{ "endpoint": "/hardware/gpus.json", "status": "beta", "as_of": "2026-05-10", "schema_version": "0.1.0-draft", "license": "CC-BY-4.0", "description": "Curated directory of accelerators relevant to autonomous agent workloads. Hand-curated; numbers are vendor-published dense peaks unless flagged sparse.", "schema": { "id": "stable slug identifier", "name": "product name", "vendor": "designer / brand owner", "architecture": "microarchitecture", "category": "datacenter_training | datacenter_inference | hyperscaler_asic | wafer_scale | inference_asic | edge | consumer | workstation", "process_nm": "fabrication node in nm (approx)", "memory_gb": "on-package memory capacity", "memory_type": "HBM3e | HBM3 | HBM2e | GDDR7 | GDDR6X | GDDR6 | LPDDR5 | SRAM | unified", "memory_bandwidth_gb_s": "peak memory bandwidth in GB/s", "fp16_tflops": "dense FP16/BF16 tensor throughput (TFLOPS)", "fp8_tflops": "dense FP8 tensor throughput (TFLOPS) — null if not native", "fp4_tflops": "dense FP4 tensor throughput (TFLOPS) — null if not native", "fp32_tflops": "non-tensor FP32 throughput", "tdp_w": "thermal design power in watts", "interconnect": "primary chip-to-chip interconnect", "released": "release year/quarter", "status": "available | sampling | announced | eol", "source": "primary spec source URL" }, "count": 50, "items": [ { "id": "nvidia-b200", "name": "B200", "vendor": "NVIDIA", "architecture": "Blackwell", "category": "datacenter_training", "process_nm": 4, "memory_gb": 192, "memory_type": "HBM3e", "memory_bandwidth_gb_s": 8000, "fp16_tflops": 2250, "fp8_tflops": 4500, "fp4_tflops": 9000, "fp32_tflops": 80, "tdp_w": 1000, "interconnect": "NVLink 5 (1.8 TB/s)", "released": "2024-Q4", "status": "available", "source": "https://www.nvidia.com/en-us/data-center/dgx-b200/" }, { "id": "nvidia-b100", "name": "B100", "vendor": "NVIDIA", "architecture": "Blackwell", "category": "datacenter_training", "process_nm": 4, "memory_gb": 192, "memory_type": "HBM3e", "memory_bandwidth_gb_s": 8000, "fp16_tflops": 1750, "fp8_tflops": 3500, "fp4_tflops": 7000, "fp32_tflops": 60, "tdp_w": 700, "interconnect": "NVLink 5 (1.8 TB/s)", "released": "2024-Q4", "status": "available", "source": "https://www.nvidia.com/en-us/data-center/dgx-b200/" }, { "id": "nvidia-gb200", "name": "GB200 Superchip", "vendor": "NVIDIA", "architecture": "Blackwell + Grace", "category": "datacenter_training", "process_nm": 4, "memory_gb": 384, "memory_type": "HBM3e + LPDDR5X", "memory_bandwidth_gb_s": 16000, "fp16_tflops": 4500, "fp8_tflops": 9000, "fp4_tflops": 18000, "fp32_tflops": 160, "tdp_w": 2700, "interconnect": "NVLink-C2C 900 GB/s CPU↔GPU; NVLink 5", "released": "2024-Q4", "status": "available", "source": "https://www.nvidia.com/en-us/data-center/gb200-nvl72/" }, { "id": "nvidia-gb300", "name": "GB300 (Blackwell Ultra) NVL72", "vendor": "NVIDIA", "architecture": "Blackwell Ultra", "category": "datacenter_training", "process_nm": 4, "memory_gb": 288, "memory_type": "HBM3e", "memory_bandwidth_gb_s": 8000, "fp16_tflops": 2500, "fp8_tflops": 5000, "fp4_tflops": 15000, "fp32_tflops": 90, "tdp_w": 1400, "interconnect": "NVLink 5", "released": "2025-H2", "status": "available", "source": "https://www.nvidia.com/en-us/data-center/gb200-nvl72/" }, { "id": "nvidia-rubin", "name": "Vera Rubin (R200)", "vendor": "NVIDIA", "architecture": "Rubin", "category": "datacenter_training", "process_nm": 3, "memory_gb": 288, "memory_type": "HBM4", "memory_bandwidth_gb_s": 22000, "fp16_tflops": null, "fp8_tflops": null, "fp4_tflops": 50000, "fp32_tflops": null, "tdp_w": null, "interconnect": "NVLink 6", "released": "2026-H2", "status": "announced", "source": "https://www.glennklockwood.com/garden/processors/R200" }, { "id": "nvidia-h200", "name": "H200", "vendor": "NVIDIA", "architecture": "Hopper", "category": "datacenter_training", "process_nm": 4, "memory_gb": 141, "memory_type": "HBM3e", "memory_bandwidth_gb_s": 4800, "fp16_tflops": 989, "fp8_tflops": 1979, "fp4_tflops": null, "fp32_tflops": 67, "tdp_w": 700, "interconnect": "NVLink 4 (900 GB/s)", "released": "2024-Q1", "status": "available", "source": "https://www.nvidia.com/en-us/data-center/h200/" }, { "id": "nvidia-h100-sxm", "name": "H100 SXM5", "vendor": "NVIDIA", "architecture": "Hopper", "category": "datacenter_training", "process_nm": 4, "memory_gb": 80, "memory_type": "HBM3", "memory_bandwidth_gb_s": 3350, "fp16_tflops": 989, "fp8_tflops": 1979, "fp4_tflops": null, "fp32_tflops": 67, "tdp_w": 700, "interconnect": "NVLink 4 (900 GB/s)", "released": "2022-Q4", "status": "available", "source": "https://www.nvidia.com/en-us/data-center/h100/" }, { "id": "nvidia-h100-pcie", "name": "H100 PCIe", "vendor": "NVIDIA", "architecture": "Hopper", "category": "datacenter_training", "process_nm": 4, "memory_gb": 80, "memory_type": "HBM2e", "memory_bandwidth_gb_s": 2000, "fp16_tflops": 756, "fp8_tflops": 1513, "fp4_tflops": null, "fp32_tflops": 51, "tdp_w": 350, "interconnect": "NVLink bridge (600 GB/s) / PCIe 5", "released": "2022-Q4", "status": "available", "source": "https://www.nvidia.com/en-us/data-center/h100/" }, { "id": "nvidia-h20", "name": "H20 (China)", "vendor": "NVIDIA", "architecture": "Hopper", "category": "datacenter_inference", "process_nm": 4, "memory_gb": 96, "memory_type": "HBM3", "memory_bandwidth_gb_s": 4000, "fp16_tflops": 148, "fp8_tflops": 296, "fp4_tflops": null, "fp32_tflops": 44, "tdp_w": 400, "interconnect": "NVLink 4 (900 GB/s)", "released": "2024-Q1", "status": "available", "source": "https://en.wikipedia.org/wiki/Hopper_(microarchitecture)" }, { "id": "nvidia-a100-80gb", "name": "A100 80GB SXM4", "vendor": "NVIDIA", "architecture": "Ampere", "category": "datacenter_training", "process_nm": 7, "memory_gb": 80, "memory_type": "HBM2e", "memory_bandwidth_gb_s": 2039, "fp16_tflops": 312, "fp8_tflops": null, "fp4_tflops": null, "fp32_tflops": 19.5, "tdp_w": 400, "interconnect": "NVLink 3 (600 GB/s)", "released": "2021-Q1", "status": "available", "source": "https://www.nvidia.com/en-us/data-center/a100/" }, { "id": "nvidia-a100-40gb", "name": "A100 40GB", "vendor": "NVIDIA", "architecture": "Ampere", "category": "datacenter_training", "process_nm": 7, "memory_gb": 40, "memory_type": "HBM2", "memory_bandwidth_gb_s": 1555, "fp16_tflops": 312, "fp8_tflops": null, "fp4_tflops": null, "fp32_tflops": 19.5, "tdp_w": 400, "interconnect": "NVLink 3 (600 GB/s)", "released": "2020-Q2", "status": "available", "source": "https://www.nvidia.com/en-us/data-center/a100/" }, { "id": "nvidia-l40s", "name": "L40S", "vendor": "NVIDIA", "architecture": "Ada Lovelace", "category": "datacenter_inference", "process_nm": 4, "memory_gb": 48, "memory_type": "GDDR6", "memory_bandwidth_gb_s": 864, "fp16_tflops": 362, "fp8_tflops": 733, "fp4_tflops": null, "fp32_tflops": 91.6, "tdp_w": 350, "interconnect": "PCIe 4", "released": "2023-Q3", "status": "available", "source": "https://www.nvidia.com/en-us/data-center/l40s/" }, { "id": "nvidia-l40", "name": "L40", "vendor": "NVIDIA", "architecture": "Ada Lovelace", "category": "datacenter_inference", "process_nm": 4, "memory_gb": 48, "memory_type": "GDDR6", "memory_bandwidth_gb_s": 864, "fp16_tflops": 181, "fp8_tflops": 362, "fp4_tflops": null, "fp32_tflops": 90.5, "tdp_w": 300, "interconnect": "PCIe 4", "released": "2022-Q4", "status": "available", "source": "https://www.nvidia.com/en-us/data-center/l40/" }, { "id": "nvidia-l4", "name": "L4", "vendor": "NVIDIA", "architecture": "Ada Lovelace", "category": "datacenter_inference", "process_nm": 4, "memory_gb": 24, "memory_type": "GDDR6", "memory_bandwidth_gb_s": 300, "fp16_tflops": 121, "fp8_tflops": 242, "fp4_tflops": null, "fp32_tflops": 30.3, "tdp_w": 72, "interconnect": "PCIe 4", "released": "2023-Q1", "status": "available", "source": "https://www.nvidia.com/en-us/data-center/l4/" }, { "id": "nvidia-a10", "name": "A10", "vendor": "NVIDIA", "architecture": "Ampere", "category": "datacenter_inference", "process_nm": 8, "memory_gb": 24, "memory_type": "GDDR6", "memory_bandwidth_gb_s": 600, "fp16_tflops": 125, "fp8_tflops": null, "fp4_tflops": null, "fp32_tflops": 31.2, "tdp_w": 150, "interconnect": "PCIe 4", "released": "2021-Q2", "status": "available", "source": "https://www.nvidia.com/en-us/data-center/products/a10-gpu/" }, { "id": "nvidia-t4", "name": "T4", "vendor": "NVIDIA", "architecture": "Turing", "category": "datacenter_inference", "process_nm": 12, "memory_gb": 16, "memory_type": "GDDR6", "memory_bandwidth_gb_s": 320, "fp16_tflops": 65, "fp8_tflops": null, "fp4_tflops": null, "fp32_tflops": 8.1, "tdp_w": 70, "interconnect": "PCIe 3", "released": "2018-Q3", "status": "available", "source": "https://www.nvidia.com/en-us/data-center/tesla-t4/" }, { "id": "nvidia-v100", "name": "V100 SXM2 32GB", "vendor": "NVIDIA", "architecture": "Volta", "category": "datacenter_training", "process_nm": 12, "memory_gb": 32, "memory_type": "HBM2", "memory_bandwidth_gb_s": 900, "fp16_tflops": 125, "fp8_tflops": null, "fp4_tflops": null, "fp32_tflops": 15.7, "tdp_w": 300, "interconnect": "NVLink 2 (300 GB/s)", "released": "2017-Q2", "status": "available", "source": "https://www.nvidia.com/en-us/data-center/v100/" }, { "id": "amd-mi355x", "name": "Instinct MI355X", "vendor": "AMD", "architecture": "CDNA 4", "category": "datacenter_training", "process_nm": 3, "memory_gb": 288, "memory_type": "HBM3e", "memory_bandwidth_gb_s": 8000, "fp16_tflops": 5000, "fp8_tflops": 10100, "fp4_tflops": 20100, "fp32_tflops": 157, "tdp_w": 1400, "interconnect": "Infinity Fabric (1075 GB/s)", "released": "2025-Q4", "status": "available", "source": "https://www.amd.com/en/products/accelerators/instinct/mi350/mi355x.html" }, { "id": "amd-mi325x", "name": "Instinct MI325X", "vendor": "AMD", "architecture": "CDNA 3", "category": "datacenter_training", "process_nm": 5, "memory_gb": 256, "memory_type": "HBM3e", "memory_bandwidth_gb_s": 6000, "fp16_tflops": 1307, "fp8_tflops": 2614, "fp4_tflops": null, "fp32_tflops": 163, "tdp_w": 1000, "interconnect": "Infinity Fabric (896 GB/s)", "released": "2024-Q4", "status": "available", "source": "https://www.amd.com/content/dam/amd/en/documents/instinct-tech-docs/product-briefs/instinct-mi325x-datasheet.pdf" }, { "id": "amd-mi300x", "name": "Instinct MI300X", "vendor": "AMD", "architecture": "CDNA 3", "category": "datacenter_training", "process_nm": 5, "memory_gb": 192, "memory_type": "HBM3", "memory_bandwidth_gb_s": 5300, "fp16_tflops": 1307, "fp8_tflops": 2614, "fp4_tflops": null, "fp32_tflops": 163, "tdp_w": 750, "interconnect": "Infinity Fabric (896 GB/s)", "released": "2023-Q4", "status": "available", "source": "https://www.amd.com/en/products/accelerators/instinct/mi300/mi300x.html" }, { "id": "amd-mi300a", "name": "Instinct MI300A (APU)", "vendor": "AMD", "architecture": "CDNA 3 + Zen 4", "category": "datacenter_training", "process_nm": 5, "memory_gb": 128, "memory_type": "HBM3 (unified)", "memory_bandwidth_gb_s": 5300, "fp16_tflops": 980, "fp8_tflops": 1960, "fp4_tflops": null, "fp32_tflops": 122, "tdp_w": 760, "interconnect": "Infinity Fabric", "released": "2023-Q4", "status": "available", "source": "https://www.amd.com/en/products/accelerators/instinct/mi300/mi300a.html" }, { "id": "amd-mi250x", "name": "Instinct MI250X", "vendor": "AMD", "architecture": "CDNA 2", "category": "datacenter_training", "process_nm": 6, "memory_gb": 128, "memory_type": "HBM2e", "memory_bandwidth_gb_s": 3276, "fp16_tflops": 383, "fp8_tflops": null, "fp4_tflops": null, "fp32_tflops": 47.9, "tdp_w": 560, "interconnect": "Infinity Fabric (800 GB/s)", "released": "2021-Q4", "status": "available", "source": "https://www.amd.com/en/products/accelerators/instinct/mi200/mi250x.html" }, { "id": "intel-gaudi-3", "name": "Gaudi 3", "vendor": "Intel (Habana)", "architecture": "Gaudi 3", "category": "datacenter_training", "process_nm": 5, "memory_gb": 128, "memory_type": "HBM2e", "memory_bandwidth_gb_s": 3700, "fp16_tflops": 1835, "fp8_tflops": 1835, "fp4_tflops": null, "fp32_tflops": null, "tdp_w": 900, "interconnect": "24x 200 GbE on-chip", "released": "2024-Q3", "status": "available", "source": "https://www.intel.com/content/www/us/en/products/details/processors/ai-processors/gaudi3.html" }, { "id": "intel-gaudi-2", "name": "Gaudi 2", "vendor": "Intel (Habana)", "architecture": "Gaudi 2", "category": "datacenter_training", "process_nm": 7, "memory_gb": 96, "memory_type": "HBM2e", "memory_bandwidth_gb_s": 2450, "fp16_tflops": 432, "fp8_tflops": 865, "fp4_tflops": null, "fp32_tflops": null, "tdp_w": 600, "interconnect": "24x 100 GbE on-chip", "released": "2022-Q2", "status": "available", "source": "https://www.intel.com/content/www/us/en/products/details/processors/ai-processors/gaudi2.html" }, { "id": "intel-gpu-max-1550", "name": "Data Center GPU Max 1550 (Ponte Vecchio)", "vendor": "Intel", "architecture": "Xe-HPC", "category": "datacenter_training", "process_nm": 7, "memory_gb": 128, "memory_type": "HBM2e", "memory_bandwidth_gb_s": 3276, "fp16_tflops": 839, "fp8_tflops": null, "fp4_tflops": null, "fp32_tflops": 52, "tdp_w": 600, "interconnect": "Xe Link", "released": "2023-Q1", "status": "available", "source": "https://www.intel.com/content/www/us/en/products/details/discrete-gpus/data-center-gpu/max-series.html" }, { "id": "google-tpu-v6e", "name": "TPU v6e (Trillium)", "vendor": "Google", "architecture": "Trillium", "category": "hyperscaler_asic", "process_nm": 5, "memory_gb": 32, "memory_type": "HBM3", "memory_bandwidth_gb_s": 1640, "fp16_tflops": 918, "fp8_tflops": 1836, "fp4_tflops": null, "fp32_tflops": null, "tdp_w": null, "interconnect": "ICI 13 TB/s per chip", "released": "2024-Q4", "status": "available", "source": "https://cloud.google.com/blog/products/compute/introducing-trillium-6th-gen-tpus" }, { "id": "google-tpu-v5p", "name": "TPU v5p", "vendor": "Google", "architecture": "TPU v5p", "category": "hyperscaler_asic", "process_nm": 5, "memory_gb": 95, "memory_type": "HBM3", "memory_bandwidth_gb_s": 2765, "fp16_tflops": 459, "fp8_tflops": 918, "fp4_tflops": null, "fp32_tflops": null, "tdp_w": null, "interconnect": "ICI 4800 Gbps/chip (3D torus)", "released": "2023-Q4", "status": "available", "source": "https://docs.cloud.google.com/tpu/docs/v5p" }, { "id": "google-tpu-v5e", "name": "TPU v5e", "vendor": "Google", "architecture": "TPU v5e", "category": "hyperscaler_asic", "process_nm": 5, "memory_gb": 16, "memory_type": "HBM2", "memory_bandwidth_gb_s": 819, "fp16_tflops": 197, "fp8_tflops": 393, "fp4_tflops": null, "fp32_tflops": null, "tdp_w": 170, "interconnect": "ICI 1600 Gbps/chip", "released": "2023-Q3", "status": "available", "source": "https://cloud.google.com/tpu/docs/v5e" }, { "id": "google-tpu-v4", "name": "TPU v4", "vendor": "Google", "architecture": "TPU v4", "category": "hyperscaler_asic", "process_nm": 7, "memory_gb": 32, "memory_type": "HBM2", "memory_bandwidth_gb_s": 1228, "fp16_tflops": 275, "fp8_tflops": null, "fp4_tflops": null, "fp32_tflops": null, "tdp_w": 192, "interconnect": "OCS 3D torus", "released": "2021-Q2", "status": "available", "source": "https://cloud.google.com/tpu/docs/v4" }, { "id": "aws-trainium-3", "name": "Trainium 3", "vendor": "AWS", "architecture": "Trainium 3", "category": "hyperscaler_asic", "process_nm": 3, "memory_gb": 144, "memory_type": "HBM3e", "memory_bandwidth_gb_s": 4900, "fp16_tflops": 1260, "fp8_tflops": 2520, "fp4_tflops": 5040, "fp32_tflops": null, "tdp_w": null, "interconnect": "NeuronLink-v3", "released": "2025-Q4", "status": "available", "source": "https://aws.amazon.com/ec2/instance-types/trn3/" }, { "id": "aws-trainium-2", "name": "Trainium 2", "vendor": "AWS", "architecture": "Trainium 2", "category": "hyperscaler_asic", "process_nm": 5, "memory_gb": 96, "memory_type": "HBM3", "memory_bandwidth_gb_s": 2900, "fp16_tflops": 667, "fp8_tflops": 1300, "fp4_tflops": null, "fp32_tflops": null, "tdp_w": null, "interconnect": "NeuronLink-v2", "released": "2024-Q4", "status": "available", "source": "https://aws.amazon.com/ec2/instance-types/trn2/" }, { "id": "aws-inferentia-2", "name": "Inferentia 2", "vendor": "AWS", "architecture": "Inferentia 2", "category": "hyperscaler_asic", "process_nm": 7, "memory_gb": 32, "memory_type": "HBM", "memory_bandwidth_gb_s": 820, "fp16_tflops": 190, "fp8_tflops": 380, "fp4_tflops": null, "fp32_tflops": null, "tdp_w": null, "interconnect": "NeuronLink", "released": "2023-Q2", "status": "available", "source": "https://aws.amazon.com/machine-learning/inferentia/" }, { "id": "microsoft-maia-100", "name": "Maia 100", "vendor": "Microsoft", "architecture": "Maia", "category": "hyperscaler_asic", "process_nm": 5, "memory_gb": 64, "memory_type": "HBM2e", "memory_bandwidth_gb_s": 1600, "fp16_tflops": null, "fp8_tflops": null, "fp4_tflops": null, "fp32_tflops": null, "tdp_w": 700, "interconnect": "Ethernet-based custom fabric", "released": "2024-Q1", "status": "available", "source": "https://news.microsoft.com/source/features/ai/in-house-chips-silicon-to-service-to-meet-ai-demand/" }, { "id": "meta-mtia-v2", "name": "MTIA v2", "vendor": "Meta", "architecture": "MTIA", "category": "hyperscaler_asic", "process_nm": 5, "memory_gb": 128, "memory_type": "LPDDR5", "memory_bandwidth_gb_s": 204, "fp16_tflops": 354, "fp8_tflops": 708, "fp4_tflops": null, "fp32_tflops": null, "tdp_w": 90, "interconnect": "PCIe Gen5 + on-rack mesh", "released": "2024-Q2", "status": "available", "source": "https://ai.meta.com/blog/next-generation-meta-training-inference-accelerator-AI-MTIA/" }, { "id": "cerebras-wse-3", "name": "WSE-3", "vendor": "Cerebras", "architecture": "WSE-3", "category": "wafer_scale", "process_nm": 5, "memory_gb": 44, "memory_type": "SRAM (on-wafer)", "memory_bandwidth_gb_s": 21000000, "fp16_tflops": 62500, "fp8_tflops": 125000, "fp4_tflops": null, "fp32_tflops": null, "tdp_w": 23000, "interconnect": "On-wafer fabric 214 Pb/s aggregate", "released": "2024-Q1", "status": "available", "source": "https://www.cerebras.ai/chip" }, { "id": "groq-lpu", "name": "LPU v1 (GroqCard)", "vendor": "Groq", "architecture": "TSP", "category": "inference_asic", "process_nm": 14, "memory_gb": 0.23, "memory_type": "SRAM (on-die)", "memory_bandwidth_gb_s": 80000, "fp16_tflops": 188, "fp8_tflops": 375, "fp4_tflops": null, "fp32_tflops": null, "tdp_w": 375, "interconnect": "Chip-to-chip via custom router", "released": "2021-Q1", "status": "available", "source": "https://groq.com/lpu-architecture" }, { "id": "sambanova-sn40l", "name": "SN40L (RDU)", "vendor": "SambaNova", "architecture": "RDU SN40L", "category": "inference_asic", "process_nm": 5, "memory_gb": 1520, "memory_type": "DRAM (1.5TB) + 64GB HBM3 + 520MB SRAM", "memory_bandwidth_gb_s": null, "fp16_tflops": 638, "fp8_tflops": 1276, "fp4_tflops": null, "fp32_tflops": null, "tdp_w": 750, "interconnect": "Dataflow mesh", "released": "2023-Q3", "status": "available", "source": "https://sambanova.ai/products/sn40l" }, { "id": "graphcore-bow-ipu", "name": "Bow IPU", "vendor": "Graphcore", "architecture": "Bow", "category": "inference_asic", "process_nm": 7, "memory_gb": 0.9, "memory_type": "SRAM (on-die)", "memory_bandwidth_gb_s": 65000, "fp16_tflops": 350, "fp8_tflops": null, "fp4_tflops": null, "fp32_tflops": 87, "tdp_w": 350, "interconnect": "IPU-Link 320 GB/s", "released": "2022-Q1", "status": "available", "source": "https://www.graphcore.ai/bow-processors" }, { "id": "tenstorrent-wormhole-n300", "name": "Wormhole n300", "vendor": "Tenstorrent", "architecture": "Wormhole", "category": "inference_asic", "process_nm": 12, "memory_gb": 24, "memory_type": "GDDR6", "memory_bandwidth_gb_s": 576, "fp16_tflops": 262, "fp8_tflops": 466, "fp4_tflops": null, "fp32_tflops": null, "tdp_w": 300, "interconnect": "16x 100 GbE on-chip", "released": "2024-Q2", "status": "available", "source": "https://tenstorrent.com/hardware/wormhole" }, { "id": "nvidia-jetson-orin-agx-64gb", "name": "Jetson AGX Orin 64GB", "vendor": "NVIDIA", "architecture": "Ampere", "category": "edge", "process_nm": 8, "memory_gb": 64, "memory_type": "LPDDR5", "memory_bandwidth_gb_s": 205, "fp16_tflops": 137, "fp8_tflops": null, "fp4_tflops": null, "fp32_tflops": 8.5, "tdp_w": 60, "interconnect": "PCIe 4 / 10 GbE", "released": "2022-Q1", "status": "available", "source": "https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-orin/" }, { "id": "google-coral-edge-tpu", "name": "Coral Edge TPU", "vendor": "Google", "architecture": "Edge TPU", "category": "edge", "process_nm": null, "memory_gb": 0.008, "memory_type": "SRAM", "memory_bandwidth_gb_s": null, "fp16_tflops": null, "fp8_tflops": null, "fp4_tflops": null, "fp32_tflops": null, "tdp_w": 2, "interconnect": "USB 3 / M.2 / PCIe", "released": "2019-Q1", "status": "available", "source": "https://coral.ai/products/" }, { "id": "hailo-10h", "name": "Hailo-10H", "vendor": "Hailo", "architecture": "Hailo-10", "category": "edge", "process_nm": null, "memory_gb": null, "memory_type": null, "memory_bandwidth_gb_s": null, "fp16_tflops": null, "fp8_tflops": null, "fp4_tflops": null, "fp32_tflops": null, "tdp_w": 3.5, "interconnect": "PCIe Gen3 / M.2", "released": "2024-Q2", "status": "available", "source": "https://hailo.ai/products/hailo-10/" }, { "id": "qualcomm-cloud-ai-100", "name": "Cloud AI 100 Ultra", "vendor": "Qualcomm", "architecture": "Cloud AI 100", "category": "datacenter_inference", "process_nm": 7, "memory_gb": 128, "memory_type": "LPDDR4X", "memory_bandwidth_gb_s": 548, "fp16_tflops": 287, "fp8_tflops": 575, "fp4_tflops": null, "fp32_tflops": null, "tdp_w": 150, "interconnect": "PCIe 4", "released": "2024-Q2", "status": "available", "source": "https://www.qualcomm.com/products/technology/processors/cloud-artificial-intelligence/cloud-ai-100" }, { "id": "nvidia-rtx-5090", "name": "GeForce RTX 5090", "vendor": "NVIDIA", "architecture": "Blackwell", "category": "consumer", "process_nm": 4, "memory_gb": 32, "memory_type": "GDDR7", "memory_bandwidth_gb_s": 1792, "fp16_tflops": 419, "fp8_tflops": 838, "fp4_tflops": 1676, "fp32_tflops": 104.8, "tdp_w": 575, "interconnect": "PCIe 5", "released": "2025-Q1", "status": "available", "source": "https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5090/" }, { "id": "nvidia-rtx-4090", "name": "GeForce RTX 4090", "vendor": "NVIDIA", "architecture": "Ada Lovelace", "category": "consumer", "process_nm": 4, "memory_gb": 24, "memory_type": "GDDR6X", "memory_bandwidth_gb_s": 1008, "fp16_tflops": 330, "fp8_tflops": 660, "fp4_tflops": null, "fp32_tflops": 82.6, "tdp_w": 450, "interconnect": "PCIe 4", "released": "2022-Q4", "status": "available", "source": "https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4090/" }, { "id": "nvidia-rtx-6000-ada", "name": "RTX 6000 Ada", "vendor": "NVIDIA", "architecture": "Ada Lovelace", "category": "workstation", "process_nm": 4, "memory_gb": 48, "memory_type": "GDDR6", "memory_bandwidth_gb_s": 960, "fp16_tflops": 365, "fp8_tflops": 730, "fp4_tflops": null, "fp32_tflops": 91.1, "tdp_w": 300, "interconnect": "PCIe 4", "released": "2022-Q4", "status": "available", "source": "https://www.nvidia.com/en-us/design-visualization/rtx-6000/" }, { "id": "nvidia-rtx-a6000", "name": "RTX A6000", "vendor": "NVIDIA", "architecture": "Ampere", "category": "workstation", "process_nm": 8, "memory_gb": 48, "memory_type": "GDDR6", "memory_bandwidth_gb_s": 768, "fp16_tflops": 155, "fp8_tflops": null, "fp4_tflops": null, "fp32_tflops": 38.7, "tdp_w": 300, "interconnect": "NVLink (112 GB/s) / PCIe 4", "released": "2020-Q4", "status": "available", "source": "https://www.nvidia.com/en-us/design-visualization/rtx-a6000/" }, { "id": "amd-radeon-rx-7900-xtx", "name": "Radeon RX 7900 XTX", "vendor": "AMD", "architecture": "RDNA 3", "category": "consumer", "process_nm": 5, "memory_gb": 24, "memory_type": "GDDR6", "memory_bandwidth_gb_s": 960, "fp16_tflops": 123, "fp8_tflops": null, "fp4_tflops": null, "fp32_tflops": 61.4, "tdp_w": 355, "interconnect": "PCIe 4", "released": "2022-Q4", "status": "available", "source": "https://www.amd.com/en/products/graphics/amd-radeon-rx-7900xtx" }, { "id": "apple-m4-max", "name": "M4 Max", "vendor": "Apple", "architecture": "Apple Silicon M4", "category": "workstation", "process_nm": 3, "memory_gb": 128, "memory_type": "LPDDR5X (unified)", "memory_bandwidth_gb_s": 546, "fp16_tflops": 34, "fp8_tflops": null, "fp4_tflops": null, "fp32_tflops": 17, "tdp_w": 70, "interconnect": "On-package fabric", "released": "2024-Q4", "status": "available", "source": "https://www.apple.com/macbook-pro/specs/" }, { "id": "apple-m3-ultra", "name": "M3 Ultra", "vendor": "Apple", "architecture": "Apple Silicon M3", "category": "workstation", "process_nm": 3, "memory_gb": 512, "memory_type": "LPDDR5 (unified)", "memory_bandwidth_gb_s": 819, "fp16_tflops": 54, "fp8_tflops": null, "fp4_tflops": null, "fp32_tflops": 27, "tdp_w": 295, "interconnect": "UltraFusion on-package", "released": "2025-Q1", "status": "available", "source": "https://www.apple.com/mac-studio/specs/" } ] }