Reviews that show their work.
Every review opens with what was checked: the specs, the pages, and whether a unit was ever on the bench.
Framework · workstation
buy-if
Framework Desktop: 128 GB that still waits on bandwidth
Soldered 128 GB on a Ryzen AI Max+ 395, in a 4.5 L Mini-ITX box. Capacity is the DGX Spark argument. AMD publishes 256 GB/s for Strix Halo, and there is no CUDA.
2026-09-22 · 7/10
NVIDIA · workstation
buy-if
RTX PRO 6000 Blackwell: 96 GB on a single workstation card
Triples the 5090's frame buffer to 96 GB ECC on the same Blackwell silicon. Same 1,792 GB/s bandwidth, so per-token throughput tracks the 5090 — you pay the multiple for capacity, ECC, and pro drivers.
2026-05-01 · 8/10
NVIDIA · gpu
buy-if
RTX 5080: 16 GB is the ceiling, not the floor
Half the 5090's VRAM at half the price. 16 GB caps you at 14B Q8 or 32B Q4 — the same ceiling a 4080 hit two years ago. A halo gaming card with AI as a side benefit.
2026-05-01 · $999 · 7/10
NVIDIA · workstation
buy-if
DGX Spark: NVIDIA in a desk-sized box
GB10 Grace Blackwell in a 1.2 kg chassis with 128 GB unified memory. Capacity wins, bandwidth loses. A development box, not a throughput box.
2026-05-01 · $4,699 · 7/10
Apple · workstation
buy-if
Mac Studio M3 Ultra: the patient context machine
Apple's M3 Ultra Mac Studio fits 100B+ class models a 5090 has to page or skip. First-token latency lags; total throughput wins when the model doesn't fit elsewhere. A specialist box, not a generalist.
2026-05-01 · $9,499 · 8/10
NVIDIA · gpu
buy
RTX 5090 for local LLM inference: the new watermark
32 GB VRAM and Blackwell sm_120, enough to run 32B-class models at high quants without paging or 70B at IQ3 with care. Worth the jump from a 4090 if you live in llama.cpp.
2026-04-24 · $1,999 · 9/10