@ivanalog_com: I severely underestimated the 5070 Ti — Strata is a streaming architecture, where all the experts and KV caches have to flow repeatedly between RAM, PCIe, and VRAM. Like flowing water, any bottleneck in this pipeline will limit the flow rate. The importance of GPUs goes without saying, but from searching community benchmarks, for older platforms…

I severely underestimated the 5070 Ti,

Strata is a streaming architecture — all experts, KV, everything must flow repeatedly between memory, PCIe, and VRAM,

Like water in a pipeline. If there’s a bottleneck in the middle, the flow speed gets limited.

The importance of the GPU goes without saying, but searching community results shows that for older platforms, memory bandwidth and PCIe speed also have a very significant impact.

The same 5070 Ti card can show read/write performance that varies by up to 2x depending on the machine it’s on — but I previously only collected results from older machines.

The conclusion can be summarized as:

• Older platforms are heavily constrained by memory and PCIe — you need to go with 90-series large-VRAM cards
• Newer platforms can freely choose dual-card setups at lower cost — dual 5070 Ti with combined VRAM and compute theoretically matches a 5090, and in practice Strata can hit over 70% of a 5090’s performance

More details here 👇

seekinganythingbutalpha (@ivanalog_com): With Strata, the optimal-value hardware configuration for running Qwen Flash Next locally on a single machine has completely flipped.

The once highly competitive Mac M5M, M5U, etc. all lost the battle,

While NVIDIA’s top-tier GPUs + large memory approach achieved an overwhelming victory.

The 3090, once universally mocked as a mining card, has become hard to come by — and the 4090 and 5090 are even more sought-after.

Even dual 5060s and dual 3060s are now future-proof.