HARDWARE

→ View original post on X — @antgrasso

Chiplet Architecture Replaces Transistor Scaling for Heat Management

By

@antgrasso

–

14 March 2026 15h01

Transistor scaling is hitting thermal and cost ceilings, pushing chip design toward modular chiplets connected by high-speed links. With shared substrates and tailored process nodes, we manage heat and sustain performance without relying on miniaturization. Microblog @antgrasso

→ View original post on X — @ninadschick

Hyperscalers Leverage Nuclear Energy for AI Infrastructure

By

@ninadschick

–

14 March 2026 14h42

Godspeed to those engineers at RR, they have their work cut out for them. PS, in the U.S., the hyperscalers are treating energy as part of their tech stack. Microsoft is restarting 3-Mile Island, and Google / Meta et all are investing heavily in the next generation of SMR.

→ View original post on X — @elonmusk

AI race hinges on scaling power and chip output

By

@elonmusk

–

14 March 2026 14h39

The AI race will come down to scaling power and chip output

Qwen 3.5 27B Matches Sonnet 4.5 on Single RTX 5090

By

–

14 March 2026 14h23

That model, Qwen 3.5 27B Dense is equal to Sonnet 4.5 Runs great on a single RTX 5090 w/ full context But we are not anywhere near Opus 4.5 even with Qwen 3.5 397B MoE

Full Model GPU Offload Required to Avoid Performance Bottlenecks

By

–

14 March 2026 13h49

Partial offloads wouldn’t work because this will create a severe bottleneck That’s why the full model has to be offloaded to the GPU to see actual performance gains

GPU Offloading Offers 1800Gbps vs DGX Spark’s 273Gbps

By

–

14 March 2026 13h36

Yeah that is definitely faster because you’re fully offloading to the GPU which runs at ~1800Gbps in comparison to DGX Spark’s 273Gbps

New Website to Benchmark Hardware Tokens per Second Performance

By

–

14 March 2026 13h20

Update: @AlpinDale and I agreed to collab on a website that does more accurate calculations for your hardware tokens/sec performance Expect more on this after I am done with GTC this week

Qwen 3.5 MoE Model Recommendations for Spark and Mac Studio

By

–

14 March 2026 13h16

For a single Spark my first recommendation would be Qwen 3.5 122B MoE (Int4) For the Mac Studio I would recommend the 397B in 4-bit (and I think @ivanfioravanti would agree with me here)

Use MoE Models on Unified Memory Hardware Like DGX Spark

By

–

14 March 2026 13h11

As I have mentioned before, stop trying to get Dense models running on the DGX Spark/Mac Studios Unified Memory is best fit for MoEs because you only make each token go through a small subset of the numbers of parameters in the model Optimize for your hardware