Compare GPUs
NVIDIA GPU comparison: specs and monthly prices, side by side
Ten GPUs from four architectures — Ampere, Ada Lovelace, Hopper and Blackwell — with NVIDIA's published figures, dense tensor throughput, and what each one costs here as a dedicated server for a month. Then 15 head-to-head comparisons for the choices people actually hesitate between.
The chart
Every GPU we rent, one table
Sorted from the largest memory to the smallest. Click a GPU for its page — full specifications, price by term and a market comparison.
| GPU | Memory | Bandwidth | FP16 dense | FP8 dense | FP4 dense | Power | GPU link | Per month | Per GB |
|---|---|---|---|---|---|---|---|---|---|
| NVIDIA B300Blackwell Ultra · 2025 | 288 GB HBM3e | 8 TB/s | ~2,250 TFLOPS | ~4,500 TFLOPS | ~14,000 TFLOPS | Up to 1,100 W | NVLink 5 · 1.8 TB/s | $3,790 | $13.16 |
| NVIDIA B200Blackwell · 2024 | 180 GB HBM3e | 8 TB/s | ~2,250 TFLOPS | ~4,500 TFLOPS | ~9,000 TFLOPS | Up to 1,000 W | NVLink 5 · 1.8 TB/s | $2,690 | $14.94 |
| NVIDIA H200Hopper · 2024 | 141 GB HBM3e | 4.8 TB/s | ~989 TFLOPS | ~1,979 TFLOPS | Not supported | Up to 700 W | NVLink 4 · 900 GB/s | $1,690 | $11.99 |
| NVIDIA RTX PRO 6000 BlackwellBlackwell · 2025 | 96 GB GDDR7 ECC | 1.6 TB/s | ~500 TFLOPS | ~1,000 TFLOPS | ~2,000 TFLOPS | Up to 600 W | PCIe only | $749 | $7.80 |
| NVIDIA H100 SXMHopper · 2022 | 80 GB HBM3 | 3.35 TB/s | ~989 TFLOPS | ~1,979 TFLOPS | Not supported | Up to 700 W | NVLink 4 · 900 GB/s | $1,249 | $15.61 |
| NVIDIA A100Ampere · 2020 | 80 GB HBM2e | 2 TB/s | 312 TFLOPS | Not supported | Not supported | 400 W | NVLink 3 · 600 GB/s | $649 | $8.11 |
| NVIDIA L40SAda Lovelace · 2023 | 48 GB GDDR6 | 864 GB/s | 362 TFLOPS | 733 TFLOPS | Not supported | 350 W | PCIe only | $429 | $8.94 |
| NVIDIA RTX 5090Blackwell · 2025 | 32 GB GDDR7 | 1.79 TB/s | 210 TFLOPS (419 with FP16 accumulate) | 419 TFLOPS (838 with FP16 accumulate) | 1,676 TFLOPS | 575 W | PCIe only | $279 | $8.72 |
| NVIDIA RTX 4090Ada Lovelace · 2022 | 24 GB GDDR6X | 1 TB/s | 165 TFLOPS (330 with FP16 accumulate) | 330 TFLOPS (661 with FP16 accumulate) | Not supported | 450 W | PCIe only | $189 | $7.88 |
| NVIDIA L4Ada Lovelace · 2023 | 24 GB GDDR6 | 300 GB/s | 121 TFLOPS | 242 TFLOPS | Not supported | 72 W | PCIe only | $169 | $7.04 |
Figures from NVIDIA datasheets and product pages; “~” marks values NVIDIA publishes per system or with sparsity, divided down to one GPU, dense. Prices are this site's flat monthly rates for a dedicated server — CPU, RAM, NVMe and 20 TB of transfer included — in USD excluding VAT.
Head to head
The comparisons people ask for
Each one has the specifications side by side, what each dollar buys, the model sizes that fit, and a straight answer on which to rent.
- H100 vs A10080 GB HBM3 · 80 GB HBM2e Same 80 GB of memory, two generations apart. The H100 SXM brings FP8, 1.6× the memory bandwidth and about 3× the dense tensor throughput; the A100 80 GB costs about half as much per month. $1,249 vs $649 a monthCompare
- H200 vs H100141 GB HBM3e · 80 GB HBM3 The same Hopper compute with more and faster memory: 141 GB of HBM3e at 4.8 TB/s on the H200 against 80 GB of HBM3 at 3.35 TB/s on the H100 SXM — for 1.4× the monthly price. $1,690 vs $1,249 a monthCompare
- B200 vs H100180 GB HBM3e · 80 GB HBM3 Blackwell against Hopper: 180 GB of HBM3e at 8 TB/s, about 2.3× the dense tensor throughput and the new FP4 format on the B200, against 80 GB at 3.35 TB/s on the H100 SXM — for 2.2× the monthly price. $2,690 vs $1,249 a monthCompare
- B200 vs H200180 GB HBM3e · 141 GB HBM3e Both answer the memory problem of the H100. The H200 does it with 141 GB on the same Hopper compute; the B200 adds Blackwell's throughput, 180 GB and 8 TB/s — for 1.6× the monthly price. $2,690 vs $1,690 a monthCompare
- B300 vs B200288 GB HBM3e · 180 GB HBM3e Blackwell Ultra against Blackwell: 288 GB of HBM3e on the B300 against 180 GB on the B200, the same 8 TB/s, and about 1.5× the dense FP4 throughput — for 1.4× the monthly price. $3,790 vs $2,690 a monthCompare
- RTX 5090 vs RTX 409032 GB GDDR7 · 24 GB GDDR6X Blackwell against Ada: 32 GB of GDDR7 at 1.79 TB/s and FP4 on the RTX 5090, against 24 GB of GDDR6X at 1 TB/s on the RTX 4090 — for 1.5× the monthly price. $279 vs $189 a monthCompare
- L40S vs A10048 GB GDDR6 · 80 GB HBM2e Two answers to the same budget: 48 GB of GDDR6 with FP8, RT cores and video engines on the L40S, against 80 GB of HBM2e, NVLink and MIG on the A100 80 GB. $429 vs $649 a monthCompare
- L40S vs H10048 GB GDDR6 · 80 GB HBM3 The inference workhorse against the training flagship: 48 GB of GDDR6 for $429 a month, or 80 GB of HBM3 with 3.9× the bandwidth for $1,249. $429 vs $1,249 a monthCompare
- L4 vs L40S24 GB GDDR6 · 48 GB GDDR6 Two Ada Lovelace data center GPUs built for inference. The L4 is the small, efficient one — 24 GB in 72 W; the L40S doubles the memory and brings nearly 3× the bandwidth and tensor throughput, for 2.5× the monthly price. $169 vs $429 a monthCompare
- RTX 4090 vs A10024 GB GDDR6X · 80 GB HBM2e A consumer flagship against a data center classic: 24 GB of GDDR6X and FP8 on the RTX 4090 for $189 a month, or 80 GB of HBM2e with ECC and NVLink on the A100 for $649. $189 vs $649 a monthCompare
- RTX 4090 vs H10024 GB GDDR6X · 80 GB HBM3 $189 against $1,249 a month: the H100 SXM has 3.3× the memory, 3.3× the bandwidth and NVLink; the RTX 4090 has the lowest price for a GPU this fast. $189 vs $1,249 a monthCompare
- RTX 5090 vs H10032 GB GDDR7 · 80 GB HBM3 Blackwell's consumer flagship against Hopper's data center flagship: 32 GB at 1.79 TB/s for $279 a month, or 80 GB at 3.35 TB/s with NVLink for $1,249. $279 vs $1,249 a monthCompare
- RTX PRO 6000 vs RTX 509096 GB GDDR7 ECC · 32 GB GDDR7 The same Blackwell generation, two very different memory sizes: 96 GB of GDDR7 with ECC on the RTX PRO 6000, 32 GB on the RTX 5090 — for 2.7× the monthly price. $749 vs $279 a monthCompare
- RTX PRO 6000 vs H10096 GB GDDR7 ECC · 80 GB HBM3 96 GB of GDDR7 for $749 a month, or 80 GB of HBM3 for $1,249: the RTX PRO 6000 Blackwell holds more for less; the H100 SXM moves its memory twice as fast. $749 vs $1,249 a monthCompare
- RTX PRO 6000 vs L40S96 GB GDDR7 ECC · 48 GB GDDR6 Workstation-class Blackwell against the Ada data center all-rounder: 96 GB at about 1.6 TB/s with FP4, or 48 GB at 864 GB/s with FP8 — for 1.7× the monthly price. $749 vs $429 a monthCompare
Choosing a GPU in four questions
- How much memory does the job need?Weights, KV cache, optimizer state and activations must fit, or the job does not run. Start with the GPU memory calculator.
- Is it bound by bandwidth or by compute?Token generation is limited by memory bandwidth; training and prompt processing by tensor throughput. HBM parts (H100, H200, B200, B300, A100) win the first.
- Which number formats will you use?FP8 needs Ada, Hopper or Blackwell; FP4 needs Blackwell. The A100 runs BF16, FP16, TF32 and INT8.
- One GPU or many?Tensor parallelism across GPUs wants NVLink (the SXM parts in 8-GPU nodes); independent replicas run fine on PCIe cards.
Recommended starting points
- Training: H100 SXM or an 8-GPU node; B200 for the largest runs.
- Fine-tuning: A100 80 GB on a budget, H200 for 30B-class full fine-tunes.
- Inference: L40S up to 30B at FP8, RTX PRO 6000 for 70B on one card.
- Image and video: RTX 4090 for value, RTX 5090 for speed.
GPU comparison: questions answered
Which NVIDIA GPU is best for LLM inference?
It depends on the model size. Up to 8B parameters, the L4 (24 GB, $169 a month) serves at FP8; up to about 30B, the L40S (48 GB, $429); 70B-class models fit on one RTX PRO 6000 (96 GB, $749) or, for high concurrency, one H200 (141 GB, $1,690). Beyond that, B200 and B300 nodes. Memory decides first; bandwidth sets the speed.
Which GPU is best for training?
For most teams, the H100 SXM ($1,249 a month) in an 8-GPU NVLink node: FP8, 80 GB of HBM3 and the most mature software. The B200 ($2,690) does about 2.3× the work per GPU for frontier-scale runs; the A100 remains the budget choice for fine-tuning that does not use FP8.
What is the cheapest GPU that runs a 70B model?
On a single card, the RTX PRO 6000 Blackwell: its 96 GB hold a 70B model at FP8 (about 70 GB of weights) with room for the KV cache, for $749 a month. At 4-bit, a 70B model (about 39 GB) also fits on one L40S or one A100.
What is the difference between data center and workstation-class GPUs?
Data center GPUs (H100, H200, B200, B300, A100, L40S, L4) are built for 24/7 duty with ECC memory, and the SXM parts add NVLink and HBM for multi-GPU work. Workstation-class GPUs (RTX PRO 6000, RTX 5090, RTX 4090) are single cards with ray-tracing cores and Windows support here — more speed per dollar on jobs that fit their memory.
Why do you quote dense rather than sparse TFLOPS?
Because dense throughput is what a normal model uses. NVIDIA often prints tensor figures “with sparsity”, which assumes 2:4 structured sparsity in the weights and doubles the number; few production models run that way. Every figure on these pages is the dense one.
Every GPU in this chart is in stock
Dedicated, single-tenant, billed by the month and live in under 10 minutes — paid in BTC, ETH or USDT, with no KYC.