Blackwell Ultra
Released 2025NVIDIA's 2025 refresh of Blackwell (B300, GB300): more HBM3e per GPU and faster FP4 for large-scale inference.
Up to 288 GB per GPUNative precisions: FP4 · FP8 · FP16 / BF16 · TF32 · INT82 of 2 rentable now
GPU architectures
Every generation we track, newest first. Each lists its GPUs with today's lowest on-demand price, the precisions its tensor cores run natively and the most memory per GPU.
Hopper vs Blackwell →Architecture, product, superchip
- Architecture
- The design a family of chips shares: tensor cores, supported precisions, memory system. Hopper, Blackwell and CDNA 3 are architectures.
- Product
- One GPU built on an architecture, in one memory size and form factor. H100 SXM, H100 PCIe and H200 are three products on Hopper, each with its own price.
- Superchip
- A CPU and one or two GPUs on one module (GH200, GB200). PetaPrice shows the price per GPU; the CPU comes with it.
Side by side
Most memory and highest power per GPU in each architecture, and its cheapest compute right now.
| Architecture | Vendor | Year | Max memory | Max power | Native precisions | Best FP16 value |
|---|---|---|---|---|---|---|
| Blackwell Ultra | NVIDIA | 2025 | 288 GB HBM3e | 1,400 W | FP4 · FP8 · FP16 / BF16 · TF32 · INT8 | $2.93/PFLOP·h (B300 SXM) |
| Blackwell | NVIDIA | 2024–2025 | 186 GB HBM3e | 1,200 W | FP4 · FP8 · FP16 / BF16 · TF32 · INT8 | $1.42/PFLOP·h (B200 SXM) |
| Hopper | NVIDIA | 2022–2024 | 141 GB HBM3e | 1,000 W | FP8 · FP16 / BF16 · TF32 · INT8 | $1.83/PFLOP·h (H100 SXM) |
| Ada Lovelace | NVIDIA | 2022–2024 | 48 GB GDDR6 | 450 W | FP8 · FP16 / BF16 · TF32 · INT8 | $0.97/PFLOP·h (RTX 4070 SUPER) |
| Ampere | NVIDIA | 2020–2022 | 80 GB HBM2e | 450 W | FP16 / BF16 · TF32 · INT8 | $1.07/PFLOP·h (RTX A4000) |
| Turing | NVIDIA | 2018 | 48 GB GDDR6 | 295 W | FP16 · INT8 | $2.53/PFLOP·h (Quadro RTX 6000) |
| Volta | NVIDIA | 2017–2019 | 32 GB HBM2 | 300 W | FP16 | $1.52/PFLOP·h (V100 SXM2 16GB) |
| Pascal | NVIDIA | 2016 | 16 GB HBM2 | 250 W | FP16 | $68.2/PFLOP·h (P100) |
| CDNA 4 | AMD | 2025 | 288 GB HBM3e | 1,400 W | FP4 · FP8 · FP16 / BF16 · INT8 | $1.03/PFLOP·h (Instinct MI355X) |
| CDNA 3 | AMD | 2023–2024 | 256 GB HBM3e | 1,000 W | FP8 · FP16 / BF16 · TF32 · INT8 | $1.42/PFLOP·h (Instinct MI300X) |
| RDNA 3 | AMD | 2022–2023 | 48 GB GDDR6 | 355 W | FP16 / BF16 · INT8 | — |
| CDNA 2 | AMD | 2021–2022 | 128 GB HBM2e | 560 W | FP16 / BF16 · INT8 | — |
| CDNA | AMD | 2020 | 32 GB HBM2 | 300 W | FP16 / BF16 · INT8 | — |
| Xe-HPC | Intel | 2023 | 128 GB HBM2e | 600 W | FP16 / BF16 · TF32 · INT8 | — |
| Gaudi | Intel | 2022–2024 | 128 GB HBM2e | 900 W | FP8 · FP16 / BF16 | $2.55/PFLOP·h (Gaudi 2) |
| TPU | 2023–2024 | 95 GB HBM2e | — | FP16 / BF16 · INT8 | $2.94/PFLOP·h (TPU v6e (Trillium)) | |
| Inferentia | AWS | 2023 | 32 GB HBM2e | — | FP8 · FP16 / BF16 · INT8 | $3.99/PFLOP·h (Inferentia2) |
| Trainium | AWS | 2022–2024 | 96 GB HBM3 | — | FP8 · FP16 / BF16 | $7.07/PFLOP·h (Trainium) |
NVIDIA
Blackwell
Released 2024–2025NVIDIA's generation after Hopper, from the B200 and GB200 in datacenters to the RTX 50 series and RTX PRO cards. Adds FP4 tensor cores; the datacenter parts join two dies into one GPU.
Up to 186 GB per GPUNative precisions: FP4 · FP8 · FP16 / BF16 · TF32 · INT810 of 10 rentable now
Hopper
Released 2022–2024H100, H200 and GH200. Introduced FP8 tensor cores with the Transformer Engine and is still the most widely rented datacenter GPU.
Up to 141 GB per GPUNative precisions: FP8 · FP16 / BF16 · TF32 · INT86 of 8 rentable now
Ada Lovelace
Released 2022–2024RTX 40 series, L4, L40S and RTX 6000 Ada. FP8 tensor cores with GDDR6 memory instead of HBM, which makes these cards cheap but bandwidth-limited.
Up to 48 GB per GPUNative precisions: FP8 · FP16 / BF16 · TF32 · INT815 of 17 rentable now
- L40Sfrom $0.74/h
- L40from $0.69/h
- L4from $0.43/h
- RTX 6000 Adafrom $0.54/h
- RTX 5880 Adanot listed
- RTX 5000 Adafrom $0.49/h
- RTX 4000 Adafrom $0.20/h
- RTX 4000 SFF Adafrom $0.18/h
- RTX 2000 Adafrom $0.24/h
- RTX 4090from $0.34/h
- RTX 4090Dnot listed
- RTX 4080 SUPERfrom $0.28/h
- RTX 4080from $0.21/h
- RTX 4070 Tifrom $0.19/h
- RTX 4070 SUPERfrom $0.07/h
- RTX 4070from $0.13/h
- RTX 4060 Tifrom $0.10/h
Ampere
Released 2020–2022A100, A10, A40 and RTX 30 series. The first generation with BF16 and TF32 tensor cores; the A100 also added MIG partitioning.
Up to 80 GB per GPUNative precisions: FP16 / BF16 · TF32 · INT819 of 23 rentable now
- A100 SXM 80GBfrom $1.07/h
- A100 SXM 40GBfrom $0.89/h
- A100 PCIe 80GBfrom $1.13/h
- A100 PCIe 40GBfrom $1.99/h
- A800not listed
- A30from $0.33/h
- A40from $0.49/h
- A10from $1.00/h
- A10Gfrom $1.01/h
- A16from $0.47/h
- RTX A6000from $0.33/h
- RTX A5000from $0.16/h
- RTX A4500from $0.19/h
- RTX A4000from $0.08/h
- RTX A2000from $0.12/h
- RTX 3090 Tifrom $0.27/h
- RTX 3090from $0.13/h
- RTX 3080 Tifrom $0.18/h
- RTX 3080from $0.17/h
- RTX 3070 Tinot listed
- RTX 3070from $0.08/h
- RTX 3060 Tinot listed
- RTX 3060not listed
Turing
Released 2018T4, RTX 20 series and Quadro RTX. Tensor cores for FP16 and INT8, but no BF16.
Up to 48 GB per GPUNative precisions: FP16 · INT82 of 5 rentable now
Volta
Released 2017–2019V100: the first GPU with tensor cores (FP16 only).
Up to 32 GB per GPUNative precisions: FP164 of 4 rentable now
Pascal
Released 2016P100: HBM memory but no tensor cores. Mostly useful for cheap FP32 and FP64 work.
Up to 16 GB per GPUNative precisions: FP161 of 1 rentable now
AMD
CDNA 4
Released 2025AMD Instinct MI350X and MI355X. Adds FP4 and FP6 to AMD's matrix cores, with 288 GB of HBM3e per GPU.
Up to 288 GB per GPUNative precisions: FP4 · FP8 · FP16 / BF16 · INT82 of 2 rentable now
CDNA 3
Released 2023–2024AMD Instinct MI300X and MI325X. FP8 support and more memory per GPU than the Hopper parts they compete with.
Up to 256 GB per GPUNative precisions: FP8 · FP16 / BF16 · TF32 · INT82 of 2 rentable now
RDNA 3
Released 2022–2023AMD's Radeon graphics architecture (RX 7900 XTX, Radeon PRO W7900), rented mostly on community marketplaces.
Up to 48 GB per GPUNative precisions: FP16 / BF16 · INT80 of 2 rentable now
CDNA 2
Released 2021–2022AMD Instinct MI210 and MI250(X). Strong FP64 for HPC; BF16 but no FP8.
Up to 128 GB per GPUNative precisions: FP16 / BF16 · INT81 of 3 rentable now
CDNA
Released 2020AMD Instinct MI100, AMD's first compute-only datacenter architecture.
Up to 32 GB per GPUNative precisions: FP16 / BF16 · INT80 of 1 rentable now
Intel
Xe-HPC
Released 2023Intel Data Center GPU Max (Ponte Vecchio), built for HPC.
Up to 128 GB per GPUNative precisions: FP16 / BF16 · TF32 · INT80 of 1 rentable now
Gaudi
Released 2022–2024Intel's Gaudi 2 and Gaudi 3 AI accelerators (from its Habana Labs acquisition).
Up to 128 GB per GPUNative precisions: FP8 · FP16 / BF161 of 2 rentable now
Google
TPU
Released 2023–2024Google's own tensor processors, rented only on Google Cloud.
Up to 95 GB per GPUNative precisions: FP16 / BF16 · INT83 of 3 rentable now
AWS
Inferentia
Released 2023AWS's own inference chips, rented only on AWS.
Up to 32 GB per GPUNative precisions: FP8 · FP16 / BF16 · INT81 of 1 rentable now
Trainium
Released 2022–2024AWS's own training chips, rented only on AWS.
Up to 96 GB per GPUNative precisions: FP8 · FP16 / BF162 of 2 rentable now
Architecture questions
Which architecture gives the most compute per dollar right now?
Ada Lovelace: the RTX 4070 SUPER costs $0.97 per PFLOP-hour of dense FP16/BF16 at $0.07 per GPU-hour. Rankings follow prices, which refresh every 2 hours.
Which architectures support FP8 and FP4?
FP8: Blackwell Ultra, Blackwell, Hopper, Ada Lovelace, CDNA 4, CDNA 3, Gaudi, Inferentia, Trainium. FP4: Blackwell Ultra, Blackwell, CDNA 4. Older architectures can still run FP8 or FP4 models, but without the speedup.
Why rent an older architecture?
Older GPUs are often cheaper per FLOP and per GB of memory, and plenty for serving smaller models, LoRA fine-tuning or experiments. Check that the precision you need runs natively; after that it's a question of price.
What's the difference between Hopper and Grace Hopper?
Hopper is the GPU architecture (H100, H200). Grace Hopper (GH200) is a superchip: NVIDIA's Arm-based Grace CPU and a Hopper GPU on one module, joined by NVLink-C2C so the GPU can also use the CPU's memory.