GeForce RTX 4090 vs GeForce RTX 5070 Ti: specs and benchmark comparison
ANVIDIA · Ada Lovelace
NVIDIA GeForce RTX 4090
- 24 GB GDDR6X
- 450 W
- PSU 850 W+
October 12, 2022 · $1,599 at launch
BNVIDIA · Blackwell
NVIDIA GeForce RTX 5070 Ti
- 16 GB GDDR7
- 300 W
- PSU 750 W+
February 20, 2025 · $749 at launch
Short verdict
NVIDIA GeForce RTX 4090 leads in every scenario with comparable benchmarks: rendering and compute. This does not imply an advantage in tasks without benchmarks.
Compare with AIRendering and compute
A +50.4%NVIDIA GeForce RTX 4090 leads in this scenario (up to 50.4% in comparable benchmarks).
No comparable benchmarks yet: gaming: rasterization, gaming: ray tracing, power efficiency.
Advantages
Only measured and rated values: a noticeable benchmark gap and specs where more is better.
Where GeForce RTX 4090 is better3
Blender Benchmark 4.5.0 46% faster11,084 samples/min vs 7,590 samples/min
Video memory 8 GB more24 GB vs 16 GB
Memory bus 128 bit more384 bit vs 256 bit
Where GeForce RTX 5070 Ti is better2
Board power 150 W lower300 W (TGP) vs 450 W (TGP)
Launch price $850 lower$749 vs $1,599
Benchmarks
Results are compared within one benchmark, one version and the same conditions.
- A ahead by 46%
Blender Benchmark 4.5.0
GPU, sum of scenes monster + junkshop + classroom (OptiX / HIP / oneAPI / Metal) · samples/min, higher is better
GeForce RTX 409011,084.1 samples/minGeForce RTX 5070 Ti7,589.8 samples/minPosition among tested models (67, OptiX)
B · 7,590A · 11,08405,00010,00015,000Each dot is one model’s median in this benchmark. The scale starts at zero, samples/min. NVIDIA GeForce RTX 4090 is ahead by 46% in this benchmark.
About the measurement and source
- A · NVIDIA GeForce RTX 4090
11,084.1 samples/min · community measurement, sample 507 · November 14, 2026
Blender Open Data (Blender Foundation)
Median of 507 submissions (20 outliers excluded), interquartile range 10646.67–11378.84; submissions 2025-07-17 to 2026-10-15; OPTIX backend. Snapshot 2026-11-14.
- B · NVIDIA GeForce RTX 5070 Ti
7,589.8 samples/min · community measurement, sample 1306 · November 14, 2026
Blender Open Data (Blender Foundation)
Median of 1306 submissions (110 outliers excluded), interquartile range 7409.15–7742.68; submissions 2025-07-17 to 2026-11-11; OPTIX backend. Snapshot 2026-11-14.
difference = |A − B| ÷ the smaller value × 100%. The 5% noticeable threshold is a UX heuristic, not statistical significance.
Cycles render in Blender Benchmark. Median of the sum of samples per minute over three scenes after removing outliers.
Other comparable benchmarks (7)
| Benchmark | A | B | Difference |
|---|---|---|---|
| Blender Benchmark 4.2.0GPU, sum of scenes monster + junkshop + classroom (OptiX / HIP / oneAPI / Metal) | 10,933.2 | 7,439.9 | A +47% |
| Blender Benchmark 5.1.1GPU, sum of scenes monster + junkshop + classroom (OptiX / HIP / oneAPI / Metal) | 11,695.2 | 8,063.8 | A +45% |
| Blender Benchmark 5.2.0GPU, sum of scenes monster + junkshop + classroom (OptiX / HIP / oneAPI / Metal) | 11,570.7 | 7,768.8 | A +48.9% |
| Blender Benchmark 4.3.0GPU, sum of scenes monster + junkshop + classroom (OptiX / HIP / oneAPI / Metal) | 10,977.5 | 7,459.2 | A +47.2% |
| Blender Benchmark 4.4.0GPU, sum of scenes monster + junkshop + classroom (OptiX / HIP / oneAPI / Metal) | 11,170 | 7,613.2 | A +46.7% |
| Blender Benchmark 5.0.0GPU, sum of scenes monster + junkshop + classroom (OptiX / HIP / oneAPI / Metal) | 10,596.1 | 7,114.9 | A +48.9% |
| Blender Benchmark 5.1.0GPU, sum of scenes monster + junkshop + classroom (OptiX / HIP / oneAPI / Metal) | 11,789.5 | 7,837.9 | A +50.4% |
Each benchmark version has its own scale: numbers are compared only within a row.
Reference results (15)
| Benchmark | A | B | Difference |
|---|---|---|---|
| Blender Benchmark 3.3.0GPU, sum of scenes monster + junkshop + classroom (CUDA) | 8,389.3 | — | no pair |
| Blender Benchmark 3.4.0GPU, sum of scenes monster + junkshop + classroom (CUDA) | 6,714.7 | — | no pair |
| Blender Benchmark 3.6.0GPU, sum of scenes monster + junkshop + classroom (CUDA) | 8,851.3 | — | no pair |
| Blender Benchmark 4.0.0GPU, sum of scenes monster + junkshop + classroom (CUDA) | 7,996.2 | — | no pair |
| Blender Benchmark 4.1.0GPU, sum of scenes monster + junkshop + classroom (CUDA) | 7,467.2 | — | no pair |
| Blender Benchmark 4.2.0GPU, sum of scenes monster + junkshop + classroom (CUDA) | 7,836 | — | no pair |
| Blender Benchmark 4.5.0GPU, sum of scenes monster + junkshop + classroom (CUDA) | 8,048.3 | — | no pair |
| Blender Benchmark 3.2.1GPU, sum of scenes monster + junkshop + classroom (OptiX / HIP / oneAPI / Metal) | 9,925.9 | — | no pair |
| Blender Benchmark 3.3.0GPU, sum of scenes monster + junkshop + classroom (OptiX / HIP / oneAPI / Metal) | 12,212 | — | no pair |
| Blender Benchmark 3.4.0GPU, sum of scenes monster + junkshop + classroom (OptiX / HIP / oneAPI / Metal) | 12,658.4 | — | no pair |
| Blender Benchmark 3.5.0GPU, sum of scenes monster + junkshop + classroom (OptiX / HIP / oneAPI / Metal) | 13,184.9 | — | no pair |
| Blender Benchmark 3.6.0GPU, sum of scenes monster + junkshop + classroom (OptiX / HIP / oneAPI / Metal) | 13,110.5 | 8,594.7 | reference |
| Blender Benchmark 4.0.0GPU, sum of scenes monster + junkshop + classroom (OptiX / HIP / oneAPI / Metal) | 11,330.1 | 7,771.3 | reference |
| Blender Benchmark 4.1.0GPU, sum of scenes monster + junkshop + classroom (OptiX / HIP / oneAPI / Metal) | 11,350.5 | 6,919.6 | reference |
| Blender Benchmark 5.3.0GPU, sum of scenes monster + junkshop + classroom (OptiX / HIP / oneAPI / Metal) | 14,347.1 | 8,479.3 | reference |
Each benchmark version has its own scale: numbers are compared only within a row.
Specifications
| Specification | A · GeForce RTX 4090 | B · GeForce RTX 5070 Ti | Difference |
|---|---|---|---|
| Chip | |||
| Architecture | Ada Lovelace | Blackwell | differs |
| Shader unitsShader unit counts are comparable only within one architecture. | 16,384 (CUDA cores) | 8,960 (CUDA cores) | not directly comparable |
| Clocks | |||
| Base clock | 2,230 MHz | 2,300 MHz | B: 70 MHz more |
| Boost clockClocks of different architectures do not reflect performance directly. For laptop chips, the clock depends on the laptop’s power limit. | 2,520 MHz | 2,450 MHz | A: 70 MHz more |
| Memory | |||
| Video memory | 24 GB | 16 GB | A: 8 GB more |
| Memory type | GDDR6X | GDDR7 | differs |
| Memory bus | 384 bit | 256 bit | A: 128 bit more |
| Power and dimensions | |||
| Board powerTGP (NVIDIA) and TBP (AMD/Intel) are calculated differently; higher is not better. For laptops, the chip maker sets the range and the laptop maker picks the value, so the same chip behaves differently in different laptops. | 450 W (TGP) | 300 W (TGP) | A: 150 W higher |
| Recommended PSU | 850 W | 750 W | A: 100 W higher |
| Interface | PCIe 4.0 | PCIe 5.0 | differs |
| Length (reference / Founders Edition)Partner cards have different dimensions; check the specific SKU. | 304 mm | no data | no data for B |
| Thickness (slots) | 3 | no data | no data for B |
| Power connectors (reference) | 3x 8-pin (adapter) or 450+ W PCIe Gen 5 — FE | no data | no data for B |
Chip
Architecture
AAda Lovelace
BBlackwell
differsShader units
A16,384 (CUDA cores)
B8,960 (CUDA cores)
not directly comparableShader unit counts are comparable only within one architecture.
Clocks
Base clock
A2,230 MHz
B2,300 MHz
B: 70 MHz moreBoost clock
A2,520 MHz
B2,450 MHz
A: 70 MHz moreClocks of different architectures do not reflect performance directly. For laptop chips, the clock depends on the laptop’s power limit.
Memory
Video memory
A24 GB
B16 GB
A: 8 GB moreMemory type
AGDDR6X
BGDDR7
differsMemory bus
A384 bit
B256 bit
A: 128 bit more
Power and dimensions
Board power
A450 W (TGP)
B300 W (TGP)
A: 150 W higherTGP (NVIDIA) and TBP (AMD/Intel) are calculated differently; higher is not better. For laptops, the chip maker sets the range and the laptop maker picks the value, so the same chip behaves differently in different laptops.
Recommended PSU
A850 W
B750 W
A: 100 W higherInterface
APCIe 4.0
BPCIe 5.0
differsLength (reference / Founders Edition)
A304 mm
Bno data
no data for BPartner cards have different dimensions; check the specific SKU.
Thickness (slots)
A3
Bno data
no data for BPower connectors (reference)
A3x 8-pin (adapter) or 450+ W PCIe Gen 5 — FE
Bno data
no data for B
“More” does not always mean “better”: higher power draw or size is a system requirement, not an advantage (such bars are dimmed).
Compare with AI
An AI model reviews the specs and benchmarks of this pair and gives a plain-language verdict on who each model suits.
Questions and answers
How much video memory do NVIDIA GeForce RTX 4090 and NVIDIA GeForce RTX 5070 Ti have?
NVIDIA GeForce RTX 4090: 24 GB GDDR6X, 384-bit bus. NVIDIA GeForce RTX 5070 Ti: 16 GB GDDR7, 256-bit bus.
What power supply does the manufacturer recommend?
For NVIDIA GeForce RTX 4090, 850 W or more; for NVIDIA GeForce RTX 5070 Ti, 750 W or more. This is a recommendation for a typical system; also check the power connectors of the specific card model.
Which PCIe interface do the cards use?
NVIDIA GeForce RTX 4090: PCIe 4.0. NVIDIA GeForce RTX 5070 Ti: PCIe 5.0.
Which is faster: NVIDIA GeForce RTX 4090 or NVIDIA GeForce RTX 5070 Ti?
NVIDIA GeForce RTX 4090 leads in every scenario with comparable benchmarks: rendering and compute. This does not imply an advantage in tasks without benchmarks.
See also
Data sources (3)
- NVIDIA — specifications of NVIDIA GeForce RTX 4090, specifications of NVIDIA GeForce RTX 5070 Tichecked November 7, 2026
- Blender Open Data (Blender Foundation) — benchmark Blender Benchmark 3.3.0, benchmark Blender Benchmark 3.4.0, benchmark Blender Benchmark 3.6.0, benchmark Blender Benchmark 4.0.0, benchmark Blender Benchmark 4.1.0, benchmark Blender Benchmark 4.2.0, benchmark Blender Benchmark 4.5.0, benchmark Blender Benchmark 5.1.1, benchmark Blender Benchmark 5.2.0, benchmark Blender Benchmark 4.3.0, benchmark Blender Benchmark 4.4.0, benchmark Blender Benchmark 5.0.0, benchmark Blender Benchmark 5.1.0, benchmark Blender Benchmark 5.3.0, benchmark Blender Benchmark 3.2.1, benchmark Blender Benchmark 3.5.0checked November 14, 2026
- Wikipedia (English) — release date and launch pricechecked November 14, 2026
How we compare: methodology. Where the data comes from: sources.