GPUs
size.py reads memory and power from this table. A GPU that is not here can still be used: the interview asks for its memory, and the design carries the value as an estimate.
| Model | Vendor | Memory (GB) | Power (W) | Evidence | Last verified |
|---|---|---|---|---|---|
| H200 | nvidia | 141 | 700 | docs |
2026-11-22 |
| H100 SXM | nvidia | 80 | 700 | docs |
2026-11-22 |
| H100 NVL | nvidia | 94 | 400 | docs |
2026-11-22 |
| A100 40GB | nvidia | 40 | 250 | docs |
2026-11-22 |
| A100 80GB | nvidia | 80 | 500 | docs |
2026-11-22 |
| L40S | nvidia | 48 | 350 | docs |
2026-11-22 |
| L4 | nvidia | 24 | 72 | docs |
2026-11-22 |
| RTX 6000 Ada | nvidia | 48 | 300 | docs |
2026-11-22 |
| RTX A6000 | nvidia | 48 | 300 | docs |
2026-11-22 |
| RTX 5090 | nvidia | 32 | 575 | docs |
2026-11-22 |
| RTX 4090 | nvidia | 24 | 450 | docs |
2026-11-22 |
| B200 | nvidia | 180 | 1000 | docs |
2026-11-22 |
| B300 | nvidia | 288 | unknown | docs |
2026-11-22 |
| RTX PRO 6000 Blackwell | nvidia | 96 | 600 | docs |
2026-11-22 |
| MI300X | amd | 192 | 750 | docs |
2026-11-22 |
| MI325X | amd | 256 | 1000 | docs |
2026-11-22 |
| MI355X | amd | 288 | 1400 | docs |
2026-11-22 |
Notes and sources
Section titled “Notes and sources”Also written as NVIDIA H200, H200 SXM, H200 NVL.
The page lists 141GB for both form factors; power is 700W (configurable) for SXM and 600W (configurable) for NVL, so the highest is used. The page marks these as preliminary specifications.
H100 SXM
Section titled “H100 SXM”Also written as NVIDIA H100 SXM.
Up to 700W (configurable). The H100 NVL has a different memory size and its own row; there is no bare H100 name.
H100 NVL
Section titled “H100 NVL”Also written as NVIDIA H100 NVL.
94GB per card (product brief, memory size 94 GB); the product page lists 350-400W (configurable) and the brief names 400W as the maximum TDP. The page text also mentions 188GB for the NVLink-bridged setup; sizing needs the per-card figure.
- https://www.nvidia.com/en-us/data-center/h100/
- https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/h100/PB-11773-001_v01.pdf
A100 40GB
Section titled “A100 40GB”Also written as NVIDIA A100 40GB, A100 40GB PCIe.
A100 40GB PCIe product brief: memory size 40 GB HBM2, TDP level 250 W. There is no bare A100 name. The row describes the PCIe card: the vendor page used for it gives no 40 GB SXM figures.
A100 80GB
Section titled “A100 80GB”Also written as NVIDIA A100 80GB, A100 80GB PCIe, A100 80GB SXM.
80GB HBM2e for PCIe and SXM. Max TDP is 300W (PCIe) and 400W (SXM, standard configuration); the page footnote says the HGX A100-80GB custom thermal solution SKU supports up to 500W, and the highest figure is used.
Also written as NVIDIA L40S.
48GB GDDR6 with ECC; max power consumption 350W.
Also written as NVIDIA L4.
24GB; max thermal design power 72W.
RTX 6000 Ada
Section titled “RTX 6000 Ada”Also written as NVIDIA RTX 6000 Ada Generation, RTX 6000 Ada Generation.
48GB GDDR6 with ECC; max power consumption 300W.
RTX A6000
Section titled “RTX A6000”Also written as NVIDIA RTX A6000.
48GB GDDR6 with ECC; max power consumption 300W.
RTX 5090
Section titled “RTX 5090”Also written as GeForce RTX 5090, NVIDIA GeForce RTX 5090.
32GB GDDR7; the page states total graphics power 575W, used as tdp_w.
RTX 4090
Section titled “RTX 4090”Also written as GeForce RTX 4090, NVIDIA GeForce RTX 4090.
24GB GDDR6X; the page states total graphics power 450W, used as tdp_w.
Also written as NVIDIA B200.
HGX B200 reference table: memory per GPU 180GB HBM3e. HGX B200 summary: each B200 GPU has 180 GB, max thermal design power per individual GPU configurable up to 1000 W. NVIDIA’s Blackwell Ultra developer blog table lists 192 GB as the Blackwell chip maximum; the product figure is used here.
- https://docs.nvidia.com/enterprise-reference-architectures/hgx-ai-factory/latest/components.html
- https://images.nvidia.com/aem-dam/Solutions/documents/HGX-B200-PCF-Summary.pdf
Also written as NVIDIA B300.
NVIDIA states 288 GB per GPU in three places: the developer blog ‘Inside NVIDIA Blackwell Ultra’ (‘With 288 GB of HBM3e per GPU’), the HGX reference table (memory per GPU 288GB HBM3e; Table 2: up to 2304 GB per eight-GPU baseboard) and the DGX B300 user guide (8 x 288 GB = 2.3 TB). The blog footnote says ‘Blackwell Ultra GPUs contain up to 160 SMs and 288GB HBM3E Memory. Available SM count and HBM capacity varies by SKU’, so 288 is a maximum: confirm the part. The HGX and DGX B300 product pages state only 2.1 TB total. 8 x 288 GB = 2,304 GB, which is about 2.1 TiB, so that figure may be in binary units; no NVIDIA page confirms this reading. No per-GPU power figure found on a product or documentation page, so tdp_w is left out.
- https://docs.nvidia.com/enterprise-reference-architectures/hgx-ai-factory/latest/components.html
- https://docs.nvidia.com/dgx/dgxb300-user-guide/introduction-to-dgxb300.html
- https://developer.nvidia.com/blog/inside-nvidia-blackwell-ultra-the-chip-powering-the-ai-factory-era/
RTX PRO 6000 Blackwell
Section titled “RTX PRO 6000 Blackwell”Also written as NVIDIA RTX PRO 6000 Blackwell, RTX PRO 6000 Blackwell Server Edition, RTX PRO 6000 Blackwell Workstation Edition.
Server Edition page: 96GB GDDR7, up to 600W (configurable). Workstation Edition datasheet: 96 GB GDDR7 and a 600W power design. Both editions share one row.
- https://www.nvidia.com/en-us/data-center/rtx-pro-6000-blackwell-server-edition/
- https://www.nvidia.com/content/dam/en-zz/Solutions/design-visualization/quadro-product-literature/workstation-datasheet-blackwell-rtx-pro6000-x-nvidia-us-3519208-web.pdf
MI300X
Section titled “MI300X”Also written as AMD Instinct MI300X, Instinct MI300X.
Dedicated memory size 192 GB HBM3; typical board power (TBP) 750W peak.
MI325X
Section titled “MI325X”Also written as AMD Instinct MI325X, Instinct MI325X.
Dedicated memory size 256 GB HBM3E; typical board power (TBP) 1000W peak.
MI355X
Section titled “MI355X”Also written as AMD Instinct MI355X, Instinct MI355X.
Dedicated memory size 288 GB HBM3E; typical board power (TBP) 1400W.