Anthony Liot

[Anthony Liot @wolfviking0]

I am thrilled to share with you our Realtime Photorealistic Virtual AI Sales Expert at the Samsung Experience Store in the Stanford Mall.

To bring this experience closer to you, here are some pictures of our virtual AI. 📸

[Itamar Golan 🤓 @ItakGol]

Whoa - LLAMA is now open-source for commercial use! 🎉

Introducing OpenLLAMA -

A team from Berkeley which is behind OpenLLAMA was inspired by LLAMA's exceptional performance and wanted to create a fully open-source version of it that is relevant for commercial use (Apache-2.0 license). So, They used a very similar training set, and also utilized identical training hyperparameters and model configurations as the original! 🤖

OpenLLAMA was trained with a whopping 200 billion tokens on the RedPajamas dataset, curated by the awesome team at @togethercompute . This dataset contains a staggering 1.2 trillion tokens, which is roughly the same number as LLAMA's. 🤯

Infrastructure-wise, to train OpenLLAMA, they used a cloud TPUv4 pod with data parallelism and FSDP or Zero3 to balance throughput and memory usage. Their training run achieved a throughput of over 1900 tokens/second/TPUv4 chip! 😮 You can see the training loss in the figure below. 📈

Benchmarking? They evaluated OpenLLAMA using the lm-evaluation-harness from @AiEleuther and found that it exhibits comparable performance to LLAMA and GPT-J in most tasks, and even outperforms them in some tasks! 🤯 The table is attached below. 📊

The team expects that the performance of OpenLLAMA will be even better once it completes its training on 1 trillion tokens. This current release is just a preview of what's to come! 🔜

👩‍💻 Code: https://t.co/PGyDsXlJoc
🏋‍ Weights: https://t.co/PYXoiN8G1w

[Alexander Zapryagaev @frustum]

Moving all scenegraph and render to the GPU was hard, but it was worth it. Now it is possible to generate geometry/nodes/objects/materials on the GPU, and nothing is going to the system memory to update BVH or other subsystems. Here is a single 2080Ti with 5 2k screens:

[Alexander Zapryagaev @frustum]

#GravityMark in Multiview mode, a single 2080Ti runs 5 2K and a single HD display at Linux. It can be 20 displays from a single PC in multi-GPU mode, only with #Tellusim. https://t.co/EWBQRoOrUV

[Alexander Zapryagaev @frustum]

Here are high-level entities of #Tellusim Engine. Everything is available on Windows/Linux/macOS/Android/iOS and can be easily embedded/integrated into your projects. Extreme performance and the same code for all platforms with great flexibility.

[chliot @chliot]

@elonmusk @Tesla I’ve just received my new Model X. This is a great car but there are two missing features that I miss from my old RAV4, the bird view 360 when I park and the blind spot window in the screen cannot be shifted. Any update coming soon?

[Emmanuel Macron @EmmanuelMacron]

Fiers de vous.

[Alexander Zapryagaev @frustum]

#Tellusim SDK 16 is ready with the following highlights: shader printf() for all platforms. Improved OIT with refraction, dispersion, and volumetric effects. FFT and real-time BC7 texture compression, resource export/import for GraphVarying.

[Alexander Zapryagaev @frustum]

We are using 21-bit floats for RGB FFT. That gives much better precision than 16-bit and saves memory bandwidth because all six channels can be packed into a single RGBAf32 texture.

[Alexander Zapryagaev @frustum]

#Sketchfab Download API plugin for #Tellusim Engine in action (https://t.co/y1vLjouKLg):

[Alexander Zapryagaev @frustum]

Thanks to the #shadertoy API, we have improved our Clay shader compiler. And now, almost all shadertoy shaders can run on VK, D3D, Metal, WGSL, and even Cuda.

[Alexander Zapryagaev @frustum]

The new #Tellusim SDK 13 with AMD ROCm API support, GPU soft-body simulation plugin, multi-GPU ray tracing, #GravityMark RT, objects asset library, GPU-marching cubes example, and SDF object rendering scene is ready.

[Alexander Zapryagaev @frustum]

Shell and hair simulation by #Tellusim engine.

[Alexander Zapryagaev @frustum]

Some of our basic WebGPU and WebGL tests are now available for everyone to use. At the moment, many of our compute tests fail, while others have unacceptable launching times. We will continue to add more tests with the progress of WebGPU. https://t.co/b01UY07Aur

[Alexander Zapryagaev @frustum]

All modern APIs support ray tracing now. Which one does it better? Find the answers here: https://t.co/Ypi6H2OmhQ
Performance comparison test of: Nvidia, AMD, Apple, Qualcomm; Direct3D12 vs Vulkan vs Metal; HW vs CS.

[Bryce, the CUDA Colonel @blelbach]

I'm not saying you have to like CMake, just that you should probably be using it.

If we continue to coalesce around it, we'll bring much needed uniformity to building C++ code.

Almost every IDE, tool, package manager, and C++ developer supports CMake.

[NEON @neondotlife]

We’ll be at #CES2021
#NEON #futureofwork

[UNIGINE @Unigine]

Sharing with you the realistic views of Antarctica created with UNIGINE 2.13. The author of the renders is Alexander Kolyasa, a 3D environment artist.

Link to the project: https://t.co/1uMIcIKJA8

#unigine #madewithunigine #realtime3d

[UNIGINE @Unigine]

Video overview of the UNIGINE 2.13 update:
https://t.co/ZIcaodJwdI
Please share your thoughts on which feature do you like most.
#unigine #gamedev #3D #RealTime

[Elon Musk @elonmusk]

Our NN is initially in Python for rapid iteration, then converted to C++/C/raw metal driver code for speed (important!). Also, tons of C++/C engineers needed for vehicle control & entire rest of car. Educational background is irrelevant, but all must pass hardcore coding test.