Skip to content

GPUs & AI Accelerators arena · specs annex

GPU comparison

Current GPUs side by side — GeForce RTX 50-series and Radeon RDNA 4 on the consumer end, RTX PRO for workstations, and H200 / B200 / Instinct MI355X in the datacenter — on the raw numbers vendors publish. For the evidence-judged leaderboard over the same roster (user stories, citations, verdicts), see the GPUs & AI Accelerators arena.

Curated 2026-09-14 · 8 GPUs · every source URL checked reachable · hover any header or value for what it means and where it came from

Source
1GeForce RTX 5090

NVIDIA · Consumer (Blackwell)

32 GBGDDR71.8 TB/s3,352AI TOPSspec ↗
2GeForce RTX 5080

NVIDIA · Consumer (Blackwell)

16 GBGDDR7960 GB/s1,801AI TOPSspec ↗
3GeForce RTX 5070 Ti

NVIDIA · Consumer (Blackwell)

16 GBGDDR7896 GB/s1,406AI TOPSspec ↗
4Radeon RX 9070 XT

AMD · Consumer (RDNA 4)

16 GBGDDR6640 GB/s195TFLOPS FP16 matrixspec ↗
5RTX PRO 6000 Blackwell

NVIDIA · Workstation (Blackwell)

96 GBGDDR7 ECC1.8 TB/s4,000AI TOPSspec ↗
6H200 (SXM)

NVIDIA · Datacenter (Hopper)

141 GBHBM3e4.8 TB/s1,979TFLOPS FP8 densespec ↗
7B200 (per GPU, DGX B200)

NVIDIA · Datacenter (Blackwell)

180 GBHBM3e8 TB/sdrv4,500TFLOPS FP8 densespec ↗
8Instinct MI355X

AMD · Datacenter (CDNA 4)

288 GBHBM3E8 TB/s5,000TFLOPS FP8 densespec ↗

Honesty notes: the AI-perf column is not apples-to-apples — NVIDIA quotes consumer cards in FP4 “AI TOPS” with sparsity, AMD quotes FP16 matrix TFLOPS for Radeon, and datacenter parts are shown as dense FP8 TFLOPS; the unit is printed on every cell and sorting mixes them. Cells marked drv are derived from published system totals or memory specs (formula in the tooltip). n/a means unpublished, not zero. MSRPs are launch list prices where announced — street prices have varied widely.