Every figure below was read from its source on 2026-10-13. A figure is either stated by the vendor, derived from stated figures (the formula is in the model), or not disclosed in the sources used: gaps are never filled with a guess. Hover a source tag for the source's title, its access date and where in it the figure is. Availability: product (a vendor product page lists the specifications), preliminary (the vendor marks them preliminary or projected), planned (the vendor gives a future date). The same table, generated by the model's script, is results.md.
GPUs
| | H100 SXM5 | H200 SXM | Blackwell (GB200 NVL72) | Blackwell Ultra (GB300 NVL72) | Rubin (Vera Rubin NVL72)preliminary |
|---|
| BF16, dense | 990 TFLOP/s (derived)[h100-page] | 990 TFLOP/s (derived)[h200-page] | 2,500 TFLOP/s (derived)[gb200-page] | 2,500 TFLOP/s (derived)[gb300-page] | 4,000 TFLOP/s[vr-page] |
|---|
| FP8, dense | 1,980 TFLOP/s (derived)[h100-page] | 1,980 TFLOP/s (derived)[h200-page] | 5,000 TFLOP/s[bwu-blog] | 5,000 TFLOP/s[bwu-blog] | 17,500 TFLOP/s[vr-page] |
|---|
| Compute dies | 1[bwu-blog] | 1[bwu-blog] | 2[bw-arch] | 2[bwu-blog] | not disclosed[vr-page] |
|---|
| HBM bandwidth | 3.35 TB/s[h100-page] | 4.8 TB/s[h200-page] | 8 TB/s (derived)[gb200-page] | 8 TB/s[bwu-blog] | 19.2 TB/s[vr-page] |
|---|
| HBM capacity | 80 GB[h100-page] | 141 GB[h200-page] | 186 GB (derived)[gb200-page] | 288 GB[bwu-blog] | 288 GB[vr-page] |
|---|
| HBM stacks | 5[hopper-blog] | not disclosed[h200-page] | not disclosed[bwu-blog] | 8[bwu-blog] | not disclosed[vr-page] |
|---|
| NVLink per GPU (both directions) | 900 GB/s[nvlink-page] | 900 GB/s[h200-page] | 1.8 TB/s[nvlink-page] | 1.8 TB/s[bwu-blog] | 3 TB/s[vr-page] |
|---|
| NVLink links per GPU | 18[nvlink-page] | 18[nvlink-page] | 18[nvlink-page] | 18[nvlink-page] | 36[nvlink-page] |
|---|
| PCIe (both directions) | 128 GB/s[bwu-blog] | 128 GB/s[h200-page] | 256 GB/s[bwu-blog] | 256 GB/s[bwu-blog] | not disclosed[vr-page] |
|---|
| Max power | 700 W[h100-page] | 700 W[h200-page] | 1.2 kW[bwu-blog] | 1.4 kW[bwu-blog] | not disclosed[vr-page] |
|---|
| Transistors | 80 billion[bwu-blog] | 80 billion[bwu-blog] | 208 billion[bw-arch] | 208 billion[bwu-blog] | not disclosed[vr-page] |
|---|
Vera Rubin NVL72 (preliminary)
Scale-out switches
| Switch | Ports | Optics | Availability |
|---|
| Quantum-2 (QM9700 class), InfiniBand NDR | 64[quantum2-page] × 400 Gb/s | pluggable | product |
| Quantum-X800, InfiniBand XDR | 144[qx800-page] × 800 Gb/s | pluggable | product |
| Spectrum-4 SN5600, Ethernet | 64[eth-switch-page] × 800 Gb/s | pluggable | product |
| Quantum-X Photonics Q3450, InfiniBand (co-packaged optics) | 144[photonics-page] × 800 Gb/s | co-packaged | plannedcommercial availability planned for early 2026 (blog dated 2025-08-18) |
| Spectrum-X Photonics SN6810, Ethernet (co-packaged optics) | 128[cpo-blog] × 800 Gb/s | co-packaged | plannedavailable in the second half of 2026 |
| Spectrum-X Photonics SN6800, Ethernet (co-packaged optics, four ASICs) | 512[cpo-blog] × 800 Gb/s | co-packaged | plannedavailable in the second half of 2026 |
DGX SuperPOD (GB200) design rules
Physical properties
Water's specific heat is computed from NIST's Shomate equation at 298.15 K (chapter 7); air's from NASA's ratio of specific heats and gas constant.
Other rack- and pod-scale designs
Google TPU v5p pod
| | Google TPU v5p pod |
|---|
| Accelerators | 8,960[tpu-v5p] |
|---|
| BF16, dense | 459 TFLOP/s[tpu-v5p] |
|---|
| HBM bandwidth | 2.765 TB/s[tpu-v5p] |
|---|
| Scale-up per chip (both directions) | 1.2 TB/s[tpu-v5p] |
|---|
| Scale-up topology | 3D torus; 4x4x4 cubes joined by optical circuit switches[tpu-v5p] |
|---|
Google TPU7x (Ironwood) pod
| | Google TPU7x (Ironwood) pod |
|---|
| Accelerators | 9,216[tpu-7x] |
|---|
| BF16, dense | 2,310 TFLOP/s[tpu-7x] |
|---|
| HBM bandwidth | 7.38 TB/s[tpu-7x] |
|---|
| Scale-up per chip (both directions) | 1.2 TB/s[tpu-7x] |
|---|
| Scale-up topology | not disclosed[tpu-7x] |
|---|
AMD Helios (72 x Instinct MI450) (planned)
AWS Trn2 UltraServer (64 x Trainium2)
AWS Trainium3 UltraServer (144 x Trainium3)
Llama 3 405B pre-training (chapter 9)
Illustrative cost inputs (chapter 11)
Round numbers chosen for the picture, not prices or quotes:
| Capital cost per rack (USD) | 3,500,000 (illustrative) |
|---|
| PUE (facility over IT power) | 1.2 (illustrative) |
|---|
| Electricity (USD per kWh) | 0.1 (illustrative) |
|---|
| Utilisation | 0.7 (illustrative) |
|---|
| Write-off period (years) | 4 (illustrative) |
|---|
Quoted claims
Shown where relevant, attributed, and never used as model inputs:
- 800 VDC 'enables 85% more power to be transmitted through the same conductor size' than 415 VAC, and 'up to 5%' better end-to-end efficiency than 54 V systems (NVIDIA 800 VDC Architecture Will Power the Next Generation of AI Factories (NVIDIA Technical Blog, 2025-05-20, updated 2025-07-31))
- co-packaged optics: '3.5x power efficiency' compared with pluggable transceivers (Scaling AI Factories with Co-Packaged Optics for Better Power Efficiency (NVIDIA Technical Blog, 2025-08-18))
- Rubin CPX: 'up to 30 petaflops of compute with NVFP4 precision', '128GB of cost-efficient GDDR7', '3x faster attention' than GB300 NVL72; 'expected to be available at the end of 2026' (NVIDIA Unveils Rubin CPX: A New Class of GPU Designed for Massive-Context Inference (NVIDIA Newsroom, 2025-09-09))
- DistServe 'can serve 7.4x more requests or 12.6x tighter SLO' than state-of-the-art systems (in its own evaluation) (Zhong et al., DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving (arXiv 2401.09670))
- GB200 NVL72: '25x more energy efficiency and 300x more water efficiency than traditional air-cooled architectures' (Chill Factor: NVIDIA Blackwell Platform Boosts Water Efficiency by Over 300x (NVIDIA Blog))
- the 72 GPUs of an NVL72 'communicate as one' (all-to-all at 1.8 TB/s per GPU) (NVIDIA GB200 NVL72 Delivers Trillion-Parameter LLM Training and Real-Time Inference (NVIDIA Technical Blog))
- 'connecting over 10,000 GPUs in a non-blocking, two-level fat-tree topology' (NVIDIA Silicon Photonics page (co-packaged optics switches))
- energy-storage power shelves: 'the peak power demand seen by the grid is reduced by 30% when training the Megatron LLM' (How New GB300 NVL72 Features Provide Steady Power for AI (NVIDIA Technical Blog))
- Vera Rubin NVL72 'delivers up to 10x more tokens per megawatt than NVIDIA GB200 NVL72' (Kimi-K2 Thinking, 32K/8K input/output) (NVIDIA Vera Rubin NVL72 product page, Specifications (marked 'Available Now' on the access date))
Sources
800v-blog: NVIDIA 800 VDC Architecture Will Power the Next Generation of AI Factories (NVIDIA Technical Blog, 2025-05-20, updated 2025-07-31) (vendor technical blog, read 2026-10-13)alfares2008: Al-Fares, Loukissas and Vahdat, A scalable, commodity data center network architecture, SIGCOMM 2008 (doi:10.1145/1402946.1402967) (paper, read 2026-10-13)bw-arch: NVIDIA Blackwell Architecture page (vendor product page, read 2026-10-13)bwu-blog: Inside NVIDIA Blackwell Ultra: The Chip Powering the AI Factory Era (NVIDIA Technical Blog; updated 2025-09-24) (vendor technical blog, read 2026-10-13)cpo-blog: Scaling AI Factories with Co-Packaged Optics for Better Power Efficiency (NVIDIA Technical Blog, 2025-08-18) (vendor technical blog, read 2026-10-13)cpx-news: NVIDIA Unveils Rubin CPX: A New Class of GPU Designed for Massive-Context Inference (NVIDIA Newsroom, 2025-09-09) (vendor press release, read 2026-10-13)deepseekv3: DeepSeek-AI, DeepSeek-V3 Technical Report (arXiv 2412.19437) (paper, read 2026-10-13)dgx-b200-page: NVIDIA DGX B200 product page, Specifications (vendor product page, read 2026-10-13)dgx-gb200-guide: NVIDIA DGX GB200 User Guide, Hardware (vendor documentation, read 2026-10-13)disagg-sim: Disaggregated_Inference_Sim: a discrete-event simulator of prefill/decode serving (this author's; the engine this site vendors) (simulator, read 2026-10-13)distserve: Zhong et al., DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving (arXiv 2401.09670) (paper, read 2026-10-13)dynamo-gh: NVIDIA Dynamo, the open-source datacenter-scale inference stack (GitHub repository README) (vendor software documentation, read 2026-10-13)eth-switch-page: NVIDIA Ethernet switching page (Spectrum-4 and Spectrum-6 switch tables) (vendor product page, read 2026-10-13)gb200-blog: NVIDIA GB200 NVL72 Delivers Trillion-Parameter LLM Training and Real-Time Inference (NVIDIA Technical Blog) (vendor technical blog, read 2026-10-13)gb200-page: NVIDIA GB200 NVL72 product page, Specifications (vendor product page, read 2026-10-13)gb300-page: NVIDIA GB300 NVL72 product page, Specifications (vendor product page, read 2026-10-13)h100-page: NVIDIA H100 product page, Product Specifications (H100 SXM column) (vendor product page, read 2026-10-13)h200-page: NVIDIA H200 product page, Specifications (H200 SXM column) (vendor product page, read 2026-10-13)helios-blog: AMD 'Helios': Advancing Openness in AI Infrastructure Built on Meta's 2025 OCP Open Rack for AI Design (AMD blog, 2025-10-14) (vendor blog, read 2026-10-13)hgx-page: NVIDIA HGX platform page, Specifications (vendor product page, read 2026-10-13)hopper-blog: NVIDIA Hopper Architecture In-Depth (NVIDIA Technical Blog) (vendor technical blog, read 2026-10-13)leiserson1985: Leiserson, Fat-trees: universal networks for hardware-efficient supercomputing, IEEE Trans. Computers 1985 (doi:10.1109/TC.1985.6312192) (paper, read 2026-10-13)llama3: Grattafiori et al., The Llama 3 Herd of Models (arXiv 2407.21783) (paper, read 2026-10-13)megatron2021: Narayanan et al., Efficient Large-Scale Language Model Training on GPU Clusters Using Megatron-LM (arXiv 2104.04473) (paper, read 2026-10-13)nasa-air: NASA Glenn Research Center, Air Properties Definitions (sea-level standard day) (reference data, read 2026-10-13)nasa-gas: NASA Glenn Research Center, Equation of State (specific gas constant of air) and Specific Heats (ratio for air) (reference data, read 2026-10-13)nist-water: NIST Chemistry WebBook, SRD 69: Water, condensed phase thermochemistry (liquid-phase Shomate equation, Chase 1998) (reference data, read 2026-10-13)nvlink-page: NVIDIA NVLink and NVLink Switch page, Specifications ('Preliminary specifications; may be subject to change') (vendor product page, read 2026-10-13)nvshmem-2026: Ma et al., 'Demystifying NVSHMEM: A System-Level Analysis on Symmetric Memory and Device-Initiated Operations in GPU Communication', arXiv:2606.05951, section VII-C and Fig. 5 (8x H200 SXM5 node, NVLink 4) (paper (measurements), read 2026-10-16)patarasuk2009: Patarasuk and Yuan, Bandwidth optimal all-reduce algorithms for clusters of workstations, JPDC 2009 (doi:10.1016/j.jpdc.2008.09.002) (paper, read 2026-10-13)photonics-page: NVIDIA Silicon Photonics page (co-packaged optics switches) (vendor product page, read 2026-10-13)quantum2-page: NVIDIA Quantum-2 InfiniBand platform page (switches and ConnectX-7 adapters) (vendor product page, read 2026-10-13)qx800-page: NVIDIA Quantum-X800 InfiniBand platform page (vendor product page, read 2026-10-13)smooth-blog: How New GB300 NVL72 Features Provide Steady Power for AI (NVIDIA Technical Blog) (vendor technical blog, read 2026-10-13)splitwise: Patel et al., Splitwise: Efficient generative LLM inference using phase splitting (arXiv 2311.18677) (paper, read 2026-10-13)supernic-page: NVIDIA Ethernet SuperNIC page (ConnectX-8) (vendor product page, read 2026-10-13)superpod-components: NVIDIA DGX SuperPOD Reference Architecture (DGX GB200), Key Components (vendor reference architecture, read 2026-10-13)superpod-network: NVIDIA DGX SuperPOD Reference Architecture (DGX GB200), Network Fabrics (Table 3) (vendor reference architecture, read 2026-10-13)tpu-7x: Google Cloud documentation, TPU7x (Ironwood) (system architecture) (vendor documentation, read 2026-10-13)tpu-v4: Jouppi et al., TPU v4: An Optically Reconfigurable Supercomputer for Machine Learning with Hardware Support for Embeddings (arXiv 2304.01433) (paper, read 2026-10-13)tpu-v5p: Google Cloud documentation, TPU v5p (system architecture and configurations) (vendor documentation, read 2026-10-13)trainium-page: AWS Trainium (product page; Trainium3 UltraServers) (vendor product page, read 2026-10-13)trn2-page: Amazon EC2 Trn2 instances and UltraServers (product page) (vendor product page, read 2026-10-13)vr-page: NVIDIA Vera Rubin NVL72 product page, Specifications (marked 'Available Now' on the access date) (vendor product page, read 2026-10-13)water-blog: Chill Factor: NVIDIA Blackwell Platform Boosts Water Efficiency by Over 300x (NVIDIA Blog) (vendor blog, read 2026-10-13)young1974: Young, A first order approximation to the optimum checkpoint interval, Communications of the ACM 1974 (doi:10.1145/361147.361115) (paper, read 2026-10-13)