{ "schema_version": "touchdown.us-ai-chip-atlas.v2", "title": "United States AI Accelerator Registry", "updated_at": "2026-08-01", "scope": "Product-family records for AI accelerators designed by U.S.-headquartered companies. This label does not imply U.S. wafer fabrication, memory, packaging, assembly, or equipment origin. Version 2 adds explicit entity-level and software-evidence classifications without changing product specifications or inferring suppliers.", "comparison_focus": { "default_product_id": "nvidia-blackwell-b200", "nvidia_product_ids": [ "nvidia-blackwell-b200", "nvidia-blackwell-b300", "nvidia-gb200-nvl72", "nvidia-gb300-nvl72", "nvidia-vera-rubin-nvl72" ], "amd_product_ids": [ "amd-instinct-mi300x", "amd-instinct-mi325x", "amd-instinct-mi350x", "amd-instinct-mi355x" ], "scope": "The active comparison keeps NVIDIA B200, B300, GB200 NVL72, GB300 NVL72, and Vera Rubin NVL72 separate, and keeps AMD MI300X, MI325X, MI350X, and MI355X separate. Hopper is historical context and is not an active product path in this dated registry.", "level_boundary": "B200, B300, MI300X, MI325X, MI350X, and MI355X are accelerator records. GB200 NVL72, GB300 NVL72, and Vera Rubin NVL72 are rack-scale system or platform records. Rack aggregate numbers must not be compared with per-accelerator numbers." }, "products": [ { "id": "nvidia-blackwell-b200", "country": "United States", "company": "NVIDIA", "product": "Blackwell B200", "accelerator_type": "GPU/GPGPU", "role": "Datacenter training and inference GPU", "architecture": "Blackwell dual-die Tensor Core GPU used in HGX and DGX B200 systems", "process": { "node_nm": 4, "node_class": "advanced", "scope": "NVIDIA states that Blackwell uses a custom TSMC 4NP process", "state": "official" }, "compute": [ { "value": null, "unit": null, "precision": "FP4 through FP64", "scope": "Peak rate depends on precision, sparsity, and HGX or DGX configuration", "state": "official" } ], "memory": { "strategy": "hbm", "hbm_free": false, "type": "HBM3e", "capacity_gb": 180, "peak_bandwidth_tb_s": 8, "bandwidth_scope": "B200 SXM GPU specification; do not read the value as rack aggregate bandwidth", "state": "official" }, "interconnect": { "peak_gb_s": 1800, "scope": "Fifth-generation NVLink per B200 GPU in HGX and DGX; topology and aggregate rate depend on the system", "state": "official" }, "package": { "description": "Two reticle-sized Blackwell dies connected by a company-stated 10 TB/s die-to-die link in one HBM-equipped accelerator package; packaging supplier is not named", "state": "official" }, "power": { "value": 1000, "unit": "W", "scope": "Configurable maximum per B200 GPU in the reviewed HGX reference; not server, rack, or wall power", "state": "official" }, "system_hierarchy": [ "B200 GPU", "GB200 Grace Blackwell Superchip", "GB200 NVL72 rack" ], "software": [ "CUDA", "TensorRT-LLM", "NCCL", "Triton" ], "availability": { "state": "official", "status": "NVIDIA documents current HGX and DGX B200 systems; this does not establish unrestricted standalone-GPU availability or shipment volume" }, "supplier_relationships": [ { "stage_id": "logic_foundry_process", "supplier": "TSMC", "supplier_country": "Taiwan", "relationship_type": "manufactured_on_process_by", "supplier_relationship_state": "known", "evidence_state": "official", "source_urls": [ "https://nvidianews.nvidia.com/news/nvidia-blackwell-platform-arrives-to-power-a-new-era-of-computing" ], "summary": "NVIDIA states that Blackwell GPUs are manufactured on a custom TSMC 4NP process.", "boundary": "This establishes the named foundry and process for the Blackwell GPU. It does not identify the facility, wafer lot, masks, equipment, materials, yield, package, assembly site, shipment volume, or every component in an HGX or DGX system." } ], "independent_validation": "Specifications and availability are NVIDIA claims. Workload throughput, usable bandwidth, power, acceptance, and delivered configuration require system-specific receipts.", "sources": [ { "label": "NVIDIA HGX B200 specifications", "url": "https://docs.nvidia.com/enterprise-reference-architectures/hgx-ai-factory/latest/components.html", "state": "official" }, { "label": "NVIDIA Blackwell platform announcement", "url": "https://nvidianews.nvidia.com/news/nvidia-blackwell-platform-arrives-to-power-a-new-era-of-computing", "state": "official" }, { "label": "NVIDIA DGX B200", "url": "https://www.nvidia.com/en-us/data-center/dgx-b200/", "state": "official" } ], "unknowns": [ "HBM manufacturer by lot", "package and substrate supplier", "standalone B200 channel availability", "shipment volume", "sustained accepted-workload economics" ], "entity_level": "datacenter_accelerator_gpu", "entity_level_evidence_state": "derived", "entity_level_was_inferred": true, "entity_level_evidence_boundary": "This is a Touchdown-derived record-level category for comparing product families. It does not establish die count, package construction, supplier, manufacturing origin, qualification, shipment volume, or deployment scale.", "accelerator_type_state": "derived", "accelerator_type_was_inferred": false, "accelerator_type_boundary": "This is a Touchdown-normalized accelerator category based on the existing product record. It does not establish undocumented microarchitecture, instruction compatibility, software parity, workload performance, or manufacturing relationships.", "software_evidence_state": "derived", "software_evidence_boundary": "The named software list is normalized from the product record's reviewed source set. It is not yet a one-source-per-software-component claim map and does not prove operator coverage, framework parity, reproducibility, workload quality, or production reliability." }, { "id": "nvidia-blackwell-b300", "country": "United States", "company": "NVIDIA", "product": "Blackwell Ultra B300", "accelerator_type": "GPU/GPGPU", "role": "Datacenter training and inference GPU", "architecture": "Blackwell Ultra Tensor Core GPU used in HGX and DGX B300 systems", "process": { "node_nm": null, "node_class": "advanced", "scope": "The reviewed B300 sources do not state a B300-specific process or foundry", "state": "unknown" }, "compute": [ { "value": null, "unit": null, "precision": "FP4 through FP64", "scope": "Peak rate depends on precision, sparsity, and HGX or DGX configuration", "state": "official" } ], "memory": { "strategy": "hbm", "hbm_free": false, "type": "HBM3e", "capacity_gb": 288, "peak_bandwidth_tb_s": 8, "bandwidth_scope": "Per B300 GPU; do not read the value as eight-GPU or rack aggregate bandwidth", "state": "official" }, "interconnect": { "peak_gb_s": 1800, "scope": "Fifth-generation NVLink per B300 GPU in HGX and DGX; topology and aggregate rate depend on the system", "state": "official" }, "package": { "description": "SXM accelerator with HBM3e; the reviewed official sources do not establish the B300 package technology, substrate, foundry, or packaging supplier", "state": "official" }, "power": { "value": 1100, "unit": "W", "scope": "Configurable maximum per B300 GPU in the reviewed HGX reference; GB300 configurations differ and this is not server, rack, or wall power", "state": "official" }, "system_hierarchy": [ "B300 GPU", "GB300 module or Superchip scope", "GB300 NVL72 rack" ], "software": [ "CUDA", "TensorRT-LLM", "NCCL", "Triton" ], "availability": { "state": "official", "status": "NVIDIA describes DGX B300 as available and documents first-customer support; this remains a vendor availability claim" }, "independent_validation": "Specifications and availability are NVIDIA claims. Production volume, customer acceptance, sustained collectives, and accepted-workload economics remain separate receipts.", "sources": [ { "label": "NVIDIA HGX B300 specifications", "url": "https://docs.nvidia.com/enterprise-reference-architectures/hgx-ai-factory/latest/components.html", "state": "official" }, { "label": "NVIDIA DGX B300", "url": "https://www.nvidia.com/en-us/data-center/dgx-b300/", "state": "official" }, { "label": "NVIDIA DGX B300 user guide", "url": "https://docs.nvidia.com/dgx/dgxb300-user-guide/dgxb300-user-guide.pdf", "state": "official" } ], "unknowns": [ "B300-specific process and foundry", "HBM manufacturer by lot", "package and substrate supplier", "standalone B300 channel availability", "shipment volume", "sustained accepted-workload economics" ], "entity_level": "datacenter_accelerator_gpu", "entity_level_evidence_state": "derived", "entity_level_was_inferred": true, "entity_level_evidence_boundary": "This is a Touchdown-derived record-level category for comparing product families. It does not establish die count, package construction, supplier, manufacturing origin, qualification, shipment volume, or deployment scale.", "accelerator_type_state": "derived", "accelerator_type_was_inferred": false, "accelerator_type_boundary": "This is a Touchdown-normalized accelerator category based on the existing product record. It does not establish undocumented microarchitecture, instruction compatibility, software parity, workload performance, or manufacturing relationships.", "software_evidence_state": "derived", "software_evidence_boundary": "The named software list is normalized from the product record's reviewed source set. It is not yet a one-source-per-software-component claim map and does not prove operator coverage, framework parity, reproducibility, workload quality, or production reliability." }, { "id": "nvidia-gb200-nvl72", "country": "United States", "company": "NVIDIA", "product": "GB200 NVL72", "accelerator_type": "GPU/GPGPU", "role": "Liquid-cooled rack-scale training and inference system", "architecture": "Grace Blackwell rack with 72 Blackwell GPUs and 36 Grace CPUs in one NVLink domain", "process": { "node_nm": 4, "node_class": "advanced", "scope": "The rack contains Blackwell GPUs that NVIDIA states use custom TSMC 4NP; this is not a process claim for every component in the rack", "state": "official" }, "compute": [ { "value": null, "unit": null, "precision": "multiple", "scope": "Rack aggregate and workload result depend on precision, sparsity, software, cooling, and utilization", "state": "official" } ], "memory": { "strategy": "hbm-plus-lpddr", "hbm_free": false, "type": "HBM3e GPU memory plus LPDDR5X CPU memory", "capacity_gb": 13400, "peak_bandwidth_tb_s": 576, "bandwidth_scope": "Up-to rack aggregate HBM3e capacity and GPU-memory bandwidth; not per-GPU bandwidth", "state": "official" }, "interconnect": { "peak_gb_s": 130000, "scope": "Rack aggregate fifth-generation NVLink across 72 GPUs; direction, collectives, and delivered topology still matter", "state": "official" }, "package": { "description": "An NVL72 rack with 18 compute trays, nine NVLink switch trays, power shelves, bus bars, and liquid-cooling manifolds; it is not one chip package", "state": "official" }, "power": { "value": null, "unit": null, "scope": "Liquid-cooled rack; the reviewed record does not normalize delivered rack wall power, facility overhead, or cooling supplier", "state": "official" }, "system_hierarchy": [ "B200 GPU", "GB200 Grace Blackwell Superchip", "GB200 NVL72 rack" ], "software": [ "CUDA", "NCCL", "TensorRT-LLM", "NVIDIA AI Enterprise" ], "availability": { "state": "official", "status": "NVIDIA documents deployed and supported GB200 rack systems; shipment and installed-base totals are not established here" }, "supplier_relationships": [ { "stage_id": "logic_foundry_process", "supplier": "TSMC", "supplier_country": "Taiwan", "relationship_type": "contains_gpu_manufactured_on_process_by", "supplier_relationship_state": "known", "evidence_state": "official", "source_urls": [ "https://nvidianews.nvidia.com/news/nvidia-blackwell-platform-arrives-to-power-a-new-era-of-computing", "https://www.nvidia.com/en-us/data-center/gb200-nvl72/" ], "summary": "GB200 NVL72 contains Blackwell GPUs, and NVIDIA states that Blackwell GPUs are manufactured on custom TSMC 4NP.", "boundary": "This is a component-level foundry relationship for the Blackwell GPUs inside GB200. It is not a foundry claim for Grace CPUs, switches, memory, power, cooling, or every component in the rack, and it does not establish facility, lot, yield, or volume." } ], "independent_validation": "Rack power, cooling, fabric utilization, model quality, and accepted workload throughput must be measured together.", "sources": [ { "label": "NVIDIA GB200 NVL72", "url": "https://www.nvidia.com/en-us/data-center/gb200-nvl72/", "state": "official" }, { "label": "NVIDIA DGX GB200 hardware guide", "url": "https://docs.nvidia.com/dgx/dgxgb200-user-guide/hardware.html", "state": "official" }, { "label": "NVIDIA multi-node NVLink documentation", "url": "https://docs.nvidia.com/multi-node-nvlink-systems/index.html", "state": "official" }, { "label": "NVIDIA Blackwell platform announcement", "url": "https://nvidianews.nvidia.com/news/nvidia-blackwell-platform-arrives-to-power-a-new-era-of-computing", "state": "official" } ], "unknowns": [ "HBM and LPDDR manufacturers by rack", "package and substrate suppliers", "OEM supplier allocation", "shipment and installed-base totals", "delivered rack power", "sustained collective efficiency" ], "entity_level": "rack_scale_accelerator_system", "entity_level_evidence_state": "derived", "entity_level_was_inferred": true, "entity_level_evidence_boundary": "This is a Touchdown-derived record-level category for comparing product families. It does not establish die count, package construction, supplier, manufacturing origin, qualification, shipment volume, or deployment scale.", "accelerator_type_state": "derived", "accelerator_type_was_inferred": false, "accelerator_type_boundary": "This is a Touchdown-normalized accelerator category based on the existing product record. It does not establish undocumented microarchitecture, instruction compatibility, software parity, workload performance, or manufacturing relationships.", "software_evidence_state": "derived", "software_evidence_boundary": "The named software list is normalized from the product record's reviewed source set. It is not yet a one-source-per-software-component claim map and does not prove operator coverage, framework parity, reproducibility, workload quality, or production reliability." }, { "id": "nvidia-gb300-nvl72", "country": "United States", "company": "NVIDIA", "product": "GB300 NVL72", "accelerator_type": "GPU/GPGPU", "role": "Liquid-cooled rack-scale training and inference system", "architecture": "Grace Blackwell Ultra rack with 72 B300 GPUs and 36 Grace CPUs in one NVLink domain", "process": { "node_nm": null, "node_class": "advanced", "scope": "The reviewed official sources do not state a B300-specific process or foundry for the rack's GPUs", "state": "unknown" }, "compute": [ { "value": null, "unit": null, "precision": "multiple", "scope": "Rack aggregate and workload result depend on precision, sparsity, software, cooling, and utilization", "state": "official" } ], "memory": { "strategy": "hbm-plus-lpddr", "hbm_free": false, "type": "HBM3e GPU memory plus LPDDR5X CPU memory", "capacity_gb": 20000, "peak_bandwidth_tb_s": 576, "bandwidth_scope": "Company-stated rack HBM3e capacity and aggregate GPU-memory bandwidth; NVIDIA separately states 37 TB of fast memory including CPU memory", "state": "official" }, "interconnect": { "peak_gb_s": 130000, "scope": "Rack aggregate fifth-generation NVLink across 72 GPUs; ConnectX-8 scale-out and delivered topology remain system-level questions", "state": "official" }, "package": { "description": "An NVL72 rack-scale system, not one chip package; the reviewed official sources do not establish component package, substrate, or supplier provenance", "state": "official" }, "power": { "value": null, "unit": null, "scope": "Liquid-cooled rack; the reviewed record does not normalize delivered rack wall power, facility overhead, or cooling supplier", "state": "official" }, "system_hierarchy": [ "B300 GPU", "GB300 module or Superchip scope", "GB300 NVL72 rack" ], "software": [ "CUDA", "NCCL", "TensorRT-LLM", "NVIDIA AI Enterprise" ], "availability": { "state": "official", "status": "NVIDIA labels GB300 NVL72 available and documents first-customer support; this remains a vendor availability claim" }, "supplier_relationships": [ { "stage_id": "memory_manufacturer", "supplier": "SK hynix", "supplier_country": "South Korea", "relationship_type": "uses_hbm3e_from", "supplier_relationship_state": "known", "evidence_state": "official", "source_urls": [ "https://news.skhynix.com/en/gtc-2026-ai-partnership/" ], "summary": "SK hynix states that a live NVIDIA GB300 GPU module at GTC 2026 used its HBM3e.", "boundary": "This establishes at least one company-reported GB300 module integration. It does not prove exclusivity, the supplier mix or lot allocation for every rack, HBM capacity by lot, rack shipment volume, customer acceptance, or the LPDDR supplier." } ], "independent_validation": "Shipment volume, customer acceptance, delivered rack configuration, power, cooling, fabric utilization, and accepted workload throughput remain separate receipts.", "sources": [ { "label": "NVIDIA GB300 NVL72", "url": "https://www.nvidia.com/en-us/data-center/gb300-nvl72/", "state": "official" }, { "label": "NVIDIA DGX GB300", "url": "https://www.nvidia.com/en-us/data-center/dgx-gb300/", "state": "official" }, { "label": "NVIDIA Mission Control release notes", "url": "https://docs.nvidia.com/mission-control/docs/systems-quick-start-guide/2.1.0/nmc-release-notes.html", "state": "official" }, { "label": "SK hynix GTC 2026 NVIDIA memory disclosure", "url": "https://news.skhynix.com/en/gtc-2026-ai-partnership/", "state": "official" } ], "unknowns": [ "B300-specific process and foundry", "complete HBM and LPDDR supplier set and rack-level lot allocation", "package and substrate suppliers", "production and shipment volume", "delivered rack power", "sustained collective efficiency" ], "entity_level": "rack_scale_accelerator_system", "entity_level_evidence_state": "derived", "entity_level_was_inferred": true, "entity_level_evidence_boundary": "This is a Touchdown-derived record-level category for comparing product families. It does not establish die count, package construction, supplier, manufacturing origin, qualification, shipment volume, or deployment scale.", "accelerator_type_state": "derived", "accelerator_type_was_inferred": false, "accelerator_type_boundary": "This is a Touchdown-normalized accelerator category based on the existing product record. It does not establish undocumented microarchitecture, instruction compatibility, software parity, workload performance, or manufacturing relationships.", "software_evidence_state": "derived", "software_evidence_boundary": "The named software list is normalized from the product record's reviewed source set. It is not yet a one-source-per-software-component claim map and does not prove operator coverage, framework parity, reproducibility, workload quality, or production reliability." }, { "id": "nvidia-vera-rubin-nvl72", "country": "United States", "company": "NVIDIA", "product": "Vera Rubin NVL72", "accelerator_type": "GPU/GPGPU", "role": "Rack-scale AI platform with preliminary published specifications", "architecture": "Vera Rubin rack with 72 Rubin GPUs, 36 Vera CPUs, NVLink 6, ConnectX-9, BlueField-4, and Quantum-X800 or Spectrum-X networking", "specification_maturity": "preliminary", "process": { "node_nm": null, "node_class": "roadmap", "scope": "The reviewed NVIDIA sources do not name the Rubin or Vera process node or foundry", "state": "unknown" }, "compute": [ { "value": null, "unit": null, "precision": "multiple", "scope": "NVIDIA labels the published rack specifications preliminary and up-to; no independent production-system result is claimed", "state": "official" } ], "memory": { "strategy": "hbm-plus-lpddr", "hbm_free": false, "type": "HBM4 GPU memory plus LPDDR5X CPU memory", "capacity_gb": 20700, "peak_bandwidth_tb_s": 1580, "bandwidth_scope": "Preliminary up-to rack HBM4 capacity and aggregate bandwidth; per Rubin GPU NVIDIA states 288 GB and 22 TB/s", "state": "official" }, "interconnect": { "peak_gb_s": 260000, "scope": "Preliminary rack NVLink 6 switch bandwidth; NVIDIA states 3.6 TB/s per GPU and ConnectX-9 scale-out at 1.6 Tb/s per GPU", "state": "official" }, "package": { "description": "Preliminary NVL72 rack platform; final component package, substrate, assembly, and supplier details are not established", "state": "unknown" }, "power": { "value": null, "unit": null, "scope": "NVIDIA states 100% liquid cooling and a preliminary 45°C warm-water inlet condition; delivered rack wall power, facility overhead, and cooling supplier are not normalized", "state": "official" }, "system_hierarchy": [ "Rubin GPU", "Vera Rubin Superchip", "Vera Rubin NVL72 rack", "Vera Rubin POD platform" ], "software": [ "CUDA", "NCCL", "TensorRT-LLM", "NVIDIA AI Enterprise" ], "availability": { "state": "official", "status": "NVIDIA says Rubin is in full production and partner systems are planned for the second half of 2026; the product page still labels specifications preliminary, so broad customer availability is not claimed" }, "supplier_relationships": [ { "stage_id": "memory_manufacturer", "supplier": "SK hynix", "supplier_country": "South Korea", "relationship_type": "planned_hbm4_supplier_for", "supplier_relationship_state": "partial", "evidence_state": "official", "source_urls": [ "https://news.skhynix.com/en/gtc-2026-ai-partnership/" ], "summary": "SK hynix states that its HBM4 is slated to be adopted in NVIDIA Rubin and displayed the product on a Vera Rubin module at GTC 2026.", "boundary": "This is a company-stated planned adoption and demonstration, not proof of exclusivity, final supplier allocation, shipped production lots, customer acceptance, yield, volume, or the complete HBM4 and LPDDR supplier set." } ], "independent_validation": "No independently verified customer shipment, accepted rack, final specification, or production workload receipt is claimed.", "sources": [ { "label": "NVIDIA Vera Rubin NVL72", "url": "https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/", "state": "official" }, { "label": "NVIDIA Vera Rubin production release", "url": "https://nvidianews.nvidia.com/news/vera-rubin-full-production-agentic-ai-factory", "state": "official" }, { "label": "NVIDIA Rubin platform announcement", "url": "https://nvidianews.nvidia.com/news/rubin-platform-ai-supercomputer", "state": "official" }, { "label": "NVIDIA NVLink 6", "url": "https://www.nvidia.com/en-us/data-center/nvlink/", "state": "official" }, { "label": "SK hynix GTC 2026 NVIDIA memory disclosure", "url": "https://news.skhynix.com/en/gtc-2026-ai-partnership/", "state": "official" } ], "unknowns": [ "final non-preliminary specifications", "process and foundry", "complete HBM4 and LPDDR supplier set and lot allocation", "package and substrate suppliers", "independently verified customer shipments", "customer acceptance", "exact general-availability date" ], "entity_level": "rack_scale_ai_platform", "entity_level_evidence_state": "derived", "entity_level_was_inferred": true, "entity_level_evidence_boundary": "This is a Touchdown-derived record-level category for comparing product families. It does not establish die count, package construction, supplier, manufacturing origin, qualification, shipment volume, or deployment scale.", "accelerator_type_state": "derived", "accelerator_type_was_inferred": false, "accelerator_type_boundary": "This is a Touchdown-normalized accelerator category based on the existing product record. It does not establish undocumented microarchitecture, instruction compatibility, software parity, workload performance, or manufacturing relationships.", "software_evidence_state": "derived", "software_evidence_boundary": "The named software list is normalized from the product record's reviewed source set. It is not yet a one-source-per-software-component claim map and does not prove operator coverage, framework parity, reproducibility, workload quality, or production reliability." }, { "id": "amd-instinct-mi300x", "country": "United States", "company": "AMD", "product": "Instinct MI300X", "accelerator_type": "GPU/GPGPU", "role": "Datacenter training and inference GPU", "architecture": "CDNA 3 chiplet accelerator", "process": { "node_nm": 5, "node_class": "advanced", "scope": "AMD lists TSMC 5 nm and 6 nm process technology for MI300X; this does not identify facilities or lots", "state": "official" }, "compute": [ { "value": null, "unit": null, "precision": "FP8 through FP64", "scope": "Precision and matrix mode dependent", "state": "official" } ], "memory": { "strategy": "hbm", "hbm_free": false, "type": "HBM3", "capacity_gb": 192, "peak_bandwidth_tb_s": 5.3, "bandwidth_scope": "Per MI300X accelerator peak", "state": "official" }, "interconnect": { "peak_gb_s": null, "scope": "Infinity Fabric; topology and delivered rate depend on the system", "state": "official" }, "package": { "description": "OAM accelerator with chiplets and HBM3; package, substrate, and assembly suppliers are not identified here", "state": "official" }, "software": [ "ROCm", "HIP", "RCCL", "MIGraphX" ], "availability": { "state": "official", "status": "AMD publishes MI300X as a current product; supplier allocation, shipment volume, and cloud availability remain separate records" }, "independent_validation": "Use model-, software-version-, system-, and workload-specific tests.", "sources": [ { "label": "AMD Instinct MI300X", "url": "https://www.amd.com/en/products/accelerators/instinct/mi300/mi300x.html", "state": "official" } ], "unknowns": [ "HBM supplier and lot allocation", "package and assembly suppliers", "shipment volume", "sustained workload bandwidth", "customer-specific economics" ], "entity_level": "datacenter_accelerator_gpu", "entity_level_evidence_state": "derived", "entity_level_was_inferred": true, "entity_level_evidence_boundary": "This is a Touchdown-derived record-level category for comparing product families. It does not establish die count, package construction, supplier, manufacturing origin, qualification, shipment volume, or deployment scale.", "accelerator_type_state": "derived", "accelerator_type_was_inferred": false, "accelerator_type_boundary": "This is a Touchdown-normalized accelerator category based on the existing product record. It does not establish undocumented microarchitecture, instruction compatibility, software parity, workload performance, or manufacturing relationships.", "software_evidence_state": "derived", "software_evidence_boundary": "The named software list is normalized from the product record's reviewed source set. It is not yet a one-source-per-software-component claim map and does not prove operator coverage, framework parity, reproducibility, workload quality, or production reliability." }, { "id": "amd-instinct-mi325x", "country": "United States", "company": "AMD", "product": "Instinct MI325X", "accelerator_type": "GPU/GPGPU", "role": "Datacenter training and inference GPU", "architecture": "CDNA 3 chiplet accelerator", "process": { "node_nm": 5, "node_class": "advanced", "scope": "AMD lists TSMC 5 nm and 6 nm process technology for MI325X; this does not identify facilities or lots", "state": "official" }, "compute": [ { "value": null, "unit": null, "precision": "FP8 through FP64", "scope": "Precision and matrix mode dependent", "state": "official" } ], "memory": { "strategy": "hbm", "hbm_free": false, "type": "HBM3e", "capacity_gb": 256, "peak_bandwidth_tb_s": 6, "bandwidth_scope": "Per MI325X accelerator peak", "state": "official" }, "interconnect": { "peak_gb_s": null, "scope": "Infinity Fabric; topology and delivered rate depend on the system", "state": "official" }, "package": { "description": "OAM accelerator with chiplets and HBM3e; package, substrate, and assembly suppliers are not identified here", "state": "official" }, "software": [ "ROCm", "HIP", "RCCL", "MIGraphX" ], "availability": { "state": "official", "status": "AMD publishes MI325X as a current product; supplier allocation, shipment volume, and system availability remain separate records" }, "independent_validation": "Use model-, software-version-, system-, and workload-specific tests.", "sources": [ { "label": "AMD Instinct MI325X", "url": "https://www.amd.com/en/products/accelerators/instinct/mi300/mi325x.html", "state": "official" } ], "unknowns": [ "HBM supplier and lot allocation", "package and assembly suppliers", "shipment volume", "sustained workload bandwidth", "customer-specific economics" ], "entity_level": "datacenter_accelerator_gpu", "entity_level_evidence_state": "derived", "entity_level_was_inferred": true, "entity_level_evidence_boundary": "This is a Touchdown-derived record-level category for comparing product families. It does not establish die count, package construction, supplier, manufacturing origin, qualification, shipment volume, or deployment scale.", "accelerator_type_state": "derived", "accelerator_type_was_inferred": false, "accelerator_type_boundary": "This is a Touchdown-normalized accelerator category based on the existing product record. It does not establish undocumented microarchitecture, instruction compatibility, software parity, workload performance, or manufacturing relationships.", "software_evidence_state": "derived", "software_evidence_boundary": "The named software list is normalized from the product record's reviewed source set. It is not yet a one-source-per-software-component claim map and does not prove operator coverage, framework parity, reproducibility, workload quality, or production reliability." }, { "id": "amd-instinct-mi350x", "country": "United States", "company": "AMD", "product": "Instinct MI350X", "accelerator_type": "GPU/GPGPU", "role": "Datacenter training and inference GPU", "architecture": "CDNA 4 chiplet accelerator", "process": { "node_nm": 3, "node_class": "advanced", "scope": "AMD lists TSMC 3 nm and 6 nm process technology for MI350X; this does not identify facilities or lots", "state": "official" }, "compute": [ { "value": null, "unit": null, "precision": "FP4 through FP64", "scope": "Precision and matrix mode dependent", "state": "official" } ], "memory": { "strategy": "hbm", "hbm_free": false, "type": "HBM3e", "capacity_gb": 288, "peak_bandwidth_tb_s": 8, "bandwidth_scope": "Per MI350X accelerator peak", "state": "official" }, "interconnect": { "peak_gb_s": null, "scope": "Infinity Fabric; topology and delivered rate depend on the system", "state": "official" }, "package": { "description": "OAM accelerator with chiplets and HBM3e; package, substrate, and assembly suppliers are not identified here", "state": "official" }, "software": [ "ROCm", "HIP", "RCCL", "MIGraphX" ], "availability": { "state": "official", "status": "AMD publishes MI350X as a current product; supplier allocation, shipment volume, and system availability remain separate records" }, "independent_validation": "Production system and workload receipts remain separate from product-page specifications.", "sources": [ { "label": "AMD Instinct MI350X", "url": "https://www.amd.com/en/products/accelerators/instinct/mi350/mi350x.html", "state": "official" } ], "unknowns": [ "HBM supplier and lot allocation", "package and assembly suppliers", "volume shipments by partner", "sustained collectives", "accepted-task economics" ], "entity_level": "datacenter_accelerator_gpu", "entity_level_evidence_state": "derived", "entity_level_was_inferred": true, "entity_level_evidence_boundary": "This is a Touchdown-derived record-level category for comparing product families. It does not establish die count, package construction, supplier, manufacturing origin, qualification, shipment volume, or deployment scale.", "accelerator_type_state": "derived", "accelerator_type_was_inferred": false, "accelerator_type_boundary": "This is a Touchdown-normalized accelerator category based on the existing product record. It does not establish undocumented microarchitecture, instruction compatibility, software parity, workload performance, or manufacturing relationships.", "software_evidence_state": "derived", "software_evidence_boundary": "The named software list is normalized from the product record's reviewed source set. It is not yet a one-source-per-software-component claim map and does not prove operator coverage, framework parity, reproducibility, workload quality, or production reliability." }, { "id": "amd-instinct-mi355x", "country": "United States", "company": "AMD", "product": "Instinct MI355X", "accelerator_type": "GPU/GPGPU", "role": "Datacenter training and inference GPU", "architecture": "CDNA 4 chiplet accelerator", "process": { "node_nm": 3, "node_class": "advanced", "scope": "AMD lists TSMC 3 nm and 6 nm process technology for MI355X; this does not identify facilities or lots", "state": "official" }, "compute": [ { "value": null, "unit": null, "precision": "FP4 through FP64", "scope": "Precision and matrix mode dependent", "state": "official" } ], "memory": { "strategy": "hbm", "hbm_free": false, "type": "HBM3e", "capacity_gb": 288, "peak_bandwidth_tb_s": 8, "bandwidth_scope": "Per MI355X accelerator peak", "state": "official" }, "interconnect": { "peak_gb_s": null, "scope": "Infinity Fabric; topology and delivered rate depend on the system", "state": "official" }, "package": { "description": "OAM accelerator with chiplets and HBM3e; package, substrate, and assembly suppliers are not identified here", "state": "official" }, "software": [ "ROCm", "HIP", "RCCL", "MIGraphX" ], "availability": { "state": "official", "status": "AMD publishes MI355X as a current product; supplier allocation, shipment volume, and partner availability remain separate records" }, "independent_validation": "Production system and workload receipts remain separate from product-page specifications.", "sources": [ { "label": "AMD Instinct MI355X", "url": "https://www.amd.com/en/products/accelerators/instinct/mi350/mi355x.html", "state": "official" } ], "unknowns": [ "HBM supplier and lot allocation", "package and assembly suppliers", "volume shipments by partner", "sustained collectives", "accepted-task economics" ], "entity_level": "datacenter_accelerator_gpu", "entity_level_evidence_state": "derived", "entity_level_was_inferred": true, "entity_level_evidence_boundary": "This is a Touchdown-derived record-level category for comparing product families. It does not establish die count, package construction, supplier, manufacturing origin, qualification, shipment volume, or deployment scale.", "accelerator_type_state": "derived", "accelerator_type_was_inferred": false, "accelerator_type_boundary": "This is a Touchdown-normalized accelerator category based on the existing product record. It does not establish undocumented microarchitecture, instruction compatibility, software parity, workload performance, or manufacturing relationships.", "software_evidence_state": "derived", "software_evidence_boundary": "The named software list is normalized from the product record's reviewed source set. It is not yet a one-source-per-software-component claim map and does not prove operator coverage, framework parity, reproducibility, workload quality, or production reliability." }, { "id": "intel-gaudi-3", "country": "United States", "company": "Intel", "product": "Gaudi 3", "accelerator_type": "NPU/tensor ASIC", "role": "Datacenter training and inference accelerator", "architecture": "Matrix-math and tensor processor accelerator with integrated Ethernet", "process": { "node_nm": 5, "node_class": "advanced", "state": "official" }, "compute": [ { "value": null, "unit": null, "precision": "BF16/FP8", "scope": "official SKU table", "state": "official" } ], "memory": { "strategy": "hbm", "hbm_free": false, "type": "HBM2e", "capacity_gb": 128, "peak_bandwidth_tb_s": 3.7, "bandwidth_scope": "official peak", "state": "official" }, "interconnect": { "peak_gb_s": null, "scope": "Integrated Ethernet ports; topology depends on system", "state": "official" }, "package": { "description": "Accelerator package with HBM2e", "state": "official" }, "software": [ "SynapseAI", "PyTorch", "DeepSpeed" ], "availability": { "state": "official", "status": "Production accelerator available through systems and cloud instances" }, "independent_validation": "Software coverage and end-to-end workload fit remain decisive.", "sources": [ { "label": "Intel Gaudi", "url": "https://www.intel.com/content/www/us/en/products/details/processors/ai-accelerators/gaudi.html", "state": "official" } ], "unknowns": [ "current volume", "system-specific network efficiency", "accepted-task cost" ], "entity_level": "datacenter_accelerator_family", "entity_level_evidence_state": "derived", "entity_level_was_inferred": true, "entity_level_evidence_boundary": "This is a Touchdown-derived record-level category for comparing product families. It does not establish die count, package construction, supplier, manufacturing origin, qualification, shipment volume, or deployment scale.", "accelerator_type_state": "derived", "accelerator_type_was_inferred": false, "accelerator_type_boundary": "This is a Touchdown-normalized accelerator category based on the existing product record. It does not establish undocumented microarchitecture, instruction compatibility, software parity, workload performance, or manufacturing relationships.", "software_evidence_state": "derived", "software_evidence_boundary": "The named software list is normalized from the product record's reviewed source set. It is not yet a one-source-per-software-component claim map and does not prove operator coverage, framework parity, reproducibility, workload quality, or production reliability." }, { "id": "intel-arc-b-series", "country": "United States", "company": "Intel", "product": "Arc B-Series", "accelerator_type": "GPU/GPGPU", "role": "Client graphics and local AI GPU", "architecture": "Xe2 graphics architecture with XMX AI engines", "process": { "node_nm": null, "node_class": "advanced", "state": "official" }, "compute": [ { "value": null, "unit": null, "precision": "mixed", "scope": "SKU dependent", "state": "official" } ], "memory": { "strategy": "gddr", "hbm_free": true, "type": "GDDR6", "capacity_gb": null, "peak_bandwidth_tb_s": null, "bandwidth_scope": "SKU dependent", "state": "official" }, "interconnect": { "peak_gb_s": null, "scope": "PCIe client GPU", "state": "official" }, "package": { "description": "Discrete client GPU board", "state": "official" }, "software": [ "oneAPI", "OpenVINO", "XeSS" ], "availability": { "state": "official", "status": "Shipping client GPU family" }, "independent_validation": "Not a datacenter HBM accelerator; included to make the chip-type boundary visible.", "sources": [ { "label": "Intel Arc graphics", "url": "https://www.intel.com/content/www/us/en/products/docs/discrete-gpus/arc/desktop/b-series/overview.html", "state": "official" } ], "unknowns": [ "family-wide memory capacity", "sustained local-AI performance by model" ], "entity_level": "client_accelerator_family", "entity_level_evidence_state": "derived", "entity_level_was_inferred": true, "entity_level_evidence_boundary": "This is a Touchdown-derived record-level category for comparing product families. It does not establish die count, package construction, supplier, manufacturing origin, qualification, shipment volume, or deployment scale.", "accelerator_type_state": "derived", "accelerator_type_was_inferred": false, "accelerator_type_boundary": "This is a Touchdown-normalized accelerator category based on the existing product record. It does not establish undocumented microarchitecture, instruction compatibility, software parity, workload performance, or manufacturing relationships.", "software_evidence_state": "derived", "software_evidence_boundary": "The named software list is normalized from the product record's reviewed source set. It is not yet a one-source-per-software-component claim map and does not prove operator coverage, framework parity, reproducibility, workload quality, or production reliability." }, { "id": "cerebras-wse-3", "country": "United States", "company": "Cerebras", "product": "WSE-3", "accelerator_type": "Wafer-scale engine", "role": "Wafer-scale training and inference accelerator", "architecture": "Entire-wafer compute engine with distributed on-wafer SRAM", "process": { "node_nm": 5, "node_class": "advanced", "state": "official" }, "compute": [ { "value": 125, "unit": "PFLOPS", "precision": "AI", "scope": "company peak", "state": "official" } ], "memory": { "strategy": "on-wafer-sram", "hbm_free": true, "type": "On-wafer SRAM plus external memory systems", "capacity_gb": 44, "peak_bandwidth_tb_s": 21000, "bandwidth_scope": "company-stated on-wafer memory bandwidth, not off-package DRAM bandwidth", "state": "official" }, "interconnect": { "peak_gb_s": null, "scope": "SwarmX and MemoryX system fabric", "state": "official" }, "package": { "description": "Wafer-scale engine in a CS-3 system", "state": "official" }, "software": [ "Cerebras Software Platform", "PyTorch integration" ], "availability": { "state": "official", "status": "Production CS-3 systems and cloud service" }, "independent_validation": "On-wafer bandwidth is a different scope from HBM bandwidth and must not be compared without the data path.", "sources": [ { "label": "Cerebras WSE-3 announcement", "url": "https://www.cerebras.ai/press-release/cerebras-announces-third-generation-wafer-scale-engine/", "state": "official" } ], "unknowns": [ "workload-specific off-wafer traffic", "customer system utilization", "accepted-task economics" ], "entity_level": "wafer_scale_accelerator_system_family", "entity_level_evidence_state": "derived", "entity_level_was_inferred": true, "entity_level_evidence_boundary": "This is a Touchdown-derived record-level category for comparing product families. It does not establish die count, package construction, supplier, manufacturing origin, qualification, shipment volume, or deployment scale.", "accelerator_type_state": "derived", "accelerator_type_was_inferred": false, "accelerator_type_boundary": "This is a Touchdown-normalized accelerator category based on the existing product record. It does not establish undocumented microarchitecture, instruction compatibility, software parity, workload performance, or manufacturing relationships.", "software_evidence_state": "derived", "software_evidence_boundary": "The named software list is normalized from the product record's reviewed source set. It is not yet a one-source-per-software-component claim map and does not prove operator coverage, framework parity, reproducibility, workload quality, or production reliability." }, { "id": "groq-lpu", "country": "United States", "company": "Groq", "product": "LPU / GroqRack", "accelerator_type": "LPU/SRAM inference ASIC", "role": "Deterministic low-latency inference accelerator", "architecture": "Compiler-scheduled streaming processor with on-chip SRAM", "process": { "node_nm": 14, "node_class": "mature-node-class", "state": "official" }, "compute": [ { "value": null, "unit": null, "precision": "mixed", "scope": "chip and system dependent", "state": "unknown" } ], "memory": { "strategy": "on-chip-sram", "hbm_free": true, "type": "On-chip SRAM with model distributed across chips", "capacity_gb": null, "peak_bandwidth_tb_s": null, "bandwidth_scope": "internal memory and chip-to-chip path differ from HBM GPU systems", "state": "official" }, "interconnect": { "peak_gb_s": null, "scope": "Chip-to-chip streaming fabric", "state": "official" }, "package": { "description": "LPU cards and rack-scale systems", "state": "official" }, "software": [ "Groq compiler", "GroqCloud API" ], "availability": { "state": "official", "status": "Production cloud inference and rack offerings" }, "independent_validation": "Model support, quality, batch behavior, and end-to-end serving are workload-specific.", "sources": [ { "label": "Groq LPU architecture", "url": "https://groq.com/lpu-architecture", "state": "official" } ], "unknowns": [ "current chip revision by service", "system memory capacity", "customer volume" ], "entity_level": "accelerator_and_rack_system_family", "entity_level_evidence_state": "derived", "entity_level_was_inferred": true, "entity_level_evidence_boundary": "This is a Touchdown-derived record-level category for comparing product families. It does not establish die count, package construction, supplier, manufacturing origin, qualification, shipment volume, or deployment scale.", "accelerator_type_state": "derived", "accelerator_type_was_inferred": false, "accelerator_type_boundary": "This is a Touchdown-normalized accelerator category based on the existing product record. It does not establish undocumented microarchitecture, instruction compatibility, software parity, workload performance, or manufacturing relationships.", "software_evidence_state": "derived", "software_evidence_boundary": "The named software list is normalized from the product record's reviewed source set. It is not yet a one-source-per-software-component claim map and does not prove operator coverage, framework parity, reproducibility, workload quality, or production reliability." }, { "id": "google-tpu", "country": "United States", "company": "Google", "product": "TPU v5e / v5p / Trillium / Ironwood", "accelerator_type": "NPU/tensor ASIC", "role": "Cloud training and inference accelerator family", "architecture": "Tensor Processing Unit family deployed through Google Cloud", "process": { "node_nm": null, "node_class": "unknown", "state": "unknown" }, "compute": [ { "value": null, "unit": null, "precision": "multiple", "scope": "generation dependent", "state": "official" } ], "memory": { "strategy": "hbm", "hbm_free": false, "type": "HBM, generation dependent", "capacity_gb": null, "peak_bandwidth_tb_s": null, "bandwidth_scope": "generation dependent", "state": "official" }, "interconnect": { "peak_gb_s": null, "scope": "ICI pod topology, generation dependent", "state": "official" }, "package": { "description": "Cloud accelerator and pod; manufacturing origin not inferred", "state": "unknown" }, "software": [ "XLA", "JAX", "TensorFlow", "PyTorch/XLA" ], "availability": { "state": "official", "status": "Multiple production Google Cloud generations; newest generations have distinct preview/availability states" }, "independent_validation": "Keep each generation and cloud availability state separate in procurement decisions.", "sources": [ { "label": "Google Cloud TPU documentation", "url": "https://cloud.google.com/tpu/docs/system-architecture-tpu-vm", "state": "official" } ], "unknowns": [ "foundry and package by generation", "current regional capacity", "generation-specific fields not yet split" ], "entity_level": "cloud_accelerator_family", "entity_level_evidence_state": "derived", "entity_level_was_inferred": true, "entity_level_evidence_boundary": "This is a Touchdown-derived record-level category for comparing product families. It does not establish die count, package construction, supplier, manufacturing origin, qualification, shipment volume, or deployment scale.", "accelerator_type_state": "derived", "accelerator_type_was_inferred": false, "accelerator_type_boundary": "This is a Touchdown-normalized accelerator category based on the existing product record. It does not establish undocumented microarchitecture, instruction compatibility, software parity, workload performance, or manufacturing relationships.", "software_evidence_state": "derived", "software_evidence_boundary": "The named software list is normalized from the product record's reviewed source set. It is not yet a one-source-per-software-component claim map and does not prove operator coverage, framework parity, reproducibility, workload quality, or production reliability." }, { "id": "aws-trainium-inferentia", "country": "United States", "company": "Amazon Web Services", "product": "Trainium / Inferentia", "accelerator_type": "NPU/tensor ASIC", "role": "Cloud training and inference accelerator families", "architecture": "AWS NeuronCore accelerators exposed through EC2 instances", "process": { "node_nm": null, "node_class": "unknown", "state": "unknown" }, "compute": [ { "value": null, "unit": null, "precision": "multiple", "scope": "generation and instance dependent", "state": "official" } ], "memory": { "strategy": "hbm-or-external", "hbm_free": null, "type": "Generation dependent", "capacity_gb": null, "peak_bandwidth_tb_s": null, "bandwidth_scope": "instance and generation dependent", "state": "official" }, "interconnect": { "peak_gb_s": null, "scope": "NeuronLink/EFA, generation and instance dependent", "state": "official" }, "package": { "description": "Cloud-only accelerator systems; manufacturing origin not inferred", "state": "unknown" }, "software": [ "AWS Neuron SDK", "PyTorch", "JAX" ], "availability": { "state": "official", "status": "Production EC2 accelerator instances; individual generations vary" }, "independent_validation": "Cloud availability, compiler coverage, and accepted-task cost are the relevant receipts.", "sources": [ { "label": "AWS Trainium", "url": "https://aws.amazon.com/machine-learning/trainium/", "state": "official" }, { "label": "AWS Inferentia", "url": "https://aws.amazon.com/machine-learning/inferentia/", "state": "official" } ], "unknowns": [ "manufacturing chain", "generation-normalized specs", "regional supply" ], "entity_level": "cloud_accelerator_family", "entity_level_evidence_state": "derived", "entity_level_was_inferred": true, "entity_level_evidence_boundary": "This is a Touchdown-derived record-level category for comparing product families. It does not establish die count, package construction, supplier, manufacturing origin, qualification, shipment volume, or deployment scale.", "accelerator_type_state": "derived", "accelerator_type_was_inferred": false, "accelerator_type_boundary": "This is a Touchdown-normalized accelerator category based on the existing product record. It does not establish undocumented microarchitecture, instruction compatibility, software parity, workload performance, or manufacturing relationships.", "software_evidence_state": "derived", "software_evidence_boundary": "The named software list is normalized from the product record's reviewed source set. It is not yet a one-source-per-software-component claim map and does not prove operator coverage, framework parity, reproducibility, workload quality, or production reliability." }, { "id": "microsoft-maia", "country": "United States", "company": "Microsoft", "product": "Maia 100 / Maia 200", "accelerator_type": "NPU/tensor ASIC", "role": "Azure AI training and inference accelerator", "architecture": "Microsoft-designed datacenter AI accelerator integrated with Azure systems", "process": { "node_nm": null, "node_class": "advanced", "state": "official" }, "compute": [ { "value": null, "unit": null, "precision": "multiple", "scope": "generation dependent", "state": "unknown" } ], "memory": { "strategy": "hbm", "hbm_free": false, "type": "HBM-class", "capacity_gb": null, "peak_bandwidth_tb_s": null, "bandwidth_scope": "generation dependent", "state": "official" }, "interconnect": { "peak_gb_s": null, "scope": "Azure system fabric", "state": "official" }, "package": { "description": "Custom Azure accelerator package; supply-chain origin not inferred", "state": "unknown" }, "software": [ "Azure AI stack", "Triton" ], "availability": { "state": "official", "status": "Microsoft deployment/roadmap family; customer access varies by generation" }, "independent_validation": "Internal deployment does not equal broadly orderable customer capacity.", "sources": [ { "label": "Microsoft Maia 100", "url": "https://news.microsoft.com/source/features/ai/in-house-chips-silicon-to-service-to-meet-ai-demand/", "state": "official" } ], "unknowns": [ "broad customer availability", "manufacturing chain", "generation-specific production volume" ], "entity_level": "cloud_accelerator_family", "entity_level_evidence_state": "derived", "entity_level_was_inferred": true, "entity_level_evidence_boundary": "This is a Touchdown-derived record-level category for comparing product families. It does not establish die count, package construction, supplier, manufacturing origin, qualification, shipment volume, or deployment scale.", "accelerator_type_state": "derived", "accelerator_type_was_inferred": false, "accelerator_type_boundary": "This is a Touchdown-normalized accelerator category based on the existing product record. It does not establish undocumented microarchitecture, instruction compatibility, software parity, workload performance, or manufacturing relationships.", "software_evidence_state": "derived", "software_evidence_boundary": "The named software list is normalized from the product record's reviewed source set. It is not yet a one-source-per-software-component claim map and does not prove operator coverage, framework parity, reproducibility, workload quality, or production reliability." }, { "id": "meta-mtia", "country": "United States", "company": "Meta", "product": "MTIA", "accelerator_type": "NPU/tensor ASIC", "role": "Internal recommendation and inference accelerator", "architecture": "Meta Training and Inference Accelerator for internal workloads", "process": { "node_nm": null, "node_class": "advanced", "state": "official" }, "compute": [ { "value": null, "unit": null, "precision": "multiple", "scope": "revision dependent", "state": "official" } ], "memory": { "strategy": "lpddr", "hbm_free": true, "type": "LPDDR-class, revision dependent", "capacity_gb": null, "peak_bandwidth_tb_s": null, "bandwidth_scope": "revision dependent", "state": "official" }, "interconnect": { "peak_gb_s": null, "scope": "Internal system design", "state": "unknown" }, "package": { "description": "Internal datacenter accelerator; manufacturing origin not inferred", "state": "unknown" }, "software": [ "Meta internal compiler and runtime", "PyTorch integration" ], "availability": { "state": "official", "status": "Deployed internally; not a merchant accelerator" }, "independent_validation": "Internal production is a different commercial state from an orderable product.", "sources": [ { "label": "Meta next-generation MTIA", "url": "https://ai.meta.com/blog/next-generation-meta-training-inference-accelerator-AI-MTIA/", "state": "official" } ], "unknowns": [ "deployment volume", "current revision mix", "external availability not applicable" ], "entity_level": "internal_datacenter_accelerator_family", "entity_level_evidence_state": "derived", "entity_level_was_inferred": true, "entity_level_evidence_boundary": "This is a Touchdown-derived record-level category for comparing product families. It does not establish die count, package construction, supplier, manufacturing origin, qualification, shipment volume, or deployment scale.", "accelerator_type_state": "derived", "accelerator_type_was_inferred": false, "accelerator_type_boundary": "This is a Touchdown-normalized accelerator category based on the existing product record. It does not establish undocumented microarchitecture, instruction compatibility, software parity, workload performance, or manufacturing relationships.", "software_evidence_state": "derived", "software_evidence_boundary": "The named software list is normalized from the product record's reviewed source set. It is not yet a one-source-per-software-component claim map and does not prove operator coverage, framework parity, reproducibility, workload quality, or production reliability." } ], "normalization_boundary": "Derived classifications organize an existing product record. They do not prove foundry, memory, packaging, assembly, equipment, customer, qualification, shipment volume, or manufacturing origin.", "provenance": { "source_dataset": "./us-data-v1.json", "source_dataset_sha256": "c548d94c1a20462fde047185a6e1c775d2611c8e8a30a027bbb421044dcc4790", "normalized_product_count": 17 } }