Eric Quinnell

Founding Chip Architect, Stealth Startup. Fmr AWS Tranium, Tesla Dojo, CPUs (x86+ARM). PhD Computer Arithmetic

[Eric Quinnell @divBy_zero]

CPU thread def: PID+context+mem

GPU thread def: whatever you feel like

[Eric Quinnell @divBy_zero]

When did .md become the documentation default? Did I miss an election between .md and .docx or something?

At least LaTeX wasn’t on the ballot

[Eric Quinnell @divBy_zero]

@opinali @corsix Definitely not. It’s all hard. Software, hardware, all of it

[Eric Quinnell @divBy_zero]

@corsix @opinali This.

Also all the different EDA tools only support different random subsections of Verilog. Part of the fun is finding the sub-sub-set that actually works in them all without losing chip intent

[Eric Quinnell @divBy_zero]

@Leik0w0 It’s usually better for hardware systolic arrays. Hw doesn’t execute matmuls in the same physical directions as paper and pencil matrices - the row and col weights come from the same physical place, so it has to be row+row or col+col. Someone is doing the xpose somewhere

[Eric Quinnell @divBy_zero]

@insane_analyst He who controls the SPICE controls the universe

[Eric Quinnell @divBy_zero]

To be fair, at low enough FP quantization it does indeed return to associative behavior.

Bc it’s INT

[SemiAnalysis @SemiAnalysis_]

Floating point math is not associative! And many of the highest performance kernels split the workload among SMs and accumulate partial results in a nondeterministic order. Many AI labs just accept this, or pay a huge performance penalty for determinism. DeepSeek decided to do neither. (1/4) 🧵

[Eric Quinnell @divBy_zero]

@itsclivetime Still true

[Clive Chan @itsclivetime]

i believe it was @divBy_zero that wisely said that, contrary to common sense, it is always easier to fix performance problems in silicon than change an entrenched sw stack

[Eric Quinnell @divBy_zero]

@yacineMTB Confirmed anecdotal data point.

(Mid 40s, but close enough. This vibe coding stuff is legit)

[Eric Quinnell @divBy_zero]

Callout to the many leads and engineers who worked this over the years, esp @rawat_ritvik @aaronsrogers and Pete.

This will be a massive (and needed) upgrade to all cars and bots.

[Eric Quinnell @divBy_zero]

Congrats AI5 team, I know it was a rocky road

[Elon Musk @elonmusk]

Congrats to the @Tesla_AI chip design team on taping out AI5!

AI6, Dojo3 & other exciting chips in work.

[NASA @NASA]

Hello, Moon. It’s great to be back.

Here’s a taste of what the Artemis II astronauts photographed during their flight around the Moon. Check out more photos from the mission: https://t.co/rzM1P0QbOl

[Eric Quinnell @divBy_zero]

@LottoLabs @ptremblay Pedantically you are correct, yes. Way less loss than current quantization, and that detail would derail non technical folks ever further, so I didn’t split the hairs

[Eric Quinnell @divBy_zero]

@CliffLattner Yes, exactly. If doing inference, the compute will hide under the dram loads, even at high batch. For training, it is extra compute to pay.

[Eric Quinnell @divBy_zero]

Two days of weird takes, so I must:

“8x perf” is 32-bit baseline vs 4-bit compressed

“KV cache” is merely a use case and hard to capitalize full perf. And yes, many already compress here.

But it’s lossless. The others aren’t.

We should have been using this all along.

[Google Research @GoogleResearch]

Introducing TurboQuant: Our new compression algorithm that reduces LLM key-value cache memory by at least 6x and delivers up to 8x speedup, all with zero accuracy loss, redefining AI efficiency. Read the blog to learn how it achieves these results: https://t.co/CDSQ8HpZoc

[Eric Quinnell @divBy_zero]

My flight lands, I exit the plane, look for the Baggage Claim sign, and then there it was:

A 448Gpbs PAM4 Keysight waveform analyzer advertisement, on an LED backlight mega screen.

Thank you SJC

[Eric Quinnell @divBy_zero]

@GoogleResearch Classic compressor, seems obvious in retrospect, we should have all done this earlier.

Well done Google Research, truly

[Eric Quinnell @divBy_zero]

@rpoo @GoogleResearch Hahaha, this