Hopper vs Blackwell
NVIDIA's two current datacenter generations, compared on specs and on what they cost to rent right now.
Dense FP8
2.3×
4.5 PF vs 1.98 PF
Memory
2.3×
180 vs 80 GB
Bandwidth
2.3×
7,700 vs 3,350 GB/s
Price per GPU-hour
1.8×
$3.20 vs $1.82
$ per PFLOP·h (FP8)
0.8×
$0.71 vs $0.92
B200 SXM vs H100 SXM
Right now the B200 SXM is the cheapest per unit of dense FP8 compute among these GPUs: $0.71 per PFLOP-hour at $3.20 per GPU-hour.
In short
Choose Hopper (H100, H200) if
- you want the most providers to choose from
- your model fits in 80–141 GB per GPU
- a low price per hour matters more than time to result
- you need FP64: Blackwell has far less of it
Choose Blackwell (B200, B300) if
- your software uses FP8 or FP4 and keeps the GPU busy
- you need 180–288 GB of memory per GPU
- you want training runs to finish sooner
- you serve large models and can quantize to FP4
Specs and prices
| H100 SXM | H200 SXM | B200 SXM | B300 SXM | |
|---|---|---|---|---|
| Architecture | Hopper | Hopper | Blackwell | Blackwell Ultra |
| Memory | 80 GB HBM3 | 141 GB HBM3e | 180 GB HBM3e | 288 GB HBM3e |
| Memory bandwidth | 3,350 GB/s | 4,800 GB/s | 7,700 GB/s | 8,000 GB/s |
| FP4 (dense) | — | — | 9 PF | 13.5 PF |
| FP8 (dense) | 1.98 PF | 1.98 PF | 4.5 PF | 4.5 PF |
| FP16/BF16 (dense) | 0.989 PF | 0.989 PF | 2.25 PF | 2.25 PF |
| FP64 (dense) | 0.067 PF | 0.067 PF | 0.037 PF | 0.0013 PF |
| Power | 700 W | 700 W | 1,000 W | 1,100 W |
| Lowest on-demand | $1.82 | $2.25 | $3.20 | $6.60 |
| $ per PFLOP·h (FP8) | $0.92 | $1.14 | $0.71 | $1.47 |
| Providers | 22 | 17 | 11 | 12 |
Every Hopper and Blackwell datacenter GPU
With its lowest on-demand price and cost per PFLOP-hour of dense FP8.
Compare two GPUs directly
Hopper vs Blackwell: questions
Is Blackwell worth it over Hopper?
The B200 SXM has 2.3× the dense FP8 compute of the H100 SXM at 1.8× the lowest hourly price right now, so it is 22% cheaper per unit of FP8 compute. It also has 180 GB of memory versus 80 GB. If your software uses FP8 or FP4 and keeps the GPU busy, that's what you pay for; if availability or FP64 matters more, Hopper is the safer choice.
What does Blackwell's two-die design mean?
The B200 joins two large dies with a 10 TB/s link so that software sees a single GPU. It gets around the size limit of a single chip.
How much more power does Blackwell draw?
The B200 SXM is rated at 1000 W, the H100 SXM at 700 W. When you rent, power is part of the hourly price.
Is the H200 a Blackwell GPU?
No. The H200 is a Hopper GPU with more and faster memory than the H100 (141 GB of HBM3e); its compute is the same.