GPU Comparisons
Specs, prices and residual value for data center GPUs, side by side.
Every GPU on one table
| GPU | Memory | Bandwidth | FP8 TFLOPS | FP4 TFLOPS | Power | Price per GPU | Residual 2027 | Residual 2029 |
|---|---|---|---|---|---|---|---|---|
| A100Ampere, 2020 | 80 GBHBM2e | 2 TB/s | - | - | 400 W | $5,400 (Used, typical realized sale) | - | - |
| H100Hopper, Late 2022 | 80 GBHBM3 | 3.35 TB/s | 1,979 | - | 700 W | $14,250 (Used, typical realized sale) | 11-19% | 2-3% |
| H200Hopper, Mid 2024 | 141 GBHBM3e | 4.8 TB/s | 1,979 | - | 700 W | $43,600 (Used 8-GPU node, per GPU) | 34-46% | 7-15% |
| MI300XCDNA 3, December 2023 | 192 GBHBM3 | 5.3 TB/s | 2,614.9 | - | 750 W | - | - | - |
| RTX PRO 6000Blackwell, March 2025 (announced) | 96 GBGDDR7 | 1.6 TB/s | - | - | 600 W | $12,500-$18,750 (New 8-GPU server, per GPU) | - | - |
| B200Blackwell, Early 2025 | 180 GBHBM3e | 8 TB/s | 4,500 | 9,000 | 1,000 W | $45,000-$58,750 (New 8-GPU node, per GPU) | 40-54% | 16-26% |
| GB200Blackwell, Early 2025 | 186 GBHBM3e | 8 TB/s | 5,000 | 10,000 | - | - | - | - |
| B300Blackwell Ultra, Late 2025 | 288 GBHBM3e | 8 TB/s | 4,500 | 13,500 | 1,400 W | $50,000-$65,000 (New 8-GPU node, per GPU) | 52-64% | 28-39% |
| GB300Blackwell Ultra, Late 2025 | 278 GBHBM3e | 8 TB/s | 5,000 | 15,000 | - | $111,000-$119,000 (Rack ballpark, per GPU) | - | - |
Dense Tensor Core TFLOPS. RTX PRO 6000 FP8 and FP4 not shown: NVIDIA does not say whether its figures are sparse. Residuals are our conservative band: what a used unit is expected to fetch, as a share of its original price. Each GPU links to its appraisal.
Sources: NVIDIA, American Compute, AMD, Managed GPUs (American Compute).
Comparisons
| A100 vs H100 | 3.2x the compute for 2.6x the used price; their resale prices are moving apart. |
| H200 vs H100 | Same compute, 76% more memory. Per GB of HBM, used nodes cost the same. |
| B200 vs H100 | A full generation apart. Per FP8 FLOP, a new B200 costs about the same as a used H100. |
| B200 vs H200 | For a 6-38% higher node price, the B200 brings 2.3x the FP8 and holds value better. |
| B300 vs B200 | More memory and FP4, no FP64. The B300 holds a higher residual band. |
| GB200 NVL72 vs HGX B200 | Same GPU, different system: a 72-GPU NVLink rack against an 8-GPU server. |
| GB300 NVL72 vs HGX B300 | Rack-scale costs roughly 1.5-2.1x per GPU; HGX goes into more data centers. |
| MI300X vs H100 | More memory and FLOPS per dollar on AMD; CUDA and resale comps on NVIDIA. |
| RTX PRO 6000 vs H100 | A PCIe inference card against an NVLink training GPU; half the server price. |
Discuss a transaction
Send us the equipment list and the loan or lease you have in mind. We will come back with an indication and the questions we would need answered.
Prefer email? hello@amcompute.com