NVIDIA A10 — Ampere Data Center GPU
24 GB
- Architecture
- Ampere
- Generation
- Ampere
- Bandwidth
- 600 GB/s
- Workload
- Inference · Graphics
- FP16 dense
- 125 TFLOPS
- L40S 48GB from $1.29/hr in New York
- NVIDIA A100 from $4.36/hr in London
- NVIDIA A100 PCIe from $1.09/hr in London
- NVIDIA A100 SXM from $1.59/hr
- NVIDIA A16 from $0.24/hr in New Jersey
- NVIDIA A40 from $0.29/hr in New Jersey
- NVIDIA B200 SXM from $5.20/hr in Helsinki
- NVIDIA B300 SXM from $7.89/hr
- NVIDIA H100 PCIe from $2.20/hr in Montreal
- NVIDIA H100 SXM from $1.89/hr in Helsinki
- NVIDIA H200 NVL from $3.79/hr
- NVIDIA H200 SXM from $4.59/hr
- NVIDIA L4 from $0.49/hr
- NVIDIA L40 from $0.88/hr in Montreal
- NVIDIA L40S from $1.27/hr in Helsinki
- NVIDIA RTX 5000 from $0.85/hr in Gravelines
- NVIDIA RTX PRO 6000 Blackwell Server Edition from $2.09/hr
- NVIDIA RTX PRO 6000 Blackwell Server Edition MIG 1g.24gb from $0.59/hr
- Radeon RX 6700 XT from $0.21/hr
Catalog / GPU Stock
GPU STOCK
Canonical NVIDIA, AMD, and Intel GPUs available through stack8s — one model name each, with VRAM, architecture, bandwidth, and typical workload. Compare dense FP16 against memory bandwidth below. AMD and Intel Gaudi are coming soon.
Get StartedCompare / Dense TFLOPS
Bandwidth × compute
Dense figures only — no sparsity doubling. Most serving work is bandwidth-bound in decode and compute-bound in prefill; treat peak FLOPS across vendors as directional, not interchangeable.
Dense FP16 vs memory bandwidth
Bubble = VRAM
- NVIDIA
- AMD
- Intel
Ranked metric
Lead: B300 at 3,500 TFLOPS
NVIDIA
NVIDIA A10G — Ampere Graphics & Inference GPU
22 GB
- Architecture
- Ampere
- Generation
- Ampere
- Bandwidth
- 600 GB/s
- Workload
- Inference · Graphics
- FP16 dense
- 70 TFLOPS
NVIDIA A16 — Ampere Virtualization GPU
16 GB
Also listed: 2 GB / 4 GB / 8 GB / 16 GB / 32 GB / 64 GB / 128 GB
- Architecture
- Ampere
- Generation
- Ampere
- Bandwidth
- 200 GB/s
- Workload
- Inference · Graphics
NVIDIA A30 — Ampere Tensor Core GPU
24 GB
- Architecture
- Ampere
- Generation
- Ampere
- Bandwidth
- 933 GB/s
- Workload
- Training · Inference
- FP16 dense
- 165 TFLOPS
NVIDIA A40 — Ampere Data Center Graphics GPU
48 GB
Also listed: 2 GB / 4 GB / 8 GB / 12 GB / 16 GB / 24 GB / 48 GB
- Architecture
- Ampere
- Generation
- Ampere
- Bandwidth
- 696 GB/s
- Workload
- Inference · Graphics
- FP16 dense
- 150 TFLOPS
NVIDIA A100 — Ampere Tensor Core AI GPU
80 GB
Also listed: 40 GB / 80 GB
- Architecture
- Ampere
- Generation
- Ampere
- Bandwidth
- 2.0 TB/s
- Workload
- Training · Inference
- FP16 dense
- 312 TFLOPS
NVIDIA B200 — Blackwell Tensor Core AI GPU
192 GB
Also listed: 179 GB / 180 GB / 192 GB
- Architecture
- Blackwell
- Generation
- Blackwell
- Bandwidth
- 8 TB/s
- Workload
- Training · Inference
- FP16 dense
- 2,250 TFLOPS
NVIDIA B300 — Blackwell Ultra Tensor Core AI GPU
288 GB
Also listed: 268 GB / 270 GB / 288 GB
- Architecture
- Blackwell
- Generation
- Blackwell
- Bandwidth
- 8 TB/s
- Workload
- Training · Inference
- FP16 dense
- 3,500 TFLOPS
NVIDIA GB200 — Grace Blackwell AI Superchip
192 GB
- Architecture
- Blackwell
- Generation
- Grace Blackwell
- Bandwidth
- 8 TB/s
- Workload
- Training · Inference
- FP16 dense
- 2,500 TFLOPS
NVIDIA GeForce RTX 3070 — Ampere Gaming GPU
8 GB
- Architecture
- Ampere
- Generation
- GeForce 30
- Bandwidth
- 448 GB/s
- Workload
- Graphics
NVIDIA GeForce RTX 3080 — Ampere Gaming GPU
10 GB
- Architecture
- Ampere
- Generation
- GeForce 30
- Bandwidth
- 760 GB/s
- Workload
- Graphics
NVIDIA GeForce RTX 3080 Ti — Ampere Gaming GPU
12 GB
- Architecture
- Ampere
- Generation
- GeForce 30
- Bandwidth
- 912 GB/s
- Workload
- Graphics
NVIDIA GeForce RTX 3090 — Ampere Enthusiast GPU
24 GB
- Architecture
- Ampere
- Generation
- GeForce 30
- Bandwidth
- 936 GB/s
- Workload
- Graphics
NVIDIA GeForce RTX 3090 Ti — Ampere Enthusiast GPU
24 GB
- Architecture
- Ampere
- Generation
- GeForce 30
- Bandwidth
- 1.0 TB/s
- Workload
- Graphics
NVIDIA GeForce RTX 4070 Ti — Ada Lovelace Gaming GPU
12 GB
- Architecture
- Ada Lovelace
- Generation
- GeForce 40
- Bandwidth
- 504 GB/s
- Workload
- Graphics
NVIDIA GeForce RTX 4080 — Ada Lovelace Gaming GPU
16 GB
- Architecture
- Ada Lovelace
- Generation
- GeForce 40
- Bandwidth
- 717 GB/s
- Workload
- Graphics
NVIDIA GeForce RTX 4080 SUPER — Ada Lovelace Gaming GPU
16 GB
- Architecture
- Ada Lovelace
- Generation
- GeForce 40
- Bandwidth
- 736 GB/s
- Workload
- Graphics
NVIDIA GeForce RTX 4090 — Ada Lovelace Enthusiast GPU
24 GB
- Architecture
- Ada Lovelace
- Generation
- GeForce 40
- Bandwidth
- 1.0 TB/s
- Workload
- Graphics
- FP16 dense
- 165 TFLOPS
NVIDIA GeForce RTX 5080 — Blackwell Gaming GPU
16 GB
- Architecture
- Blackwell
- Generation
- GeForce 50
- Bandwidth
- 960 GB/s
- Workload
- Graphics
NVIDIA GeForce RTX 5090 — Blackwell Enthusiast GPU
32 GB
- Architecture
- Blackwell
- Generation
- GeForce 50
- Bandwidth
- 1.8 TB/s
- Workload
- Graphics
- FP16 dense
- 419 TFLOPS
NVIDIA GH200 — Grace Hopper AI Superchip
96 GB
- Architecture
- Hopper
- Generation
- Hopper
- Bandwidth
- 4 TB/s
- Workload
- Training · Inference
- FP16 dense
- 990 TFLOPS
NVIDIA H100 — Hopper Tensor Core AI GPU
80 GB
Also listed: 80 GB / 94 GB
- Architecture
- Hopper
- Generation
- Hopper
- Bandwidth
- 3.4 TB/s
- Workload
- Training · Inference
- FP16 dense
- 989 TFLOPS
NVIDIA H200 — Hopper Tensor Core AI GPU
141 GB
Also listed: 141 GB / 143 GB
- Architecture
- Hopper
- Generation
- Hopper
- Bandwidth
- 4.8 TB/s
- Workload
- Training · Inference
- FP16 dense
- 989 TFLOPS
NVIDIA L4 — Ada Lovelace Inference GPU
24 GB
Also listed: 22 GB / 24 GB
- Architecture
- Ada Lovelace
- Generation
- Ada Lovelace
- Bandwidth
- 300 GB/s
- Workload
- Inference · Graphics
- FP16 dense
- 121 TFLOPS
NVIDIA L40 — Ada Lovelace Data Center GPU
48 GB
- Architecture
- Ada Lovelace
- Generation
- Ada Lovelace
- Bandwidth
- 864 GB/s
- Workload
- Inference · Graphics
- FP16 dense
- 181 TFLOPS
NVIDIA L40S — Ada Lovelace AI & Graphics GPU
48 GB
Also listed: 44 GB / 48 GB / 96 GB / 192 GB
- Architecture
- Ada Lovelace
- Generation
- Ada Lovelace
- Bandwidth
- 864 GB/s
- Workload
- Inference · Graphics
- FP16 dense
- 181 TFLOPS
NVIDIA RTX 2000 Ada — Ada Lovelace Professional GPU
16 GB
- Architecture
- Ada Lovelace
- Generation
- RTX Ada
- Bandwidth
- 224 GB/s
- Workload
- Inference · Graphics
NVIDIA RTX 4000 Ada — Ada Lovelace Professional GPU
20 GB
- Architecture
- Ada Lovelace
- Generation
- RTX Ada
- Bandwidth
- 360 GB/s
- Workload
- Inference · Graphics
NVIDIA RTX 5000 Ada — Ada Lovelace Professional GPU
32 GB
- Architecture
- Ada Lovelace
- Generation
- RTX Ada
- Bandwidth
- 576 GB/s
- Workload
- Inference · Graphics
NVIDIA RTX 6000 Ada — Ada Lovelace Professional GPU
48 GB
- Architecture
- Ada Lovelace
- Generation
- RTX Ada
- Bandwidth
- 960 GB/s
- Workload
- Inference · Graphics
- FP16 dense
- 182 TFLOPS
NVIDIA RTX 5000 — Turing Professional GPU
32 GB
- Architecture
- Turing
- Generation
- Quadro RTX
- Bandwidth
- 448 GB/s
- Workload
- Inference · Graphics
NVIDIA RTX 6000 — Turing Professional GPU
24 GB
- Architecture
- Turing
- Generation
- Quadro RTX
- Bandwidth
- 672 GB/s
- Workload
- Inference · Graphics
NVIDIA RTX A2000 — Ampere Professional GPU
6 GB
- Architecture
- Ampere
- Generation
- RTX Ampere
- Bandwidth
- 288 GB/s
- Workload
- Inference · Graphics
NVIDIA RTX A4000 — Ampere Professional GPU
16 GB
- Architecture
- Ampere
- Generation
- RTX Ampere
- Bandwidth
- 448 GB/s
- Workload
- Inference · Graphics
- FP16 dense
- 77 TFLOPS
NVIDIA RTX A4500 — Ampere Professional GPU
20 GB
- Architecture
- Ampere
- Generation
- RTX Ampere
- Bandwidth
- 640 GB/s
- Workload
- Inference · Graphics
NVIDIA RTX A5000 — Ampere Professional GPU
24 GB
- Architecture
- Ampere
- Generation
- RTX Ampere
- Bandwidth
- 768 GB/s
- Workload
- Inference · Graphics
- FP16 dense
- 111 TFLOPS
NVIDIA RTX A6000 — Ampere Professional GPU
48 GB
- Architecture
- Ampere
- Generation
- RTX Ampere
- Bandwidth
- 768 GB/s
- Workload
- Inference · Graphics
- FP16 dense
- 155 TFLOPS
NVIDIA RTX PRO 4000 — Blackwell Professional GPU
24 GB
- Architecture
- Blackwell
- Generation
- RTX PRO Blackwell
- Bandwidth
- 672 GB/s
- Workload
- Inference · Graphics
NVIDIA RTX PRO 4500 — Blackwell Professional GPU
32 GB
- Architecture
- Blackwell
- Generation
- RTX PRO Blackwell
- Bandwidth
- 896 GB/s
- Workload
- Inference · Graphics
NVIDIA RTX PRO 5000 — Blackwell Professional GPU
48 GB
- Architecture
- Blackwell
- Generation
- RTX PRO Blackwell
- Bandwidth
- 1.3 TB/s
- Workload
- Inference · Graphics
NVIDIA RTX PRO 6000 — Blackwell Professional GPU
96 GB
- Architecture
- Blackwell
- Generation
- RTX PRO Blackwell
- Bandwidth
- 1.8 TB/s
- Workload
- Inference · Graphics
- FP16 dense
- 500 TFLOPS
NVIDIA T4 — Turing Tensor Core Inference GPU
16 GB
- Architecture
- Turing
- Generation
- Turing
- Bandwidth
- 320 GB/s
- Workload
- Inference
- FP16 dense
- 65 TFLOPS
NVIDIA V100 — Volta Tensor Core AI GPU
32 GB
Also listed: 16 GB / 32 GB
- Architecture
- Volta
- Generation
- Volta
- Bandwidth
- 900 GB/s
- Workload
- Training · Inference
- FP16 dense
- 125 TFLOPS
AMD
AMD Instinct MI25 — GCN AI & HPC GPU
Coming Soon16 GB
- Architecture
- GCN
- Generation
- Instinct
- Bandwidth
- 484 GB/s
- Workload
- Training · Inference
AMD Instinct MI300X — CDNA 3 AI Accelerator
Coming Soon192 GB
- Architecture
- CDNA 3
- Generation
- Instinct
- Bandwidth
- 5.3 TB/s
- Workload
- Training · Inference
- FP16 dense
- 1,307 TFLOPS
AMD Instinct MI325X — CDNA 3 AI Accelerator
Coming Soon256 GB
- Architecture
- CDNA 3
- Generation
- Instinct
- Bandwidth
- 6 TB/s
- Workload
- Training · Inference
- FP16 dense
- 1,307 TFLOPS
AMD Instinct MI350X — CDNA 4 AI Accelerator
Coming Soon288 GB
- Architecture
- CDNA 4
- Generation
- Instinct
- Bandwidth
- 8 TB/s
- Workload
- Training · Inference
- FP16 dense
- 2,300 TFLOPS
AMD Instinct MI355X — CDNA 4 AI Accelerator
Coming Soon288 GB
- Architecture
- CDNA 4
- Generation
- Instinct
- Bandwidth
- 8 TB/s
- Workload
- Training · Inference
- FP16 dense
- 2,500 TFLOPS
AMD Radeon Pro V520 — RDNA Professional GPU
Coming Soon8 GB
- Architecture
- RDNA
- Generation
- Radeon Pro
- Bandwidth
- 448 GB/s
- Workload
- Graphics
AMD Radeon Pro V710 — RDNA 3 Professional GPU
Coming Soon24 GB
- Architecture
- RDNA 3
- Generation
- Radeon Pro
- Bandwidth
- 432 GB/s
- Workload
- Graphics
AMD Radeon RX 6700 XT — RDNA 2 Gaming GPU
Coming Soon12 GB
- Architecture
- RDNA 2
- Generation
- Radeon RX 6000
- Bandwidth
- 384 GB/s
- Workload
- Graphics
Intel
Intel Gaudi 2 — Gaudi 2 AI Accelerator
Coming Soon96 GB
- Architecture
- Gaudi
- Generation
- Gaudi 2
- Bandwidth
- 2.5 TB/s
- Workload
- Training · Inference
- FP16 dense
- 838 TFLOPS
Intel Gaudi HL-205 — Gaudi AI Accelerator
Coming Soon32 GB
- Architecture
- Gaudi
- Generation
- Gaudi
- Bandwidth
- 1 TB/s
- Workload
- Training · Inference