BTC
ETH
USDT
XMR
LTC
GPU comparison · Ada Lovelace vs Ada Lovelace
NVIDIA L40S vs RTX 4090.
Two Ada Lovelace GPUs. The L40S is the data-center card: 48 GB of GDDR6 with ECC, about 2.2× the dense FP8 throughput and servers of up to 8 GPUs. The GeForce RTX 4090 has 24 GB, slightly more memory bandwidth (1.01 against 0.86 TB/s) and costs about a third as much: $269 against $789 per GPU per month.
- 48 vs 24 GBGPU memory
- 0.86 vs 1.01 TB/smemory bandwidth
- $789 vs $269per GPU, per month
L40S vs RTX 4090: which one to rent.
Choose the L40S if
- Your model needs more than 24 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 330 on the RTX 4090.
Choose the RTX 4090 if
- Your model fits in 24 GB: about 18B parameters in FP8 or 34B in 4-bit per card.
- You want the same 48 GB for less: two RTX 4090 cost $538 a month against $789 for one L40S, with the memory split over two cards.
- Your workload is bound by memory bandwidth rather than tensor throughput, such as single-user token generation on a model that fits.
L40S vs RTX 4090 specs, side by side.
Figures from NVIDIA’s published specifications, per GPU. Tensor figures are peak dense throughput.
| Per GPU | NVIDIA L40S | NVIDIA RTX 4090 | RTX 4090 / L40S |
|---|---|---|---|
| Architecture | Ada Lovelace | Ada Lovelace, TSMC 4N | — |
| Form factor | Dual-slot PCIe card, passive | PCIe Gen4 graphics card | — |
| GPU memory | 48 GB GDDR6 with ECC | 24 GB GDDR6X, 384-bit | 0.5× |
| Memory bandwidth | 864 GB/s | 1,008 GB/s | 1.17× |
| GPU-to-GPU link | None (PCIe Gen4 x16) | None (PCIe Gen4) | — |
| FP4 tensor (dense) | Not supported | Not supported | — |
| FP8 tensor (dense) | 733 TFLOPS | 330 TFLOPS | 0.45× |
| FP16 / BF16 tensor (dense) | 362 TFLOPS | 165 TFLOPS | 0.46× |
| FP64 | — | — | — |
| Max power | 350 W | 450 W | — |
| GPUs per server | 1, 2, 4, 8 | 1, 2, 4 | — |
| Per GPU in our servers | 12 vCPU · 96 GB RAM · 1 TB | 12 vCPU · 64 GB RAM · 1 TB | — |
Sources: NVIDIA L40S · NVIDIA GeForce RTX 4090 · Ada whitepaper. Full sheets: L40S · RTX 4090
L40S vs RTX 4090 rental price.
At CryptGPU, the L40S rents for $789 and the RTX 4090 for $269 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.
| CryptGPU | NVIDIA L40S | NVIDIA RTX 4090 | RTX 4090 / L40S |
|---|---|---|---|
| Price per GPU, per month | $789 | $269 | 0.34× |
| Equivalent per GPU-hour | $1.08 | $0.37 | — |
| Per GB of GPU memory, per month | $16.44 | $11.21 | 0.68× |
| Per PFLOPS of dense FP8, per month | $1,076 | $815 | 0.76× |
| Largest server | 8× · $6,312/mo | 4× · $1,076/mo | — |
| Market median, per GPU | $1,139 | $387 | — |
| Below the median | −31% | −30% | — |
Same price per GPU at every server size, no hourly metering. How we compare prices · Configure an L40S server · Configure a RTX 4090 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) | L40S · FP8 | RTX 4090 · FP8 | L40S · 4-bit | RTX 4090 · 4-bit |
|---|---|---|---|---|
| Qwen3.5-9B | 1× | 1× | 1× | 1× |
| Gemma 4 31B | 1× | 2× | 1× | 1× |
| Llama 3.3 70B | 2× | 4× | 2× | 4× |
| gpt-oss-120b | 4× | — | 2× | 4× |
| 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?
L40S vs RTX 4090 FAQ.
More on each GPU: NVIDIA L40S · NVIDIA RTX 4090.
Why does the L40S cost more than the RTX 4090?
It is a data-center card: twice the memory with ECC, passive cooling for dense servers, about 2.2× the dense FP8 throughput and servers of up to 8 GPUs. We price both 30% below their own market median.
Is the RTX 4090 good enough for inference?
For models that fit in 24 GB, yes: one card holds about an 18B model in FP8 or a 34B model in 4-bit. It has no ECC, and its servers stop at 4 cards.
Do both support FP8?
Yes. Both are Ada Lovelace GPUs with FP8 tensor cores, so engines such as vLLM can serve FP8 checkpoints on either. Neither supports FP4.
Other GPU comparisons.
- Ada Lovelace vs Hopper RTX 4090 vs H100 24 vs 80 GB · 1 TB/s vs 3.35 TB/s $269 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 Ada Lovelace RTX 5080 vs RTX 4090 16 vs 24 GB · 960 GB/s vs 1 TB/s $199 vs $269/mo per GPU
Rent the L40S or the RTX 4090.
Dedicated servers with 1 to 8 GPUs, one monthly price, paid in crypto. Online in under 10 minutes, no KYC.