GPU comparisons
Compare the hardware.
Side-by-side comparisons of NVIDIA GPUs for AI inference, fine-tuning and image generation, with specs, memory and live rental prices.
A100 80GB vs RTX 5090 for AI
A100 vs 5090 for AI: rent 80 GB of HBM2e by the hour or buy a 32 GB card. Specs, what fits in each, and the break-even at the $1,999 launch price.
A100 vs H200 NVL: is the upgrade worth it?
A100 80GB vs H200 NVL: 80 against 141 GB, 2.0 against 4.8 TB/s, no FP8 against FP8. When the upgrade pays for itself, and when the A100 still wins.
B200 vs H100: what Blackwell changes
B200 vs H100 on NVIDIA's figures: 180 vs 80 GB, 2.3x the dense FP8 and BF16, FP4, NVLink 5 and power, and the H100 you can rent by the hour today.
B300 vs B200: what Blackwell Ultra changes
B300 vs B200 on NVIDIA's figures: 270 vs 180 GB, dense FP4, FP8, FP64, power and CUDA needs, and which to build for inference or training.
H100 PCIe vs RTX 5090 for AI
H100 vs 5090 for AI: close on memory bandwidth, far apart on memory and FP8. When the RTX 5090 wins, when only an H100 works, and what renting one costs.
H200 vs B200 for training and inference
H200 vs B200 on NVIDIA's figures: 141 vs 180 GB, FP8, FP4, NVLink 4 vs 5 and power, and when to start on an H200 NVL while a B200 build is quoted.
L40 vs L40S: what the extra S buys you
L40 vs L40S: same 48 GB and 864 GB/s, but NVIDIA rates the L40S at twice the Tensor Core throughput. Specs, live prices and which to rent.
L40S vs A100: specs, speed and price per hour
NVIDIA L40S vs A100 80GB: FP8 against none, 48 GB against 80 GB, 864 against 2,039 GB/s. Live prices and a verdict for each workload.
L40S vs H100 PCIe: which to rent for inference and fine-tuning
NVIDIA L40S vs H100 PCIe: 48 against 80 GB, 864 against 2,000 GB/s, FP8 on both. Where 48 GB is enough, where it is not, and live hourly prices.
L40S vs RTX 6000 Ada: a data-center card and a workstation card
L40S vs RTX 6000 Ada: the same core counts and 48 GB, FP8 within 1%, but 11% more bandwidth on the RTX 6000 Ada. Specs, live prices and which to rent.
RTX 6000 Ada vs RTX 4090: one AD102 chip, two cards
RTX 6000 Ada vs RTX 4090: one AD102 chip, 48 GB against 24, and full-rate FP32 accumulate against half. Specs, what fits, and renting vs buying.
RTX 6000 Ada vs RTX 5090 for AI
RTX 6000 Ada vs RTX 5090 for AI: both run FP8, but one has 48 GB at 300 W and the other 1.9x the bandwidth and FP4. Specs, fit and buy vs rent.
RTX 6000 Ada vs RTX A6000: FP8 and bandwidth at the same 48 GB
RTX 6000 Ada vs RTX A6000: the same 48 GB, but the Ada card adds FP8, 2.35x the Tensor math and 25% more bandwidth. Live prices and when each wins.
RTX A6000 vs A100: 48 GB workstation card vs 80 GB datacenter GPU
RTX A6000 vs A100 80GB: specs, live hourly prices, what fits in 48 GB against 80 GB, and when the cheaper A6000 does the same job for less.
RTX A6000 vs RTX 4090 for AI: rent 48 GB or buy 24 GB
A6000 vs 4090 for AI: twice the memory against 31% more bandwidth and FP8. What 48 GB fits that 24 GB cannot, and renting vs buying a 4090 now.
RTX A6000 vs RTX 5090: what 48 GB fits that 32 GB does not
A6000 vs 5090 for AI: the RTX 5090 is faster, the RTX A6000 holds 16 GB more. What fits in 48 GB but not 32 GB, and when renting beats buying.
RTX PRO 6000 Blackwell vs A100 80GB
RTX PRO 6000 Blackwell vs A100 80GB: 96 GB with FP8 and FP4 against 80 GB of faster HBM2e. Which to rent for BF16, FP8, FP4 and multi-GPU jobs.
RTX PRO 6000 Blackwell vs H100 PCIe: which is better value for inference
H100 PCIe vs RTX PRO 6000 Blackwell for LLM inference: 80 GB of faster HBM2e against 96 GB with FP4, at a similar hourly price. Tokens per dollar on paper.
RTX PRO 6000 Blackwell vs H200 NVL for large-model inference
RTX PRO 6000 Blackwell vs H200 NVL: 96 GB of GDDR7 against 141 GB of HBM3e at 4.8 TB/s. Cost per token, what fits, and when the cheaper hour wins.
RTX PRO 6000 Blackwell vs RTX 5090: one GB202 chip, 96 GB against 32
RTX PRO 6000 vs RTX 5090: the same GB202 chip with 96 GB against 32 and full-rate FP32 accumulate. What 96 GB fits, and when renting beats buying.