Hardware

72-GPU Blackwell rack with 36 Grace CPUs. Hyperscale-only — relevant context here for understanding 'what frontier training runs on'.

Latest CDNA 4. 288GB HBM3e — currently the highest VRAM per chip on the market.

256GB HBM3e — direct competitor to NVIDIA H200 with more memory.

Datacenter Blackwell. 192GB HBM3e per chip, ~8 TB/s bandwidth. Cloud-tier — you rent these by the hour.

192GB HBM3 datacenter card. Used by Microsoft, Oracle, Meta cloud deployments.

Dual-card H100 with 188GB combined memory. Built for LLM serving.

Hopper refresh — 141GB HBM3e at ~4.8 TB/s. Datacenter-class; rentable on RunPod, Lambda, etc.

Intel's enterprise AI accelerator. 128GB HBM2e. Habana stack required — limited ecosystem support.

Previous-gen CDNA 2. 128GB HBM2e. Powered the Frontier supercomputer.

Pro Blackwell — 96GB GDDR7 ECC. The single-card answer to 70B and 100B+ local inference.

Previous-gen Habana accelerator. 96GB HBM2e.

PCIe Hopper. Lower power, lower bandwidth than SXM. Server-tier.

Hopper SXM5 — 80GB HBM3 at 3.35 TB/s. The original GPU that trained GPT-4. Cloud-rentable.

Ampere datacenter flagship. 80GB HBM2e at 2 TB/s. Still common at cloud providers.

64GB CDNA 2. Lower-power AMD datacenter option.

Ada-gen datacenter card. 48GB GDDR6 — popular at cloud GPU rentals as a budget H100 alternative.

Original Ada datacenter. Slower than L40S. 48GB GDDR6.

Pro Ada — 48GB ECC. Pre-Blackwell workstation default.

Ampere workstation/datacenter hybrid. 48GB GDDR6.

Ampere-gen workstation card with 48GB. Common in AI labs; used market is reasonable for 48GB at this point.

Original A100. 40GB HBM2 at 1.55 TB/s. Trained the early generation of frontier models.

Blackwell flagship. 32GB GDDR7 on a 512-bit bus delivers ~1.79 TB/s memory bandwidth — the new top of consumer hardware for local LLM infere

32GB workstation Ada. Mid-tier pro card.

Mobile Blackwell flagship. 24GB GDDR7 in a laptop is the new high-water mark.

Inference-focused Ada datacenter card. Low-power 24GB suitable for 7B-14B serving.

AMD's 24GB challenger to the 4090. ROCm Linux now solid for llama.cpp and vLLM. Best price-per-VRAM-GB on the new market.

Highest-tier Ampere consumer card. Used market gold for AI: 24GB at sub-$1200 in 2026.

The community-default high-end local-AI card from 2022 to 2025. 24GB GDDR6X at ~1 TB/s makes 70B Q4 comfortably loadable.

24GB Ampere workstation card. Tighter power envelope than RTX 3090.

The original 24GB CUDA value pick. Used market still strong in 2026 — many AI hobbyists run dual 3090 setups for 70B inference.

20GB RDNA 3. Cheaper alternative to XTX.

16GB Blackwell at the upper-mid price tier. Strong 14B–32B model performance.

Second-tier Blackwell. 16GB GDDR7, ~960 GB/s bandwidth. Fastest 16GB consumer card on the market.

16GB RDNA 4 at sub-$600. ROCm + Vulkan supported.

The 16GB sub-$500 sweet spot. Best value for entering local AI seriously.

RDNA 4 flagship. 16GB at $599 — best AMD value for local AI in 2026.

Refreshed 4080 with 16GB GDDR6X. Slightly behind 5080 but well-supported.

16GB upgrade of the 4070 Ti. Solid mid-high pick for local AI.

Sub-$330 16GB AMD. Memory-bandwidth-limited but great VRAM-per-dollar.

16GB RDNA 3 mid-range.

Mobile Ada flagship. 16GB VRAM in a laptop. Premium gaming and AI laptop default.

The poster child of 'cheap 16GB CUDA card'. Memory bandwidth is mediocre but 16GB at $400-something opens up 14B Q4.

Laptop variant of Ampere. 16GB VRAM in a portable form factor was rare and remains a sleeper pick on the used market.

Original 4080. 16GB GDDR6X. Still capable for 14B–32B Q4 work.

Alchemist 16GB. Cheapest path to that VRAM tier. Vulkan llama.cpp is the most-tested route.

Mid-range Blackwell with 12GB. 7B-14B Q4 territory.

Refreshed 4070. Strong mid-range value for 12GB-tier local AI.

Battlemage architecture. 12GB at $250 — the budget compute card. IPEX-LLM and Vulkan are usable paths for AI.

Original 4070. 12GB Ada. Now eclipsed by 4070 Super at the same price.

12GB RDNA 3.

12GB Ada — fits 7B–14B Q4 with usable context.

Mid-life 12GB refresh of the 3080. Decent 7B–14B card on the used market.

The community pick for 'cheapest CUDA card with serious VRAM'. The value floor for local AI in 2026.

10GB Battlemage at sub-$220. Entry budget compute.

Original 10GB 3080. Tight on VRAM for AI but still capable for 7B work.

Entry Blackwell. 8GB limits to 7B Q4 with limited context.

8GB Blackwell. Capable of 7B Q4 only — go 16GB SKU instead for AI work.

Entry-level Ada. 8GB limits to 7B Q4.

8GB version — go 16GB SKU for AI work.

8GB Ampere. Fits 7B Q4 only.

Combined CPU + GPU APU with 128GB unified HBM3. Powers the El Capitan supercomputer.

Desktop-replacement gaming/AI laptop with cooler thermals than ultraslims.

Top-end Windows AI laptop with 24GB RTX 5090 Mobile.

Ryzen 7 6800H + RTX 3080 16GB Mobile. The reference 'serious local-AI laptop' build. Look for the 16GB SKU.

Modular AMD laptop. Limited GPU but the platform is the appeal.

16-inch M4 Max — 128GB unified at 546 GB/s. The most capable AI laptop in 2026.

NVIDIA's desktop AI box — Grace Blackwell GB10 with 128GB unified LPDDR5X. The closest consumer can get to running 200B-class models locally

Top-spec Mac Studio with M3 Ultra. Up to 512GB unified memory in custom configs.

Two-chip Ultra fusing two M4 Max dies. Up to 256GB unified memory at 1.1 TB/s. The single highest-VRAM consumer rig you can buy in a Mac Stu

M3 Ultra — up to 512GB unified in Mac Studio top spec. 819 GB/s bandwidth.

Lower-tier Snapdragon X. 45 TOPS NPU.

Windows-on-ARM SoC with a 45 TOPS NPU. Limited LLM ecosystem in 2026 but improving via DirectML and ONNX paths.

Mid-tier M4 — 273 GB/s bandwidth, up to 48GB.

M4 Max — 546 GB/s memory bandwidth, up to 128GB unified. Most capable laptop SoC for 70B+ models.

M3 Max — 400 GB/s bandwidth, up to 128GB.

M2 Ultra — up to 192GB at 800 GB/s. Mac Studio and Mac Pro hosting models.

M2 Max — 400 GB/s bandwidth, up to 96GB.

Original Ultra — 800 GB/s. 64–128GB unified. Still capable for 70B Q4.

Original M1 Max. 400 GB/s. 32–64GB unified.