[ { "id": "nvidia-h100-sxm", "name": "NVIDIA H100 SXM5", "manufacturer": "NVIDIA", "architecture": "Hopper", "memory": "80GB HBM3", "memory_bandwidth": "3.35 TB/s", "fp16_tflops": 1979, "fp8_tflops": 3958, "interconnect": "NVLink4 (900GB/s)", "tpb_watt": 700, "released": "2022-10-01", "status": "Shipping (allocation constrained)", "pricing": { "retail_usd": "30000-40000", "cloud_per_hour": { "aws_p5_xlarge": 12.99, "azure_nd96asr_v4": 11.91, "gcp_a3_highgpu": 11.8, "coreweave": 2.23 } }, "availability": "Constrained — major allocations to hyperscalers and AI labs", "tags": [ "training", "inference", "flagship" ], "last_updated": "2026-06-23" }, { "id": "nvidia-h100-pcie", "name": "NVIDIA H100 PCIe", "manufacturer": "NVIDIA", "architecture": "Hopper", "memory": "80GB HBM3", "memory_bandwidth": "2.0 TB/s", "fp16_tflops": 1513, "fp8_tflops": 3026, "interconnect": "PCIe Gen5 (128GB/s)", "tpb_watt": 350, "released": "2023-04-01", "status": "Shipping", "pricing": { "retail_usd": "25000-35000", "cloud_per_hour": { "coreweave": 1.99, "runpod": 1.98 } }, "availability": "Available — lower-power variant for inference and smaller training clusters", "tags": [ "training", "inference", "pcie" ], "last_updated": "2026-06-24" }, { "id": "nvidia-h200-sxm", "name": "NVIDIA H200 SXM5", "manufacturer": "NVIDIA", "architecture": "Hopper (refresh)", "memory": "141GB HBM3e", "memory_bandwidth": "4.8 TB/s", "fp16_tflops": 1979, "fp8_tflops": 3958, "interconnect": "NVLink4 (900GB/s)", "tpb_watt": 700, "released": "2024-11-01", "status": "Shipping", "pricing": { "retail_usd": "35000-45000", "cloud_per_hour": { "coreweave": 2.49, "lambda_labs": 2.49 } }, "availability": "Available via neoclouds; hyperscaler rollouts ongoing", "tags": [ "training", "inference", "flagship", "hbm3e" ], "last_updated": "2026-06-23" }, { "id": "nvidia-h200-pcie", "name": "NVIDIA H200 PCIe", "manufacturer": "NVIDIA", "architecture": "Hopper (refresh)", "memory": "141GB HBM3e", "memory_bandwidth": "4.0 TB/s", "fp16_tflops": 1513, "fp8_tflops": 3026, "interconnect": "PCIe Gen5 (128GB/s)", "tpb_watt": 350, "released": "2025-04-01", "status": "Shipping", "pricing": { "retail_usd": "30000-40000", "cloud_per_hour": { "coreweave": 2.1, "runpod": 2.15 } }, "availability": "Available — inference-focused variant with high memory capacity", "tags": [ "inference", "hbm3e", "pcie" ], "last_updated": "2026-06-24" }, { "id": "nvidia-b200", "name": "NVIDIA B200", "manufacturer": "NVIDIA", "architecture": "Blackwell", "memory": "192GB HBM3e", "memory_bandwidth": "8.0 TB/s", "fp16_tflops": 2250, "fp8_tflops": 4500, "fp4_tflops": 9000, "interconnect": "NVLink5 (1.8TB/s)", "tpb_watt": 1000, "released": "2025-03-01", "status": "Shipping (limited initial allocation)", "pricing": { "retail_usd": "40000-60000", "cloud_per_hour": { "coreweave": 3.5 } }, "availability": "Ramping — Q2 2026 volume increasing; GB200 systems shipping", "tags": [ "training", "inference", "flagship", "blackwell", "fp4" ], "last_updated": "2026-06-23" }, { "id": "nvidia-b100", "name": "NVIDIA B100", "manufacturer": "NVIDIA", "architecture": "Blackwell", "memory": "192GB HBM3e", "memory_bandwidth": "8.0 TB/s", "fp16_tflops": 1800, "fp8_tflops": 3600, "fp4_tflops": 7200, "interconnect": "NVLink5 (1.8TB/s)", "tpb_watt": 700, "released": "2025-06-01", "status": "Shipping", "pricing": { "retail_usd": "30000-45000", "cloud_per_hour": { "coreweave": 2.8 } }, "availability": "Available — lower-power Blackwell variant for mainstream training and inference", "tags": [ "training", "inference", "blackwell", "fp4" ], "last_updated": "2026-06-24" }, { "id": "nvidia-gb200-nvl72", "name": "NVIDIA GB200 NVL72", "manufacturer": "NVIDIA", "architecture": "Blackwell (Grace+Blackwell)", "memory": "13.8TB HBM3e (72 GPU system)", "memory_bandwidth": "8.0 TB/s per GPU", "fp16_tflops": 162000, "fp8_tflops": 324000, "fp4_tflops": 648000, "interconnect": "NVLink5 (1.8TB/s) + NVSwitch", "tpb_watt": "27000 (system)", "released": "2025-06-01", "status": "Shipping to select customers", "pricing": { "retail_usd": "3000000-3500000 (per NVL72 rack)", "cloud_per_hour": { "coreweave": "TBD" } }, "availability": "Limited — Dell, Supermicro, Cisco shipping systems; volume ramp H2 2026", "tags": [ "training", "frontier", "blackwell", "grace", "rack-scale" ], "last_updated": "2026-06-23" }, { "id": "nvidia-gb200-nvl36", "name": "NVIDIA GB200 NVL36", "manufacturer": "NVIDIA", "architecture": "Blackwell (Grace+Blackwell)", "memory": "6.9TB HBM3e (36 GPU system)", "memory_bandwidth": "8.0 TB/s per GPU", "fp16_tflops": 81000, "fp8_tflops": 162000, "fp4_tflops": 324000, "interconnect": "NVLink5 (1.8TB/s) + NVSwitch", "tpb_watt": "13500 (system)", "released": "2025-09-01", "status": "Shipping", "pricing": { "retail_usd": "1500000-1800000 (per NVL36 rack)", "cloud_per_hour": { "coreweave": "TBD" } }, "availability": "Available — smaller rack-scale system for mid-size deployments", "tags": [ "training", "inference", "blackwell", "grace", "rack-scale" ], "last_updated": "2026-06-24" }, { "id": "nvidia-l40s", "name": "NVIDIA L40S", "manufacturer": "NVIDIA", "architecture": "Ada Lovelace", "memory": "48GB GDDR6", "memory_bandwidth": "0.864 TB/s", "fp16_tflops": 362, "fp8_tflops": 733, "interconnect": "PCIe Gen4", "tpb_watt": 350, "released": "2023-08-01", "status": "Shipping (widely available)", "pricing": { "retail_usd": "8000-12000", "cloud_per_hour": { "coreweave": 0.8, "runpod": 0.79 } }, "availability": "Widely available — primary choice for inference workloads", "tags": [ "inference", "available", "mid-range" ], "last_updated": "2026-06-23" }, { "id": "nvidia-l40", "name": "NVIDIA L40", "manufacturer": "NVIDIA", "architecture": "Ada Lovelace", "memory": "48GB GDDR6", "memory_bandwidth": "0.864 TB/s", "fp16_tflops": 181, "fp8_tflops": 362, "interconnect": "PCIe Gen4", "tpb_watt": 300, "released": "2022-10-01", "status": "Shipping", "pricing": { "retail_usd": "6000-9000", "cloud_per_hour": { "runpod": 0.65 } }, "availability": "Available — graphics and inference workloads, lower FP8 than L40S", "tags": [ "inference", "available", "mid-range" ], "last_updated": "2026-06-24" }, { "id": "nvidia-l4", "name": "NVIDIA L4", "manufacturer": "NVIDIA", "architecture": "Ada Lovelace", "memory": "24GB GDDR6", "memory_bandwidth": "0.300 TB/s", "fp16_tflops": 121, "fp8_tflops": 242, "interconnect": "PCIe Gen4", "tpb_watt": 72, "released": "2023-01-01", "status": "Shipping (widely available)", "pricing": { "retail_usd": "2000-3500", "cloud_per_hour": { "gcp_t4_l4": 0.75, "runpod": 0.49 } }, "availability": "Widely available — low-power inference GPU for edge and cloud", "tags": [ "inference", "available", "low-power", "edge" ], "last_updated": "2026-06-24" }, { "id": "nvidia-a100-sxm", "name": "NVIDIA A100 SXM4 80GB", "manufacturer": "NVIDIA", "architecture": "Ampere", "memory": "80GB HBM2e", "memory_bandwidth": "2.0 TB/s", "fp16_tflops": 624, "fp8_tflops": "N/A", "interconnect": "NVLink3 (600GB/s)", "tpb_watt": 400, "released": "2021-06-01", "status": "Shipping (mature, widely available)", "pricing": { "retail_usd": "10000-16000", "cloud_per_hour": { "aws_p4d": 3.22, "gcp_a2_megagpu": 3.67, "runpod": 0.75 } }, "availability": "Widely available — decreasing as Hopper/Blackwell ramp", "tags": [ "training", "inference", "legacy", "available" ], "last_updated": "2026-06-23" }, { "id": "nvidia-a100-pcie", "name": "NVIDIA A100 PCIe 80GB", "manufacturer": "NVIDIA", "architecture": "Ampere", "memory": "80GB HBM2e", "memory_bandwidth": "1.94 TB/s", "fp16_tflops": 312, "fp8_tflops": "N/A", "interconnect": "PCIe Gen4 (64GB/s)", "tpb_watt": 300, "released": "2021-06-01", "status": "Shipping (widely available)", "pricing": { "retail_usd": "8000-14000", "cloud_per_hour": { "runpod": 0.65, "vast_ai": 0.55 } }, "availability": "Widely available — PCIe variant for inference and smaller clusters", "tags": [ "inference", "legacy", "available", "pcie" ], "last_updated": "2026-06-24" }, { "id": "nvidia-a30", "name": "NVIDIA A30", "manufacturer": "NVIDIA", "architecture": "Ampere", "memory": "24GB HBM2", "memory_bandwidth": "0.933 TB/s", "fp16_tflops": 165, "fp8_tflops": "N/A", "interconnect": "NVLink3 (600GB/s)", "tpb_watt": 165, "released": "2021-04-01", "status": "Shipping", "pricing": { "retail_usd": "3000-5000", "cloud_per_hour": {} }, "availability": "Available — mainstream inference GPU, lower cost than A100", "tags": [ "inference", "available", "mid-range" ], "last_updated": "2026-06-24" }, { "id": "nvidia-a10", "name": "NVIDIA A10", "manufacturer": "NVIDIA", "architecture": "Ampere", "memory": "24GB GDDR6", "memory_bandwidth": "0.600 TB/s", "fp16_tflops": 125, "fp8_tflops": "N/A", "interconnect": "PCIe Gen4", "tpb_watt": 150, "released": "2021-04-01", "status": "Shipping", "pricing": { "retail_usd": "2500-4000", "cloud_per_hour": { "runpod": 0.4 } }, "availability": "Available — inference and visualization GPU", "tags": [ "inference", "available", "mid-range" ], "last_updated": "2026-06-24" }, { "id": "nvidia-t4", "name": "NVIDIA T4", "manufacturer": "NVIDIA", "architecture": "Turing", "memory": "16GB GDDR6", "memory_bandwidth": "0.320 TB/s", "fp16_tflops": 65, "fp8_tflops": "N/A", "interconnect": "PCIe Gen3", "tpb_watt": 70, "released": "2018-09-01", "status": "Shipping (legacy)", "pricing": { "retail_usd": "1000-2000", "cloud_per_hour": { "aws_g4dn": 0.526, "gcp_t4": 0.35 } }, "availability": "Widely available — entry-level inference, being phased out for L4", "tags": [ "inference", "legacy", "low-power", "edge" ], "last_updated": "2026-06-24" }, { "id": "nvidia-v100", "name": "NVIDIA V100 SXM2", "manufacturer": "NVIDIA", "architecture": "Volta", "memory": "32GB HBM2", "memory_bandwidth": "0.900 TB/s", "fp16_tflops": 125, "fp8_tflops": "N/A", "interconnect": "NVLink2 (300GB/s)", "tpb_watt": 300, "released": "2017-06-01", "status": "End of life (still deployed)", "pricing": { "retail_usd": "2000-5000 (used)", "cloud_per_hour": { "vast_ai": 0.3 } }, "availability": "Legacy — still available on secondary market and some clouds; not recommended for new deployments", "tags": [ "training", "inference", "legacy", "eol" ], "last_updated": "2026-06-24" }, { "id": "nvidia-b300", "name": "NVIDIA B300", "manufacturer": "NVIDIA", "architecture": "Blackwell Ultra", "memory": "288GB HBM3e", "memory_bandwidth": "8.0 TB/s", "fp16_tflops": 2500, "fp8_tflops": 5000, "fp4_tflops": 10000, "interconnect": "NVLink5 (1.8TB/s)", "tpb_watt": 1200, "released": "2026-05-01", "status": "Shipping (HGX B300 available now)", "pricing": { "retail_usd": "TBD", "cloud_per_hour": { "coreweave": "TBD" } }, "availability": "Shipping now — HGX B300 platform in production. Current high-end shipping GPU, successor to B200 with 50% more HBM3e memory (288GB vs 192GB). Increased compute density for frontier-scale training.", "tags": [ "training", "inference", "flagship", "blackwell-ultra", "fp4", "hbm3e" ], "last_updated": "2026-06-24" }, { "id": "nvidia-vera-rubin", "name": "NVIDIA Vera Rubin", "manufacturer": "NVIDIA", "architecture": "Vera Rubin", "memory": "TBD (expected 288GB+ HBM4)", "memory_bandwidth": "TBD (expected 13+ TB/s)", "fp16_tflops": "TBD", "fp8_tflops": "TBD", "fp4_tflops": "TBD (expected 40+ PFLOPS)", "interconnect": "NVLink6 (expected)", "tpb_watt": "TBD (expected 1400W+)", "released": "2026-H2 (expected)", "status": "Announced — shipping H2 2026 to hyperscalers only", "pricing": { "retail_usd": "TBD", "cloud_per_hour": {} }, "availability": "Announced at GTC 2026. Renamed from 'Rubin' to 'Vera Rubin' at GTC 2026. Ships H2 2026 to hyperscalers ONLY (no general availability expected in 2026). Next-generation architecture succeeding Blackwell, featuring HBM4 memory.", "tags": [ "training", "inference", "flagship", "vera-rubin", "upcoming", "hbm4" ], "last_updated": "2026-06-24" }, { "id": "nvidia-rubin-ultra", "name": "NVIDIA Rubin Ultra", "manufacturer": "NVIDIA", "architecture": "Rubin Ultra", "memory": "TBD (expected 576GB+ HBM4e)", "memory_bandwidth": "TBD (expected 13+ TB/s)", "fp16_tflops": "TBD", "fp8_tflops": "TBD", "fp4_tflops": "TBD (expected 50+ PFLOPS)", "interconnect": "NVLink6 (expected)", "tpb_watt": "TBD (expected 1800W+)", "released": "2027-Q3 (expected)", "status": "Announced (not yet shipping)", "pricing": { "retail_usd": "TBD", "cloud_per_hour": {} }, "availability": "Not yet available — announced at GTC 2025. Rubin Ultra is the follow-on to Vera Rubin, expected to push the 1-exaflop-per-rack boundary.", "tags": [ "training", "inference", "flagship", "rubin", "upcoming", "hbm4e" ], "last_updated": "2026-06-24" }, { "id": "amd-mi300x", "name": "AMD Instinct MI300X", "manufacturer": "AMD", "architecture": "CDNA 3", "memory": "192GB HBM3", "memory_bandwidth": "5.3 TB/s", "fp16_tflops": 1307, "fp8_tflops": 2614, "interconnect": "Infinity Fabric (896GB/s)", "tpb_watt": 750, "released": "2023-12-01", "status": "Shipping", "pricing": { "retail_usd": "15000-20000", "cloud_per_hour": { "coreweave": 1.99 } }, "availability": "Available — gaining adoption; Meta, Microsoft, Oracle deploying", "tags": [ "training", "inference", "amd", "alternative" ], "last_updated": "2026-06-23" }, { "id": "amd-mi300a", "name": "AMD Instinct MI300A (APU)", "manufacturer": "AMD", "architecture": "CDNA 3", "memory": "128GB HBM3", "memory_bandwidth": "5.3 TB/s", "fp16_tflops": 1307, "fp8_tflops": 2614, "interconnect": "Infinity Fabric (896GB/s)", "tpb_watt": 760, "released": "2023-12-01", "status": "Shipping", "pricing": { "retail_usd": "TBD (system-level)", "cloud_per_hour": {} }, "availability": "Available — CPU+GPU APU used in El Capitan supercomputer (Lawrence Livermore)", "tags": [ "training", "amd", "apu", "supercomputer", "alternative" ], "last_updated": "2026-06-24" }, { "id": "amd-mi350", "name": "AMD Instinct MI350", "manufacturer": "AMD", "architecture": "CDNA 4 (3nm)", "memory": "288GB HBM3E", "memory_bandwidth": "8.0 TB/s", "fp16_tflops": 5000, "fp8_tflops": 10000, "interconnect": "Infinity Fabric (UPI)", "tpb_watt": 1000, "released": "2025-06-01", "status": "Shipping (limited initial allocation)", "pricing": { "retail_usd": "TBD", "cloud_per_hour": {} }, "availability": "Ramping — 185 billion transistors, 3nm process. Direct B200 competitor. Early deployments at Meta, Microsoft.", "tags": [ "training", "inference", "amd", "alternative", "flagship", "cdna4" ], "last_updated": "2026-06-23" }, { "id": "amd-mi325x", "name": "AMD Instinct MI325X", "manufacturer": "AMD", "architecture": "CDNA 3 (refresh)", "memory": "288GB HBM3E", "memory_bandwidth": "6.0 TB/s", "fp16_tflops": 1307, "fp8_tflops": 2614, "interconnect": "Infinity Fabric (896GB/s)", "tpb_watt": 750, "released": "2024-11-01", "status": "Shipping", "pricing": { "retail_usd": "TBD", "cloud_per_hour": { "coreweave": 2.49 } }, "availability": "Available — MI300X refresh with 288GB HBM3E for large-model inference. Competitive with H200 on memory.", "tags": [ "training", "inference", "amd", "alternative", "hbm3e" ], "last_updated": "2026-06-24" }, { "id": "amd-mi400", "name": "AMD Instinct MI400 (Announced)", "manufacturer": "AMD", "architecture": "CDNA 5 (next-gen)", "memory": "TBD (expected 384GB+ HBM4)", "memory_bandwidth": "TBD (expected 10+ TB/s)", "fp16_tflops": "TBD", "fp8_tflops": "TBD", "interconnect": "Infinity Fabric (next-gen)", "tpb_watt": "TBD (expected 1200W+)", "released": "2027-Q2 (expected)", "status": "Announced (not yet shipping)", "pricing": { "retail_usd": "TBD", "cloud_per_hour": {} }, "availability": "Not yet available — announced as Rubin competitor. Expected to feature HBM4 and compete with NVIDIA's next-gen Blackwell/Rubin.", "tags": [ "training", "inference", "amd", "alternative", "flagship", "upcoming" ], "last_updated": "2026-06-24" }, { "id": "intel-gaudi3", "name": "Intel Gaudi 3", "manufacturer": "Intel", "architecture": "Habana", "memory": "128GB HBM2e", "memory_bandwidth": "3.2 TB/s", "fp16_tflops": 1835, "fp8_tflops": 1835, "interconnect": "24x 200G RoCEv2", "tpb_watt": 600, "released": "2024-09-01", "status": "Shipping (limited)", "pricing": { "retail_usd": "125000 (8-GPU kit)", "cloud_per_hour": {} }, "availability": "Limited — IBM Cloud and AWS announced; adoption unclear", "tags": [ "training", "inference", "intel", "alternative", "limited" ], "last_updated": "2026-06-23" }, { "id": "intel-gaudi2", "name": "Intel Gaudi 2", "manufacturer": "Intel", "architecture": "Habana", "memory": "96GB HBM2e", "memory_bandwidth": "2.45 TB/s", "fp16_tflops": 729, "fp8_tflops": 1458, "interconnect": "24x 100G RoCEv2", "tpb_watt": 600, "released": "2022-10-01", "status": "Shipping (mature)", "pricing": { "retail_usd": "TBD", "cloud_per_hour": {} }, "availability": "Available — prior-gen Habana accelerator, deployed at Supermicro and Intel DevCloud", "tags": [ "training", "inference", "intel", "alternative", "legacy" ], "last_updated": "2026-06-24" }, { "id": "google-tpu-trillium", "name": "Google TPU Trillium (v6e)", "manufacturer": "Google", "architecture": "Trillium", "memory": "32GB HBM3", "memory_bandwidth": "1.6 TB/s", "fp16_tflops": 918, "fp8_tflops": 1836, "interconnect": "ICI (3D Torus)", "tpb_watt": "TBD", "released": "2024-10-01", "status": "Operational (GCP only)", "pricing": { "cloud_per_hour": { "gcp_ct6e": 2.7 } }, "availability": "GCP only — available to selected customers; expandable pods", "tags": [ "training", "inference", "tpu", "google", "alternative" ], "last_updated": "2026-06-23" }, { "id": "google-tpu-v5p", "name": "Google TPU v5p", "manufacturer": "Google", "architecture": "TPU v5 (Pod)", "memory": "95GB HBM3", "memory_bandwidth": "2.77 TB/s", "fp16_tflops": 459, "fp8_tflops": 918, "interconnect": "ICI (3D Torus, 6K pod)", "tpb_watt": "TBD", "released": "2023-12-01", "status": "Operational (GCP only)", "pricing": { "cloud_per_hour": { "gcp_v5p": 4.2 } }, "availability": "GCP only — largest TPU pod configuration (6,144 chips); available to selected customers", "tags": [ "training", "inference", "tpu", "google", "alternative", "pod" ], "last_updated": "2026-06-24" }, { "id": "google-tpu-v5e", "name": "Google TPU v5e", "manufacturer": "Google", "architecture": "TPU v5 (Lite)", "memory": "16GB HBM3", "memory_bandwidth": "0.81 TB/s", "fp16_tflops": 197, "fp8_tflops": 393, "interconnect": "ICI (2D Torus)", "tpb_watt": "TBD", "released": "2023-08-01", "status": "Operational (GCP only)", "pricing": { "cloud_per_hour": { "gcp_v5e": 1.2 } }, "availability": "GCP only — cost-efficient inference TPU for large-scale serving", "tags": [ "inference", "tpu", "google", "alternative", "cost-efficient" ], "last_updated": "2026-06-24" }, { "id": "google-tpu-v4", "name": "Google TPU v4", "manufacturer": "Google", "architecture": "TPU v4", "memory": "32GB HBM3", "memory_bandwidth": "1.2 TB/s", "fp16_tflops": 275, "fp8_tflops": "N/A", "interconnect": "ICI (3D Torus, 4K pod)", "tpb_watt": "TBD", "released": "2021-06-01", "status": "Operational (GCP only)", "pricing": { "cloud_per_hour": { "gcp_v4": 3.22 } }, "availability": "GCP only — prior-gen, still widely deployed for training and inference", "tags": [ "training", "inference", "tpu", "google", "alternative", "legacy" ], "last_updated": "2026-06-24" }, { "id": "google-ironwood", "name": "Google Ironwood (TPU v7)", "manufacturer": "Google", "architecture": "Ironwood", "memory": "TBD (expected 64GB+ HBM3e)", "memory_bandwidth": "TBD (expected 3+ TB/s)", "fp16_tflops": "TBD", "fp8_tflops": "TBD", "interconnect": "ICI (next-gen, expected)", "tpb_watt": "TBD", "released": "2027-Q1 (expected)", "status": "Announced (not yet shipping)", "pricing": { "cloud_per_hour": {} }, "availability": "Not yet available — Google's next-gen TPU after Trillium. Expected to power Gemini 3 era training.", "tags": [ "training", "inference", "tpu", "google", "alternative", "upcoming" ], "last_updated": "2026-06-24" }, { "id": "cerebras-cs3", "name": "Cerebras CS-3 (Wafer-Scale)", "manufacturer": "Cerebras Systems", "architecture": "WSE-3 (wafer-scale)", "memory": "44GB SRAM (on-die) + external", "memory_bandwidth": "21 PB/s (on-die)", "fp16_tflops": 125, "fp8_tflops": 250, "interconnect": "MemoryX + SwarmX", "tpb_watt": "TBD (system)", "released": "2024-03-01", "status": "Shipping (limited)", "pricing": { "retail_usd": "TBD (system-level)", "cloud_per_hour": {} }, "availability": "Limited — wafer-scale engine with 4 trillion transistors. Deployed at G42, Argonne National Lab, and select AI labs.", "tags": [ "training", "inference", "cerebras", "wafer-scale", "alternative", "innovative" ], "last_updated": "2026-06-24" }, { "id": "groq-lpu", "name": "Groq LPU (Inference Engine)", "manufacturer": "Groq", "architecture": "LPU (Tensor Streaming)", "memory": "TBD (SRAM-based)", "memory_bandwidth": "80 TB/s (on-die SRAM)", "fp16_tflops": 750, "fp8_tflops": 1500, "interconnect": "GroqLink (multi-chip)", "tpb_watt": "300", "released": "2024-02-01", "status": "Shipping (limited)", "pricing": { "retail_usd": "TBD (system-level)", "cloud_per_hour": {} }, "availability": "Limited — extreme inference throughput (tokens/sec) for LLMs. Deployed at GroqCloud for fast inference API.", "tags": [ "inference", "groq", "alternative", "innovative", "high-throughput" ], "last_updated": "2026-06-24" }, { "id": "sambanova-sn40l", "name": "SambaNova SN40L", "manufacturer": "SambaNova Systems", "architecture": "RDU (Reconfigurable Dataflow)", "memory": "64GB HBM3 + 1.5TB DDR5", "memory_bandwidth": "1.7 TB/s (HBM)", "fp16_tflops": 638, "fp8_tflops": 1275, "interconnect": "RDU-Connect", "tpb_watt": "TBD", "released": "2023-12-01", "status": "Shipping (limited)", "pricing": { "retail_usd": "TBD (system-level)", "cloud_per_hour": {} }, "availability": "Limited — dataflow architecture optimized for large-model inference. Deployed at Argonne National Lab and LLNL.", "tags": [ "training", "inference", "sambanova", "alternative", "reconfigurable" ], "last_updated": "2026-06-24" }, { "id": "tenstorrent-grayskull", "name": "Tenstorrent Grayskull", "manufacturer": "Tenstorrent", "architecture": "Wormhole (RISC-V)", "memory": "32GB LPDDR4X", "memory_bandwidth": "0.137 TB/s", "fp16_tflops": 233, "fp8_tflops": 466, "interconnect": "WH-Connect (Ethernet-based)", "tpb_watt": "200", "released": "2024-06-01", "status": "Shipping (limited)", "pricing": { "retail_usd": "1500 (dev kit)", "cloud_per_hour": {} }, "availability": "Limited — RISC-V based AI accelerator focused on cost-efficient training and inference. Developer kits available.", "tags": [ "training", "inference", "tenstorrent", "alternative", "risc-v", "low-cost" ], "last_updated": "2026-06-24" }, { "id": "huawei-ascend-910b", "name": "Huawei Ascend 910B", "manufacturer": "Huawei", "architecture": "Da Vinci (7nm)", "memory": "64GB HBM2e", "memory_bandwidth": "1.6 TB/s", "fp16_tflops": 320, "fp8_tflops": 640, "interconnect": "HCCS (300GB/s)", "tpb_watt": "400", "released": "2023-08-01", "status": "Shipping (China domestic only)", "pricing": { "retail_usd": "TBD (China only)", "cloud_per_hour": {} }, "availability": "China domestic market only — Huawei's primary AI accelerator replacing NVIDIA GPUs under US export controls", "tags": [ "training", "inference", "huawei", "china", "alternative", "export-controlled" ], "last_updated": "2026-06-24" }, { "id": "huawei-ascend-910c", "name": "Huawei Ascend 910C", "manufacturer": "Huawei", "architecture": "Da Vinci (enhanced)", "memory": "80GB HBM3", "memory_bandwidth": "2.0 TB/s", "fp16_tflops": 400, "fp8_tflops": 800, "interconnect": "HCCS (next-gen)", "tpb_watt": "450", "released": "2025-12-01", "status": "Shipping (China domestic only)", "pricing": { "retail_usd": "TBD (China only)", "cloud_per_hour": {} }, "availability": "China domestic market only — enhanced Ascend for frontier-scale training; reportedly used by DeepSeek and other Chinese AI labs", "tags": [ "training", "inference", "huawei", "china", "alternative", "export-controlled" ], "last_updated": "2026-06-24" } ]