Cloud GPUs with 1 TB/s+ Memory Bandwidth — May 2026
Memory-bound workloads (LLM inference, large-batch training) live or die by bandwidth. Every cloud GPU pushing 1 TB/s or more.
NVIDIA
384 GB
GB200 Superchip
HBM3e
Blackwell
NVIDIA
288 GB
B300
HBM3e
Blackwell Ultra
AMD
288 GB
MI350X
HBM3e
CDNA 4
AMD
288 GB
MI355X
HBM3e
CDNA 4
$2.59/hr
AMD
256 GB
MI325X
HBM3e
CDNA 3
$2.00/hr
NVIDIA
192 GB
B200
HBM3e
Blackwell
$1.99/hr
NVIDIA
192 GB
B100
HBM3e
Blackwell
AMD
192 GB
MI300X
HBM3
CDNA 3
$1.85/hr
NVIDIA
141 GB
H200 SXM
HBM3e
Hopper
$2.05/hr
NVIDIA
96 GB
GH200 Superchip
HBM3
Hopper
NVIDIA
80 GB
H100 SXM
HBM3
Hopper
$1.57/hr
NVIDIA
80 GB
A100 SXM (80GB)
HBM2e
Ampere
$1.10/hr
NVIDIA
40 GB
A100 SXM (40GB)
HBM2e
Ampere
$0.80/hr
NVIDIA
96 GB
RTX PRO 6000
GDDR7
Blackwell
$1.71/hr
NVIDIA
32 GB
RTX 5090
GDDR7
Blackwell
$0.34/hr
NVIDIA
24 GB
RTX 4090
GDDR6X
Ada Lovelace
$0.28/hr
NVIDIA
24 GB
RTX 3090 Ti
GDDR6X
Ampere
GB200 Superchip vs B300 vs MI350X — top picks from this guide
|
GB200 Superchip
Blackwell · 384 GB
|
B300
Blackwell Ultra · 288 GB
|
MI350X
CDNA 4 · 288 GB
|
|
|---|---|---|---|
| Specifications | |||
| Manufacturer | NVIDIA | NVIDIA | AMD |
| Architecture | Blackwell | Blackwell Ultra | CDNA 4 |
| VRAM | 384 GB HBM3e | 288 GB HBM3e | 288 GB HBM3e |
| Memory Bandwidth | 16,000 GB/s | 8,000 GB/s | 8,000 GB/s |
| FP16 (Tensor) | 4,500 TFLOPS | 2,250 TFLOPS | 1,800 TFLOPS |
| FP32 | 150 TFLOPS | 75 TFLOPS | 72 TFLOPS |
| TDP | 2700 W | 1400 W | 1000 W |
| Release Year | 2024 | 2025 | 2025 |
| Segment | Data center | Data center | Data center |
| Cloud Pricing | |||
| Cheapest On-Demand | — | — | — |
| Providers | 0 | 1 | 1 |
Build your own GPU comparison
Select any 2 GPUs from this guide and open them side-by-side.
Tip: GPU comparisons run in pairs. Pick exactly 2 — if you skip selection, we open the top 2 from this guide.