GPU memory bandwidth: how it is calculated, when it limits FPS and how caches change the picture
Memory bandwidth is how many gigabytes per second a GPU can move between its cores and VRAM: bus width in bytes × memory speed in Gbps (256-bit × 20 Gbps = 640 GB/s). High resolution, ray tracing and large textures are bandwidth-hungry, and when the cores wait for memory the frame rate drops — which modern GPUs offset with big on-die caches (AMD Infinity Cache, NVIDIA's 32–96 MB L2).
bandwidth · GB/s · effective bandwidth · Infinity Cache
Key figures
From the FPSHQ database, as of Oct 14, 2026
The formula and typical values
Vendors quote "effective bandwidth" that adds cache hits; it is a marketing estimate, but the underlying idea is real: a cache hit rate of 50–70 % at 1080p halves memory traffic.
| Card | Bus | Memory | Bandwidth | Last-level cache |
|---|---|---|---|---|
| RTX 4060 | 128-bit | GDDR6 17 Gbps | 272 GB/s | 24 MB L2 |
| RX 7600 | 128-bit | GDDR6 18 Gbps | 288 GB/s | 32 MB IC |
| RTX 4070 Super | 192-bit | GDDR6X 21 Gbps | 504 GB/s | 48 MB L2 |
| RX 9070 XT | 256-bit | GDDR6 20 Gbps | 640 GB/s | 64 MB IC |
| RTX 5080 | 256-bit | GDDR7 30 Gbps | 960 GB/s | 64 MB L2 |
| RTX 5090 | 512-bit | GDDR7 28 Gbps | 1792 GB/s | 96 MB L2 |
When bandwidth is the limit
Symptoms: a card that is competitive at 1080p loses more than the usual 35 % going to 1440p and more than 55–60 % going to 4K; memory overclocking yields noticeable gains; ray tracing and MSAA hit harder than expected. Cache hit rates fall as resolution rises because the working set (frame buffers, textures per frame) outgrows the cache, which is why 128-bit cards with 24–32 MB caches are 1080p products and 4K cards need 256-bit+ or GDDR7.
The FPSHQ measured dataset shows this scaling per card: the ratio of 4K to 1080p FPS for narrow-bus cards is consistently worse than for wide-bus cards of the same tier.
Improving effective bandwidth
CPU-side memory bandwidth (DDR4/DDR5) is a different topic — see RAM frequency and dual channel.
On your PC
Enter your GPU (and CPU) — we calculate the effect of "Memory bandwidth" with the unified FPSHQ model using Cyberpunk 2077 as the example.
Save your PC in the account · Detect hardware with the app · Unified FPSHQ model estimate, not a measurement.
Where to use this on FPSHQ
Forum discussions
No topics with this term yet — be the first to ask.
What is Memory bandwidth?
Memory bandwidth is how many gigabytes per second a GPU can move between its cores and VRAM: bus width in bytes × memory speed in Gbps (256-bit × 20 Gbps = 640 GB/s). High resolution, ray tracing and large textures are bandwidth-hungry, and when the cores wait for memory the frame rate drops — which modern GPUs offset with big on-die caches (AMD Infinity Cache, NVIDIA's 32–96 MB L2).
- of GPUs since 2020 have ≥ 12 GB
- 67%
- tier S GPUs
- 52
FAQ
How do I calculate GPU memory bandwidth?
Is bandwidth more important than VRAM size?
What is Infinity Cache?
Does memory overclocking help?
Sources
- TechPowerUp GPU database (2025)
- AMD Infinity Cache (2024)
FPSHQ Team · methodology · methodology · Updated
✏️ Found a mistake? Suggest an edit
Get the FPSHQ app — it detects your hardware and shows FPS in your games.
💻 Windows app