Skip to main content

AI Accelerator Chip Glossary

Performance Metrics

TFLOPS (Tera FLOPs)

Trillions of floating-point operations per second, measuring a chip's floating-point computing capability. Common precisions include:

  • FP64 (double precision): Scientific computing, HPC domain
  • FP32 (single precision): Traditional AI training precision
  • FP16 (half precision) / BF16: Mainstream mixed-precision training formats
  • FP8: Next-gen training/inference precision, supported by Blackwell and Hopper
  • FP4: Inference optimization precision, introduced with Blackwell architecture

Example: NVIDIA H100 SXM5 FP8 compute is 1,979 TFLOPS

TOPS (Tera Operations/s)

Trillions of integer operations per second, typically used for INT8 precision inference scenarios. INT8 compute is usually 2× FP16 and 4× FP32.

Example: Ascend 910B INT8 compute is 640 TOPS

Memory and Bandwidth

TermDescription
HBM (High Bandwidth Memory)High-bandwidth memory using 3D stacking technology for extreme bandwidth. Mainstream: HBM2e / HBM3 / HBM3e
GDDR (Graphics DDR)Graphics-dedicated memory, lower cost than HBM. Common in consumer and professional GPUs (GDDR6 / GDDR7)
Memory BandwidthBytes per second the memory can read/write, measured in GB/s. Critical for large model inference
SRAM (Static RAM)On-chip static cache, extremely fast but small capacity. Groq LPU uses 230MB on-chip SRAM instead of DRAM

Bandwidth formula: Bandwidth = Memory Frequency × Bus Width ÷ 8

Interconnect Technologies

TermDescription
NVLinkNVIDIA's proprietary high-speed GPU interconnect. 5th gen reaches 1.8 TB/s bidirectional bandwidth
NVLink-C2CNVIDIA chip-level interconnect for Grace CPU + Hopper GPU superchips
InfiniBandHigh-performance network interconnect standard used in AI clusters for cross-node communication (400Gb/s NDR mainstream)
PCIe (PCI Express)General peripheral interconnect, main interface between GPU and host. PCIe 5.0 x16 bandwidth ~64 GB/s
CXL (Compute Express Link)New CPU-memory/accelerator interconnect standard, based on PCIe physical layer
OAM (OCP Accelerator Module)Accelerator module form factor standard defined by the Open Compute Project

Compute Architectures

TermDescription
Tensor CoreNVIDIA GPU dedicated matrix operation unit, introduced starting with Volta architecture, now the core of AI computing
Transformer EngineDedicated Transformer acceleration unit in NVIDIA Hopper/Blackwell architectures, automatically managing FP8/FP16 precision switching
MIG (Multi-Instance GPU)NVIDIA A100/H100 GPU virtualization technology, partitioning one physical GPU into multiple independent instances
3D CubeMatrix compute unit in Huawei's Da Vinci architecture, purpose-built for matrix multiplication acceleration
TSP (Tensor Streaming Processor)Groq LPU processor architecture based on deterministic temporal execution, extremely low latency

Software Stack

TermDescription
CUDANVIDIA's parallel computing platform and programming model, de facto standard in AI computing
ROCmAMD's open source GPU compute platform, compatible with CUDA programming model
oneAPIIntel's unified programming model supporting heterogeneous CPU/GPU/FPGA computing
CANNHuawei Ascend AI computing framework, aligned with CUDA
MUSAMoore Threads GPU compute platform, compatible with CUDA API
cuDNNNVIDIA deep neural network acceleration library, provides optimized implementations for convolution, normalization, and other operators
TensorRTNVIDIA inference optimization engine, supports model quantization, layer fusion, and other optimizations
vLLMHigh-performance LLM inference engine, supports PagedAttention continuous batching
llama.cppLightweight LLM inference framework, supports CPU/GPU hybrid inference, focused on quantized model deployment

Deployment Architectures

TermDescription
SXM (Server eXpansion Module)NVIDIA data center GPU board-mount interface form factor, higher bandwidth than PCIe
NVL (NVLink)NVIDIA multi-GPU configuration connected via NVLink (e.g., H100 NVL dual-card)
SuperchipPacking CPU and GPU together via high-speed interconnect (e.g., NVIDIA Grace Hopper, GB200)
TDP (Thermal Design Power)Thermal design power in watts. In AI clusters: H100 ~700W, B200 ~1000W
HPC (High Performance Computing)High performance computing, typically referring to scientific computing rather than AI inference
TermDescription
LLM (Large Language Model)Large language model such as GPT-4, Llama 3, Qwen, etc.
MoE (Mixture of Experts)Mixture of experts architecture, splitting the model into multiple expert sub-networks, activating only relevant experts during inference to reduce computation
QuantizationCompressing model weights from FP16 to INT8/FP4/INT4, reducing memory usage and computation
DistillationTraining a small model using a large model, retaining most capability while dramatically reducing compute requirements
BatchProcessing multiple inference requests simultaneously to improve GPU utilization and throughput
TTFT (Time to First Token)First token latency, key metric for measuring inference response speed
TPOT (Time per Output Token)Time to produce each output token, key metric for measuring inference throughput

Chip Classification

ClassificationFull NameTypical Application
GPUGraphics Processing UnitAI training and inference (broadest generality)
NPUNeural Processing UnitEdge AI inference, edge computing
TPUTensor Processing UnitTraining and inference within Google ecosystem
LPULanguage Processing UnitOptimized for LLM inference
IPUIntelligence Processing UnitAI training accelerator designed by Graphcore
DPUData Processing UnitData center networking and data offload
FPGAField-Programmable Gate ArrayReconfigurable AI inference/signal processing
ASICApplication-Specific ICDedicated AI training/inference acceleration