🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Ordered a RTX 5090 🤤

🤣 the lost art of shipping binaries

Required memory bandwidth --> determines chip perimeter (shoreline) --> determines minimum chip size

Funny how -20% drop in a day is a shock these days 😂💸 .

I predict GPU accelerated STWO will get even more crazy fast within 6 months!

Le Chat

1/n What’s the speed limit for GPUs? Why do some teams achieve 10x proving speedups while others hit 100x on the same hardware and algorithms? NVIDIA calls it "Speed of Light Analysis", a topic near and dear to my heart. Let’s unpack it.

NVIDIA GB10 - 10P + 10E core ARM CPU - RTX 5070 class GPU - 128GB Unified Memory Now put this in a 16-inch laptop with a Sensel haptic touchpad. Take my money 💸

👀 Initial testing on RTX 5090... Real world data transfer speeds over PCIe 5.0 - 52GB/s It's so fast that on a desktop machine, you start worrying that your system RAM won't be fast enough to keep up. Make sure you turn on XMP/EXPO on your DDR5! #rtx5090

My new development workhorse: - Ryzen 9950x (16-core Zen 5, full AVX512 datapath!) - 192GB of DDR5 - 100GbE Nvidia Connectx-6 Network card (running at 25Gb since I want all my PCIe lanes to the GPU) - Sipeed NanoKVM for remote management Just waiting for my RTX 5090 now

A pretty nice #locallama machine! Looking forward to what Alex Cheema - e/acc does with this one 🙂. Note for cryptography, in particular #ZKProofs : - 128GB unified memory eliminates CPU<>GPU transfer bottleneck - GPU about on par with RTX 4070 - However, AMD GPUs is half as efficient

So this is where the entire supply went 😂

The amount of open research Ingonyama does is awesome! Check out more Ingo papers here: github.com/ingonyama-zk/p…

The new Mac Studio is the perfect machine for running Deepseek R1 locally. Now we await Nvidia's answer, if they care to.

Super excited about Mojo language. - Ever wanted to write GPU kernels in Python? - Easily leverage SIMD without the code turning into a horrible mess? - Import regular Python libraries into your high performance compiled code without FFI boilerplate?

Hardware-friendliness of HyperPlonk, Part2 hackmd.io/@Ingonyama/Har…

Is this the first time that AMD will be on the leading edge node before Apple?

My RTX 5090 finally arrived!! 🤤 1.8 TB/s memory bandwidth!

First benchmarks on the NVIDIA 5090 are in. Most meaningful data point: Groth16 on 5090 is now 1.5× faster compared to 4090. Same code, no GPU-specific tuning. Measured with ICICLE-Snark, the current state-of-the-art Groth16 GPU implementation. H/t Emir Soytürk , Tony Wu

Amazing! GPU accelerated RTL simulator. Quick, someone wrap Cocotb around this!