BTC
ETH
USDT
XMR
LTC
GPU comparison · Blackwell vs Ada Lovelace
NVIDIA RTX 5090 vs L40S.
A Blackwell GeForce card against the Ada data-center GPU. The L40S has 48 GB of GDDR6 with ECC, about 1.75× the dense FP8 throughput and servers of up to 8 GPUs. The RTX 5090 has 32 GB of GDDR7, about twice the memory bandwidth (1.79 against 0.86 TB/s), FP4 tensor cores and costs $329 against $789 per GPU per month.
- 32 vs 48 GBGPU memory
- 1.79 vs 0.86 TB/smemory bandwidth
- $329 vs $789per GPU, per month
RTX 5090 vs L40S: which one to rent.
Choose the RTX 5090 if
- Your model fits in 32 GB per card: up to about 24B parameters in FP8 or 45B in 4-bit by our sizing rule.
- Your workload is bound by memory bandwidth rather than tensor throughput, such as single-user token generation on a model that fits.
- You want the lower price per GB of GPU memory: $10.28 against $16.44 a month.
Choose the L40S if
- Your model needs more than 32 GB per GPU: 48 GB holds about 36B parameters in FP8 or 67B in 4-bit by our sizing rule.
- You need ECC memory or a server of 8 GPUs: GeForce servers stop at 4 cards.
- You serve in FP8 at volume: 733 TFLOPS of dense FP8 tensor throughput against 419 on the RTX 5090.
RTX 5090 vs L40S specs, side by side.
Figures from NVIDIA’s published specifications, per GPU. Tensor figures are peak dense throughput.
| Per GPU | NVIDIA RTX 5090 | NVIDIA L40S | L40S / RTX 5090 |
|---|---|---|---|
| Architecture | Blackwell | Ada Lovelace | — |
| Form factor | PCIe Gen5 graphics card | Dual-slot PCIe card, passive | — |
| GPU memory | 32 GB GDDR7, 512-bit | 48 GB GDDR6 with ECC | 1.5× |
| Memory bandwidth | 1,792 GB/s | 864 GB/s | 0.48× |
| GPU-to-GPU link | None (PCIe Gen5) | None (PCIe Gen4 x16) | — |
| FP4 tensor (dense) | 1.7 PFLOPS | Not supported | — |
| FP8 tensor (dense) | 419 TFLOPS | 733 TFLOPS | 1.75× |
| FP16 / BF16 tensor (dense) | 209.5 TFLOPS | 362 TFLOPS | 1.73× |
| FP64 | — | — | — |
| Max power | 575 W | 350 W | — |
| GPUs per server | 1, 2, 4 | 1, 2, 4, 8 | — |
| Per GPU in our servers | 16 vCPU · 64 GB RAM · 1 TB | 12 vCPU · 96 GB RAM · 1 TB | — |
Sources: NVIDIA GeForce RTX 5090 · RTX Blackwell whitepaper · NVIDIA L40S. Full sheets: RTX 5090 · L40S
RTX 5090 vs L40S rental price.
At GPUCenter, the RTX 5090 rents for $329 and the L40S for $789 per GPU per month, 30% below the market median of public on-demand prices, paid in crypto with no KYC. Hourly figures are the monthly price divided by 730 hours.
| GPUCenter | NVIDIA RTX 5090 | NVIDIA L40S | L40S / RTX 5090 |
|---|---|---|---|
| Price per GPU, per month | $329 | $789 | 2.4× |
| Equivalent per GPU-hour | $0.45 | $1.08 | — |
| Per GB of GPU memory, per month | $10.28 | $16.44 | 1.6× |
| Per PFLOPS of dense FP8, per month | $785 | $1,076 | 1.37× |
| Largest server | 4× · $1,316/mo | 8× · $6,312/mo | — |
| Market median, per GPU | $481 | $1,139 | — |
| Below the median | −32% | −31% | — |
Same price per GPU at every server size, no hourly metering. How we compare prices · Configure an RTX 5090 server · Configure a L40S server
Which models fit on each.
Smallest server that holds each model, by precision. Rule of thumb with headroom for the KV cache, 92% of GPU memory usable.
| Model (total parameters) | RTX 5090 · FP8 | L40S · FP8 | RTX 5090 · 4-bit | L40S · 4-bit |
|---|---|---|---|---|
| Qwen3.5-9B | 1× | 1× | 1× | 1× |
| Gemma 4 31B | 2× | 1× | 1× | 1× |
| Llama 3.3 70B | 4× | 2× | 2× | 2× |
| gpt-oss-120b | — | 4× | 4× | 2× |
| DeepSeek-V4-Flash | — | 8× | — | 8× |
| Qwen3.5-397B-A17B | — | — | — | 8× |
| DeepSeek-R1 (671B) | — | — | — | — |
| Kimi K2.6 (1T) | — | — | — | — |
FP8 ≈ 1.2 bytes and 4-bit ≈ 0.65 bytes per parameter. — = larger than the biggest server of that GPU. Size another model · How much VRAM does an LLM need?
RTX 5090 vs L40S FAQ.
More on each GPU: NVIDIA RTX 5090 · NVIDIA L40S.
Is the RTX 5090 faster than the L40S?
It depends on what limits the job. The RTX 5090 has about twice the memory bandwidth, which sets the pace of token generation for a model that fits. The L40S has about 1.75× the dense FP8 tensor throughput and 16 GB more memory. We do not publish benchmarks.
Why does the L40S cost more than the RTX 5090?
It is a data-center card: 48 GB of memory with ECC, passive cooling for dense servers and servers of up to 8 GPUs. We price both 30% below their own market median.
Which one fits a 32B model?
In 4-bit (about 21 GB), both on a single card. In FP8 (about 38 GB), one L40S holds it; on RTX 5090 it takes two cards.
Other GPU comparisons.
- Blackwell vs Hopper RTX 5090 vs H100 32 vs 80 GB · 1.8 TB/s vs 3.35 TB/s $329 vs $1,779/mo per GPU
- Blackwell vs Ada Lovelace RTX 5090 vs RTX 4090 32 vs 24 GB · 1.8 TB/s vs 1 TB/s $329 vs $269/mo per GPU
- Blackwell vs Blackwell RTX PRO 6000 vs RTX 5090 96 vs 32 GB · 1.6 TB/s vs 1.8 TB/s $959 vs $329/mo per GPU
Rent the RTX 5090 or the L40S.
Dedicated servers with 1 to 8 GPUs, one monthly price, paid in crypto. Online in under 10 minutes, no KYC.