@pixelio

I ❤️ GPUs

 South Florida, USA
Joined September 2009
  • Tweets1.1K
  • Following1.9K
  • Followers1.1K
197Media
Vulkan `vkpeak` benchmark for the new Arduino Ventuno Q reaches 400+ FP32 GFLOPS. ./vkpeak 0 device = Adreno623 driver = turnip Mesa driver / Mesa 25.2.8-0ubuntu0.24.04.2 fp32-scalar = 417.44 GFLOPS fp32-vec4 = 438.42 GFLOPS fp16-scalar = 594.99 GFLOPS fp16-vec4 = 853.58 GFLOPS int32-scalar = 109.97 GIOPS int32-vec4 = 111.09 GIOPS int16-scalar = 283.50 GIOPS int16-vec4 = 294.14 GIOPS int64-scalar = 11.31 GIOPS int64-vec4 = 6.33 GIOPS int8-dotprod = 621.29 GIOPS copy-h2h = 10.11 GBPS copy-h2d = 11.41 GBPS copy-d2h = 9.12 GBPS copy-d2d = 13.66 GBPS @arduino
1
1
4
793
I’m late to the DGX Spark party but I finally got an NVIDIA Blackwell GB10 box!
1
4
441
Given the TPU buzz, here's a reminder: "Why systolic architectures?", H.T. Kung, Jan 1982 eecs.harvard.edu/~htk/public…
Allan MacKinnon
@pixelio22 May 2015
Kung, H.T. "Why systolic architectures?" Jan 1982: csis.bits-pilani.ac.in/facul…
18
117
17,553
Solid `vkpeak` results from the tiny Mali G310 GPU on an Amlogic S905X5M: 𝚏𝚙32-𝚜𝚌𝚊𝚕𝚊𝚛 = 52.58 𝙶𝙵𝙻𝙾𝙿𝚂 𝚏𝚙16-𝚟𝚎𝚌4 = 99.07 𝙶𝙵𝙻𝙾𝙿𝚂 𝚒𝚗𝚝8-𝚍𝚘𝚝𝚙𝚛𝚘𝚍 = 213.88 𝙶𝙸𝙾𝙿𝚂 Gist with `vulkaninfo` output here: gist.github.com/allanmac/5c1…

vulkaninfo ODROID C5 S905X5M Mali G310

vulkaninfo ODROID C5 S905X5M Mali G310. GitHub Gist: instantly share code, notes, and snippets.

gist.github.com
3
381
Good `vkpeak` benchmarks from an RK3588 Arm Mali G610 GPU: fp32-scalar = 467.89 GFLOPS fp32-vec4 = 496.97 GFLOPS fp16-scalar = 471.15 GFLOPS fp16-vec4 = 978.09 GFLOPS int8-dotprod = 1884.12 GIOPS
1
1
9
4,849
Given $INTC is in the news, it's nice to see that the latest LM Studio and the OpenAI gpt-oss-20b model run at ~47 tok/s on an A770 using Vulkan.
13
3
19
1,584
Fun with Google Sheets AI: =AI("Should I buy this stock? Only answer yes or no."). Every symbol returns, "I can't provide financial advice" except for NVIDIA and ADOBE where the answer is "No." I love it!🤷‍♂️
1
2
2
564
The terse Section 3.1 of the DeepSeek V3 paper will be interesting to CUDA ninjas. Hide communications overhead while keeping the 8x GPU node busy: arxiv.org/pdf/2412.19437
1
8
462
nextplatform.com/2025/01/27/…

How Did DeepSeek Train Its AI Model On A Lot Less – And Crippled – Hardware?

Maybe they should have called it DeepFake, or DeepState, or better still Deep Selloff. Or maybe the other obvious deep thing that the indigenous AI

nextplatform.com
2
286
Allan MacKinnon retweeted
Tim@timheuer1 Jan 2025
Last night I watched an elderly fam member navigate her iPhone for ‘important things and emails’. She is 96 and holding on to be tech relevant. Here is what I observed… 🧵
332
488
101
6,187
1,377,330
Excellent GNSS reception for a flexible antenna!
3
4
350
My dad found this button from his time working on the Lunar Rover. Apollo 15 was the first to use the moon buggy. Good ink, it hasn’t faded very much!
5
362
You all said “Numerical Linear Algebra, L. Trefethen, and D. Bau. SIAM, (1997)” is invaluable so I got the paperback (PDF out there too).
Numerical Linear Algebra, L. Trefethen, and D. Bau. SIAM, (1997)

ALT Numerical Linear Algebra, L. Trefethen, and D. Bau. SIAM, (1997)

7
464
Allan MacKinnon retweeted
HGPU group@hgpu9 Jul 2023
Optimization Techniques for GPU Programming #OpenCL #CUDA #Survey hgpu.org/?p=28413

Optimization Techniques for GPU Programming

In the past decade, Graphics Processing Units have played an important role in the field of high-performance computing and they still advance new fields such as IoT, autonomous vehicles, and exasca…

hgpu.org
3
5
689
Allan MacKinnon retweeted
François Chollet
@fchollet23 Jul 2022
I've been asked about this a lot, so let me provide a quick FAQ. Q: What's the nature of the issue? A: Anyone who has bought my book from Amazon in the past few month hasn't bought a genuine copy, but a lower-quality counterfeit copy printed by various fraudulent sellers.
François Chollet
@fchollet8 Jul 2022
Replying to @fchollet
For instance, if you go to the page of DLwP2 on Amazon, you see that it's being sold by a 3rd party seller named "Sacred Gamez". If you click "buy", you won't get the actual book from Manning. You get a low-quality counterfeit printed by the fraudulent seller (from the book PDF)
88
1,007
446
2,716
Allan MacKinnon retweeted
The Wall Street Journal
@WSJ14 Jul 2022
A former Wall Street bond trader now coaches the dominant team in competitive high school math. “I used to get paid money,” says Will Frazer. “Now we get trophies.” on.wsj.com/3cgzqBG

How a Public School in Florida Built America’s Greatest Math Team

A teacher obsessed with identifying talent, maximizing potential and optimizing education has created a dynasty. A former Wall Street bond trader, Buchholz High School coach Will Frazer says: ‘I used...

wsj.com
6
26
10
127
My very fast GPU-accelerated radix sort Vulkan library got pulled into the MESA RADV Vulkan driver: phoronix.com/scan.php?page=n… #vulkan #gpu #gpucompute #mesa #amd #radeon
3
8
DietGPU: fast specialized lossless compression on Nvidia GPUs If you have slow network fabric, this can speedup distributed training by a lot. Authored by Jeff Johnson (gh:wickedfoo) who wrote a lot of the PyTorch CUDA code and faiss-gpu. github.com/facebookresearch/…
4
38
1
248
Allan MacKinnon retweeted
HGPU group@hgpu31 Jan 2021
C-for-Metal: High Performance SIMD Programming on Intel GPUs #OpenCL #CUDA #Package hgpu.org/?p=24469

C-for-Metal: High Performance SIMD Programming on Intel GPUs

The SIMT execution model is commonly used for general GPU development. CUDA and OpenCL developers write scalar code that is implicitly parallelized by compiler and hardware. On Intel GPUs, however,…

hgpu.org
1
1
13
Load more