of compute for one desk. That compute is what lets the system read long prompts quickly, serve over a hundred simultaneous AI requests and fine-tune models in-house. It is not what makes one person's chatbot answer appear faster. Single-user reply speed depends mainly on memory bandwidth, and the GB300 is built around that too.
What is actually inside the GB300 superchip?
The GB300 Grace Blackwell Ultra Desktop Superchip combines two processors in one package, joined by NVIDIA's NVLink-C2C link:
- Blackwell Ultra GPU: 252 GB of HBM3e memory at 7.1 TB/s, with Blackwell-generation Tensor Cores that handle FP4, FP6, FP8, FP16 and BF16 maths.
- Grace CPU: 72 Arm Neoverse V2 cores with SVE2 vector units and 114 MB of L3 cache, attached to 496 GB of LPDDR5X memory at 396 GB/s.
- NVLink-C2C: a 900 GB/s coherent link, so the GPU and CPU can address each other's memory. NVIDIA quotes the total as 748 GB of coherent memory.
- ConnectX-8 SuperNIC: two 400 Gb/s QSFP112 ports for linking a second system or fast storage.
The whole system runs within a 1,600 W budget, which the CPU, GPU and any add-in graphics card share. Our posts on what "coherent" memory means and on the 72-core Grace CPU go deeper on those two parts. This article is about the compute.
How close does real hardware get to the headline numbers?
| Precision | NVIDIA peak (with sparsity) | Measured matrix throughput | Measured as share of peak |
|---|---|---|---|
| NVFP4 (4-bit) | 20 PFLOPS | 6,134 TFLOPS | 31% |
| FP8 | 10 PFLOPS | 3,909 TFLOPS | 39% |
| BF16 | 5 PFLOPS | 1,967 TFLOPS | 39% |
| HBM3e read bandwidth | 7.1 TB/s | about 6,875 GB/s | 97% |
Source: NVIDIA DGX Station datasheet (Aug 2026) for peak figures; StorageReview's MSI XpertStation WS300 review (24 Aug 2026), using the MAMF matrix benchmark, for measured figures. Percentages are SourceIT's arithmetic. We have not tested the system ourselves.
The gap is not a defect. NVIDIA's headline figures assume 2:4 structured sparsity, where half the values in a matrix are zero and the hardware skips them. Most inference workloads don't use structured sparsity, and dense maths typically runs at about half the sparse rating. NVIDIA's own spec table appears to give about 15 PFLOPS for dense FP4, and against that the measured 6.1 PFLOPS is about 41%. Real kernels also lose time to data movement, and the GPU shares its power budget with the CPU. The memory figure tells the more useful story: the HBM3e delivered 97% of its rated bandwidth, and bandwidth is what governs token generation.
Petaflops tell you how fast the system can think in bulk. Memory bandwidth tells you how fast it can talk to one person.
Which jobs actually use all that compute?
AI inference has two phases, and they stress the hardware differently.
- Prefill (reading the prompt): the model processes every input token at once. That is large matrix maths, so it is compute-bound and benefits directly from the Tensor Cores. In StorageReview's ASUS ExpertCenter Pro ET900N G3 review, GPT-OSS-120B on an 8,192-token-in, 1,024-token-out workload at 128 concurrent requests processed about 45,205 tokens per second in total, of which 5,023 were output tokens. Most of the work was reading long prompts.
- Decode (writing the answer): the model produces one token at a time per request, reading its weights from memory for each step. With one user, that is limited by memory bandwidth, not compute. StorageReview measured MiniMax M2.7 at 214 output tokens per second for a single stream, rising to 4,793 at 128 concurrent streams, because batching lets each weight read serve many requests.
So the compute shows up when many people use the system at once, when prompts are long (document analysis, retrieval-augmented generation, code review), and when you train or fine-tune. Typical workloads that keep the Tensor Cores busy include:
- Shared inference for a team or department: dozens of concurrent chat, coding or agent sessions.
- Long-document work: summarising contracts, analysing reports, answering questions over a document library.
- Embedding and indexing: converting thousands of documents into vectors for search, which is mostly prefill.
- Fine-tuning: adapting an open-weight model to your data with techniques such as LoRA, where BF16 and FP8 throughput matter.
- Batch jobs: overnight classification, extraction or translation across large volumes of text.
Our separate guide What Can You Run on a DGX Station GB300? covers which models fit. For how compute turns into a token count, see How Many AI Tokens Can DGX Station Actually Generate?
Where does the Grace CPU fit in?
The Grace CPU is not just a host. It runs the serving software, tokenises text, prepares data and handles everything outside the GPU, and its 496 GB of LPDDR5X can hold data the GPU reaches over NVLink-C2C. StorageReview measured about 390 GB/s from Grace memory into the GPU, limited by the LPDDR5X itself rather than the 900 GB/s link. That makes Grace memory useful for overflow, but much slower than HBM3e for anything read every token. 252GB HBM3e vs 496GB LPDDR5X explains how the two memory tiers divide the work.
Because Grace is an Arm processor, software has to run natively on aarch64. Most AI frameworks and NVIDIA's containers do; some engineering and x86-only tools don't. Check your stack before you buy.
What does the power budget mean for performance?
NVIDIA's documentation describes "dynamic power sloshing": the 1,600 W budget moves between CPU and GPU as load changes. If you add an RTX PRO display card, the GB300's cap falls to 1,250 W for a card drawing 76–300 W and to 950 W for a card above 300 W. A card at 75 W or less leaves the cap unchanged. For maximum AI throughput, keep any display card small. Our post on why a DGX Station might need another graphics card sets out the options. Independent reviewers have found the different OEM versions within about 2% of each other on the same tests, so the OEM you choose matters less for raw speed than the configuration (see why two DGX Stations might still differ).
So what can 20 petaflops actually do for a business?
It turns one deskside machine into shared AI infrastructure. The compute lets a single system serve a department's concurrent requests, process long documents in bulk and fine-tune open-weight models without renting cloud GPUs. It will not make a single chat session dramatically faster than a smaller system with similar memory bandwidth per model. If your use is one person at a time with short prompts, a smaller system such as the NVIDIA DGX Spark may be enough. If you need high concurrency, long context or in-house training, the GB300's compute is the reason to choose it. For the memory side of that decision, see Can 748GB Really Handle a Trillion-Parameter AI Model?
Buying in Singapore
SourceIT lists five DGX Station GB300 systems, all price on request: the HP ZGX Fury G1n, MSI XpertStation WS300, ASUS ExpertCenter Pro ET900N G3, GIGABYTE W775-V10-L01 and Supermicro ARS-511GD-NB-LCC. Systems in this class typically come to around S$150,000 depending on OEM and configuration. For smaller single-user work, the NVIDIA DGX Spark 4TB is S$9,300 (listed 5 Nov 2026). Browse the full range in our NVIDIA DGX Spark and DGX Station collection. For a quotation, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
Does the DGX Station GB300 really deliver 20 petaflops? Only as a theoretical peak. The 20 PFLOPS figure is NVIDIA's FP4 rating with structured sparsity. StorageReview measured about 6,134 TFLOPS of NVFP4 matrix throughput on an MSI WS300, roughly 31% of the headline figure.
Why is the measured figure so much lower than the spec? The spec assumes 2:4 sparsity, which most inference doesn't use, and dense maths typically runs at about half the sparse rating. Real software also loses time to data movement, and the GPU shares a 1,600 W power budget with the CPU.
Does more compute make chatbot answers faster? Not much for a single user. Writing each answer token is limited by memory bandwidth. Compute matters most for long prompts, many simultaneous users, batch processing and fine-tuning.
What precision should I run models at? Many open-weight models now ship in FP8 or 4-bit formats such as NVFP4 and MXFP4, which halve or quarter memory use compared with BF16 and use the GB300's fastest Tensor Core paths. Test answer quality on your own tasks before standardising on 4-bit.
Do the HP, MSI, ASUS, GIGABYTE and Supermicro versions perform differently? Very little on the same tests. Reviewers found them within about 2% of each other. Configuration choices, such as adding a high-power display card, matter more because they reduce the GPU's power budget.
SourceIT Approved
Researched and verified by the SourceIT team for Singapore buyers.
