{"ok":true,"source":"tensorfeed.ai","lastUpdated":"2026-10-26","count":26,"hardware":[{"id":"rubin","name":"NVIDIA Rubin (Vera Rubin)","manufacturer":"NVIDIA","family":"Rubin","process":"Not disclosed by NVIDIA","released":"2026-08","memoryGB":288,"memoryBandwidthTBs":19.2,"fp16TFLOPS":4000,"fp8TFLOPS":17500,"fp4TFLOPS":35000,"tdpWatts":null,"interconnect":"NVLink 6 (3 TB/s per GPU) + Vera CPU NVLink-C2C (1.8 TB/s)","listPriceUSD":null,"availability":"Cloud and rack-scale (Vera Rubin NVL72: 72 Rubin GPUs + 36 Vera CPUs); early racks at CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure, Nebius","notes":"Blackwell successor. 288GB HBM4 per GPU. NVIDIA began production shipments in August 2026. FLOPS are NVIDIA dense figures (FP8 is FP8/FP6 training, FP4 is NVFP4 training); the 50 PFLOPS NVFP4 inference headline is a sparse figure and is not used here.","url":"https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/"},{"id":"b300","name":"NVIDIA B300 (Blackwell Ultra)","manufacturer":"NVIDIA","family":"Blackwell Ultra","process":"TSMC 4NP","released":"2025","memoryGB":288,"memoryBandwidthTBs":8,"fp16TFLOPS":2500,"fp8TFLOPS":5000,"fp4TFLOPS":15000,"tdpWatts":1400,"interconnect":"NVLink 5 (1.8 TB/s)","listPriceUSD":null,"availability":"Cloud and rack-scale (GB300 NVL72: 72 Blackwell Ultra GPUs + 36 Grace CPUs), listed by NVIDIA as available now","notes":"Mid-cycle Blackwell refresh: 208 billion transistors across two dies, with NVIDIA quoting 1.5x the HBM3e capacity and 1.5x the dense NVFP4 rate of Blackwell. Dense figures shown; NVIDIA rates the NVL72 rack at 1,440 PFLOPS NVFP4 with sparsity and 1.1 exaFLOPS dense.","url":"https://www.nvidia.com/en-us/data-center/gb300-nvl72/"},{"id":"gb200","name":"NVIDIA GB200 (Blackwell)","manufacturer":"NVIDIA","family":"Blackwell","process":"TSMC 4NP","released":"2025","memoryGB":372,"memoryBandwidthTBs":16,"fp16TFLOPS":5000,"fp8TFLOPS":10000,"fp4TFLOPS":20000,"tdpWatts":2700,"interconnect":"NVLink 5 (1.8 TB/s) + Grace CPU C2C (900 GB/s)","listPriceUSD":null,"availability":"Cloud only (CoreWeave, AWS, Azure, GCP, Oracle, Lambda)","notes":"Blackwell flagship: dual B200 + Grace CPU. FP4 native. The chip every major lab is buying for 2025-2026 frontier training.","url":"https://www.nvidia.com/en-us/data-center/gb200-nvl72/"},{"id":"b200","name":"NVIDIA B200","manufacturer":"NVIDIA","family":"Blackwell","process":"TSMC 4NP","released":"2025","memoryGB":180,"memoryBandwidthTBs":8,"fp16TFLOPS":2500,"fp8TFLOPS":5000,"fp4TFLOPS":10000,"tdpWatts":1000,"interconnect":"NVLink 5 (1.8 TB/s)","listPriceUSD":null,"availability":"Cloud only","notes":"Single-die Blackwell. 180GB HBM3e per GPU (1,440GB across an 8-GPU DGX B200). Inference workhorse for serving 100B+ MoE models on a single GPU.","url":"https://www.nvidia.com/en-us/data-center/blackwell-architecture/"},{"id":"h200","name":"NVIDIA H200","manufacturer":"NVIDIA","family":"Hopper","process":"TSMC 4N","released":"2024","memoryGB":141,"memoryBandwidthTBs":4.8,"fp16TFLOPS":989,"fp8TFLOPS":1979,"fp4TFLOPS":null,"tdpWatts":700,"interconnect":"NVLink 4 (900 GB/s)","listPriceUSD":null,"availability":"Cloud and direct sale","notes":"Hopper refresh with HBM3e. The sweet spot for production inference in 2025. Extra memory means 70B-class models fit single-GPU at FP8.","url":"https://www.nvidia.com/en-us/data-center/h200/"},{"id":"h100","name":"NVIDIA H100","manufacturer":"NVIDIA","family":"Hopper","process":"TSMC 4N","released":"2022","memoryGB":80,"memoryBandwidthTBs":3.35,"fp16TFLOPS":989,"fp8TFLOPS":1979,"fp4TFLOPS":null,"tdpWatts":700,"interconnect":"NVLink 4 (900 GB/s)","listPriceUSD":30000,"availability":"Cloud and direct sale (~$30k retail, declining)","notes":"The chip that built the 2023-2024 LLM boom. Still the most-deployed AI GPU. FP8 native. 80GB HBM3.","url":"https://www.nvidia.com/en-us/data-center/h100/"},{"id":"a100","name":"NVIDIA A100 80GB","manufacturer":"NVIDIA","family":"Ampere","process":"TSMC 7nm","released":"2020","memoryGB":80,"memoryBandwidthTBs":2,"fp16TFLOPS":312,"fp8TFLOPS":null,"fp4TFLOPS":null,"tdpWatts":400,"interconnect":"NVLink 3 (600 GB/s)","listPriceUSD":18000,"availability":"Cloud and used market (~$18k retail, declining)","notes":"Workhorse of the late-2010s ML wave. Still widely used for training smaller models and inference where H100/H200 supply is tight.","url":"https://www.nvidia.com/en-us/data-center/a100/"},{"id":"rtx-pro-6000-blackwell","name":"NVIDIA RTX PRO 6000 Blackwell","manufacturer":"NVIDIA","family":"Blackwell","process":"TSMC 4NP","released":"2025","memoryGB":96,"memoryBandwidthTBs":1.79,"fp16TFLOPS":360,"fp8TFLOPS":720,"fp4TFLOPS":1440,"tdpWatts":600,"interconnect":"PCIe 5.0","listPriceUSD":8500,"availability":"Direct sale (workstation card)","notes":"Workstation Blackwell. 96GB GDDR7. Good fit for on-prem agent dev environments where multi-H100 rentals are overkill.","url":"https://www.nvidia.com/en-us/products/workstations/professional-desktop-gpus/rtx-pro-6000-blackwell/"},{"id":"rtx-4090","name":"NVIDIA RTX 4090","manufacturer":"NVIDIA","family":"Ada Lovelace","process":"TSMC 4N","released":"2022","memoryGB":24,"memoryBandwidthTBs":1.01,"fp16TFLOPS":165,"fp8TFLOPS":660,"fp4TFLOPS":null,"tdpWatts":450,"interconnect":"PCIe 4.0","listPriceUSD":1600,"availability":"Consumer retail","notes":"Best price/perf for local agent dev. 24GB VRAM fits Q4-quantized 70B models. The default consumer-tier choice for self-hosted Ollama/llama.cpp.","url":"https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4090/"},{"id":"rtx-5090","name":"NVIDIA RTX 5090","manufacturer":"NVIDIA","family":"Blackwell","process":"TSMC 4NP","released":"2025","memoryGB":32,"memoryBandwidthTBs":1.79,"fp16TFLOPS":209,"fp8TFLOPS":838,"fp4TFLOPS":3352,"tdpWatts":575,"interconnect":"PCIe 5.0","listPriceUSD":2000,"availability":"Consumer retail","notes":"Consumer Blackwell flagship. 32GB GDDR7. FP4 native. Strong pick for local agent dev with current-frontier features.","url":"https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5090/"},{"id":"mi455x","name":"AMD MI455X","manufacturer":"AMD","family":"Instinct","process":"Not disclosed by AMD","released":"2026-07","memoryGB":432,"memoryBandwidthTBs":23.3,"fp16TFLOPS":5030,"fp8TFLOPS":20130,"fp4TFLOPS":40260,"tdpWatts":null,"interconnect":"UALink over Ethernet (3.6 TB/s bidirectional scale-up per GPU) plus 2,400 Gb/s scale-out","listPriceUSD":null,"availability":"Rack-scale (AMD Helios: 72 MI455X GPUs and 18 6th Gen EPYC CPUs per rack) via Bull, HPE, Lenovo and Supermicro","notes":"AMD's MI400-series flagship, launched at Advancing AI on July 23, 2026 with Helios racks already in production. 432GB HBM4. Per-GPU figures are MXFP4, MXFP8 and dense BF16 peaks; AMD publishes the rack numbers of 2.9 exaflops FP4, 1.4 exaflops FP8, 31TB HBM4 and 1.7 PB/s across 72 GPUs. OpenAI expects to bring Helios online starting in Q4 2026.","url":"https://www.amd.com/en/products/accelerators/instinct/mi400/mi455x.html"},{"id":"mi355x","name":"AMD MI355X","manufacturer":"AMD","family":"Instinct","process":"TSMC N3 / N6","released":"2025-06","memoryGB":288,"memoryBandwidthTBs":8,"fp16TFLOPS":2500,"fp8TFLOPS":5000,"fp4TFLOPS":10100,"tdpWatts":1400,"interconnect":"Infinity Fabric (7x 153 GB/s scale-up + 1x 128 GB/s scale-out)","listPriceUSD":null,"availability":"Cloud (Oracle Cloud Infrastructure, Crusoe; DigitalOcean spot)","notes":"AMD's MI350-series flagship. 288GB HBM3e with native MXFP4 and MXFP6 support. 1400W typical board power. The step between MI325X and the Helios-rack MI400 series.","url":"https://www.amd.com/en/products/accelerators/instinct/mi350/mi355x.html"},{"id":"mi350p","name":"AMD Instinct MI350P","manufacturer":"AMD","family":"Instinct","process":"TSMC N3 / N6","released":"2026-05","memoryGB":144,"memoryBandwidthTBs":4,"fp16TFLOPS":1150,"fp8TFLOPS":2300,"fp4TFLOPS":4600,"tdpWatts":600,"interconnect":"PCIe 5.0 x16 (no GPU-to-GPU Infinity Fabric links)","listPriceUSD":null,"availability":"Direct sale and OEM servers (dual-slot, full-length air-cooled PCIe card; announced May 2026)","notes":"CDNA 4 in a standard PCIe slot: 128 compute units, roughly half an MI350X, with 144GB HBM3e and a configurable 450W power mode. Peak dense figures shown; AMD doubles its 8-bit and 16-bit ratings with structured sparsity. Built to drop inference capacity into existing air-cooled enterprise servers instead of liquid-cooled racks.","url":"https://www.amd.com/en/products/accelerators/instinct/mi350/mi350p.html"},{"id":"mi325x","name":"AMD MI325X","manufacturer":"AMD","family":"Instinct","process":"TSMC N5 / N6","released":"2024-Q4","memoryGB":256,"memoryBandwidthTBs":6,"fp16TFLOPS":1300,"fp8TFLOPS":2600,"fp4TFLOPS":null,"tdpWatts":1000,"interconnect":"Infinity Fabric (8x 153 GB/s)","listPriceUSD":null,"availability":"Cloud (Microsoft Azure, Oracle), direct sale","notes":"AMD's answer to H200. 256GB HBM3e (highest in the catalog at single-die). Strong inference value when supply is constrained on NVIDIA.","url":"https://www.amd.com/en/products/accelerators/instinct/mi300/mi325x.html"},{"id":"mi300x","name":"AMD MI300X","manufacturer":"AMD","family":"Instinct","process":"TSMC N5 / N6","released":"2023-Q4","memoryGB":192,"memoryBandwidthTBs":5.3,"fp16TFLOPS":1300,"fp8TFLOPS":2600,"fp4TFLOPS":null,"tdpWatts":750,"interconnect":"Infinity Fabric","listPriceUSD":15000,"availability":"Cloud (Vultr, RunPod, Microsoft Azure), direct sale (~$15k)","notes":"Wider availability than NVIDIA flagship; cheaper to rent. ROCm software stack matures; vLLM and PyTorch first-class on MI300X in 2026.","url":"https://www.amd.com/en/products/accelerators/instinct/mi300/mi300x.html"},{"id":"tpu7x-ironwood","name":"Google TPU7x (Ironwood)","manufacturer":"Google","family":"TPU","process":"Not disclosed by Google","released":"2026-03","memoryGB":192,"memoryBandwidthTBs":7.38,"fp16TFLOPS":2307,"fp8TFLOPS":4614,"fp4TFLOPS":null,"tdpWatts":null,"interconnect":"ICI (1.2 TB/s bidirectional per chip), 9,216-chip pods","listPriceUSD":null,"availability":"Google Cloud only (TPU7x VMs; preview November 2025, GA March 2026)","notes":"Seventh-generation TPU and the first in the Ironwood family. 192 GiB HBM per chip, six times Trillium. Google positions it for large-scale training and inference of LLMs, MoE, and diffusion models.","url":"https://docs.cloud.google.com/tpu/docs/tpu7x"},{"id":"tpu-v6e-trillium","name":"Google TPU v6e (Trillium)","manufacturer":"Google","family":"TPU","process":"Not disclosed by Google","released":"2024-12","memoryGB":32,"memoryBandwidthTBs":1.64,"fp16TFLOPS":918,"fp8TFLOPS":918,"fp4TFLOPS":null,"tdpWatts":null,"interconnect":"ICI (800 GB/s bidirectional per chip), 256-chip pods","listPriceUSD":null,"availability":"Google Cloud only (GA December 2024)","notes":"Sixth-generation TPU. Twice the per-chip BF16 compute of v5p with a third of the memory; Google used it to train Gemini 2.0.","url":"https://docs.cloud.google.com/compute/docs/tpus/tpu-machines"},{"id":"tpu-v5p","name":"Google TPU v5p","manufacturer":"Google","family":"TPU","process":"TSMC 3nm","released":"2023-12","memoryGB":95,"memoryBandwidthTBs":2.77,"fp16TFLOPS":459,"fp8TFLOPS":459,"fp4TFLOPS":null,"tdpWatts":700,"interconnect":"ICI (Inter-Chip Interconnect, 1.2 TB/s bidirectional per chip)","listPriceUSD":null,"availability":"Google Cloud only (Vertex AI, GCE)","notes":"Google's previous-generation training TPU, now behind Trillium and Ironwood. Used for Gemini training. Only available on GCP; pod-scale (8960 chips) is a different beast than per-chip rentals.","url":"https://cloud.google.com/tpu/docs/v5p"},{"id":"tpu-v5e","name":"Google TPU v5e","manufacturer":"Google","family":"TPU","process":"TSMC 5nm","released":"2023-08","memoryGB":16,"memoryBandwidthTBs":0.8,"fp16TFLOPS":197,"fp8TFLOPS":393,"fp4TFLOPS":null,"tdpWatts":170,"interconnect":"ICI (400 GB/s bidirectional per chip)","listPriceUSD":null,"availability":"Google Cloud only","notes":"TPU inference tier. The 8-bit figure is Google's INT8 rating (393 TOPS). Cheaper than v5p; 256-chip pods. Best fit for Google Cloud customers serving Gemini-style inference at scale.","url":"https://cloud.google.com/tpu/docs/v5e"},{"id":"trainium-2","name":"AWS Trainium 2","manufacturer":"AWS","family":"Trainium","process":"TSMC 5nm","released":"2024-12","memoryGB":96,"memoryBandwidthTBs":2.9,"fp16TFLOPS":667,"fp8TFLOPS":1299,"fp4TFLOPS":null,"tdpWatts":500,"interconnect":"NeuronLink-v3 (1.28 TB/s per chip)","listPriceUSD":null,"availability":"AWS only (Trn2 instances)","notes":"AWS custom training/inference silicon. Dense figures shown; AWS rates it at 2,563 TFLOPS with sparsity. Anthropic uses these for Claude training. Cheaper TCO on AWS than NVIDIA flagship for committed workloads.","url":"https://aws.amazon.com/ai/machine-learning/trainium/"},{"id":"trainium-3","name":"AWS Trainium3","manufacturer":"AWS","family":"Trainium","process":"TSMC 3nm","released":"2025-12","memoryGB":144,"memoryBandwidthTBs":4.9,"fp16TFLOPS":671,"fp8TFLOPS":2517,"fp4TFLOPS":2517,"tdpWatts":null,"interconnect":"NeuronLink-v4 (2.56 TB/s per chip)","listPriceUSD":null,"availability":"AWS only (Trn3 instances)","notes":"Third-generation AWS training and inference silicon. 144 GB HBM3e, 4.9 TB/s bandwidth. 8-bit and 4-bit figures are MXFP8 and MXFP4 per AWS; BF16 is roughly flat with Trainium 2. Roughly 30-50% lower cost per hour than H100/H200 on committed AWS workloads.","url":"https://aws.amazon.com/ai/machine-learning/trainium/"},{"id":"inferentia-2","name":"AWS Inferentia 2","manufacturer":"AWS","family":"Inferentia","process":"TSMC 5nm","released":"2023-04","memoryGB":32,"memoryBandwidthTBs":0.82,"fp16TFLOPS":190,"fp8TFLOPS":null,"fp4TFLOPS":null,"tdpWatts":null,"interconnect":"NeuronLink","listPriceUSD":null,"availability":"AWS only (Inf2 instances)","notes":"AWS inference silicon. Lower compute than Trainium 2 but cheaper per-token for high-volume serving workloads on AWS.","url":"https://aws.amazon.com/ai/machine-learning/inferentia/"},{"id":"apple-m4-max","name":"Apple M4 Max","manufacturer":"Apple","family":"Apple Silicon","process":"TSMC 3nm","released":"2024-10","memoryGB":128,"memoryBandwidthTBs":0.546,"fp16TFLOPS":38,"fp8TFLOPS":null,"fp4TFLOPS":null,"tdpWatts":60,"interconnect":"On-die (unified memory)","listPriceUSD":4699,"availability":"Prior generation: Apple ships M5 Max (up to 128GB, 614 GB/s) in the MacBook Pro from March 11, 2026 and in the Mac Studio from September 22, 2026, where M5 Ultra reaches 512GB of unified memory at 1.2 TB/s","notes":"128GB unified memory means a 2024-2025 Mac Studio or MacBook Pro runs 70B-class models at Q4 quantization with usable speed. The dark-horse local-inference platform. Apple publishes memory and bandwidth for Apple Silicon but no TFLOPS rating for the M5 generation, so no M5 row is listed here.","url":"https://www.apple.com/newsroom/2024/10/apple-introduces-m4-pro-and-m4-max/"},{"id":"cerebras-wse-3-turbo","name":"Cerebras WSE-3 Turbo","manufacturer":"Cerebras","family":"WSE","process":"Not disclosed by Cerebras","released":"2026-08","memoryGB":44,"memoryBandwidthTBs":43200,"fp16TFLOPS":250000,"fp8TFLOPS":null,"fp4TFLOPS":null,"tdpWatts":null,"interconnect":"On-wafer fabric (53.5 PB/s), 2.4 Tb/s off-wafer I/O","listPriceUSD":null,"availability":"Cerebras Cloud and CS-4 systems (three WSE-3T wafers per CS-4 rack)","notes":"Announced August 18, 2026. Same 46,225 mm^2 wafer, 4 trillion transistors and 900,000 cores as WSE-3, but Cerebras doubles AI compute to 250 PFLOPS and memory bandwidth to 43.2 PB/s (shown here as 43,200 TB/s). A CS-4 rack holds three wafers for 750 PFLOPS. Compute figure is on the same basis as the 125 PFLOPS FP16 rating of WSE-3.","url":"https://www.cerebras.ai/chip"},{"id":"cerebras-wse-3","name":"Cerebras WSE-3","manufacturer":"Cerebras","family":"WSE","process":"TSMC 5nm","released":"2024","memoryGB":44,"memoryBandwidthTBs":21000,"fp16TFLOPS":125000,"fp8TFLOPS":null,"fp4TFLOPS":null,"tdpWatts":23000,"interconnect":"On-wafer (no interconnect needed)","listPriceUSD":null,"availability":"Cerebras Cloud only; superseded by WSE-3 Turbo in August 2026","notes":"Single 46,225 mm^2 wafer. Holds entire activations on-chip; eliminates many distributed-training pains. Cerebras rates aggregate memory bandwidth at 21 PB/s (shown here as 21,000 TB/s).","url":"https://cerebras.ai/product-chip/"},{"id":"groq-lpu","name":"Groq LPU","manufacturer":"Groq","family":"LPU","process":"GlobalFoundries 14nm (gen 1)","released":"2023","memoryGB":0.23,"memoryBandwidthTBs":80,"fp16TFLOPS":188,"fp8TFLOPS":null,"fp4TFLOPS":null,"tdpWatts":350,"interconnect":"GroqLink","listPriceUSD":null,"availability":"Groq Cloud only","notes":"Deterministic single-thread inference silicon. SRAM-only memory model means small per-chip capacity but massive bandwidth. Behind Groq's 700+ tokens/sec Llama 4 Scout serving.","url":"https://groq.com/lpu-language-processing-unit/"}]}