A complete scenario for hw AI specialized
“looking for the way to attach in high definition..”
Vendor: Nvidia
Architecture: GPU
AI Focus: Everything (Training & Inference)
Claim to Fame: Performance, SW Stack, Ecosystem
Performance: The best for all AI models, per MLPerf
Market Focus: Everything AI
Vendor: AMD
Architecture: GPU
AI Focus: Everything
Claim to Fame: More HBM, More FLOPs
Performance: Excellent, esp for HPC
Market Focus: Everything AI
Vendor: Intel Gaudi
Architecture: ASIC
AI Focus: Everything
Claim to Fame: Price/Perf
Performance: Excellent
Market Focus: Enterprise AI
Vendor: Google TPU V4p
Architecture: ASIC
AI Focus: Everything
Claim to Fame: Google Gemini
Performance: Excellent
Market Focus: Internal Google workloads
Vendor: AWS Trainium2
Architecture: ASIC
AI Focus: Training
Claim to Fame: Low cost in AWS Cloud
Performance: Unknown. No Benchmarks
Market Focus: Enterprise small-scale training
Vendor: AWS Inferentia
Architecture: ASIC
AI Focus: Inference
Claim to Fame: Low cost in AWS Cloud
Performance: Unknown. No Benchmarks
Market Focus: Enterprise inference
Vendor: Microsoft Maia
Architecture: ASIC
AI Focus: Everything
Claim to Fame: Microsoft 360
Performance: Should be good
Market Focus: Microsoft 360 Apps, enterprise fine-tuning
Vendor: Qualcomm Cloud AI100
Architecture: ASIC
AI Focus: Inference
Claim to Fame: Best Perf/Watt
Performance: Needs GenAI Refresh
Market Focus: 2nd-tier Cloud
Vendor: Cerebras CS2
Architecture: Wafer-Scale ASIC
AI Focus: Training
Claim to Fame: EZ to Scale New Models
Performance: Excellent per node (@ >$2M)
Market Focus: 2nd-tier Cloud and Enterprise Training
Vendor: Groq LPU
Architecture: ASIC
AI Focus: Inference
Claim to Fame: Fast LLM Inference
Performance: > 300 tokens per second on Llama 2 70B
Market Focus: Inference As a Service
Vendor: SambaNova SN30 & SN40
Architecture: ASIC
AI Focus: Training
Claim to Fame: Train fast and efficiently
Performance: Unknown. No Benchmarks
Market Focus: Training as a Service
Vendor: Tenstorrent
Architecture: AI ASIC, RISC-V
AI Focus: IP for Custom chips
Claim to Fame: Jim Keller designs
Performance: Depends on Partner
Market Focus: Embedded Chips, Data Center