Tesla G80 is NVIDIA’s first unified shader architecture and the first GPU to expose the CUDA programming model. It introduces SIMT execution: scalar ALUs execute threads in fixed warps of 32, issued in lockstep by a single in-order warp scheduler per SM. Control-flow divergence is handled via warp masking, with no independent thread progress.

Each Texture Processor Cluster (TPC) includes a small L1 texture/read cache, while off-chip memory traffic is serviced via per–memory-partition L2 caches tightly coupled to ROP partitions. There is no general-purpose per-SM data cache; latency hiding relies on high warp occupancy and explicit shared memory usage.

• Compute capability: 1.0 (SM_10)
• Execution model: Warp-synchronous SIMT, single warp scheduler per SM

• SM-visible memory model:

  • Per-TPC L1 texture/read cache
  • Per memory-partition (ROP-coupled) L2 cache
  • Explicitly managed shared memory
• Memory support: GDDR3

CUDA / ISA feature set (SM_10)

• Baseline PTX and native ISA

• Integer and FP32 arithmetic only

• No atomic operations

• No warp vote / ballot instructions

• No FP64 (double precision)

• Strict global memory coalescing rules (half-warp, aligned, sequential)

• Denormalized FP values flushed to zero

• Architectural significance: Establishes NVIDIA’s SIMT execution model and baseline CUDA SM semantics