Arquitecturas Multimedia y de Propósito Específico

Encantado de anunciar q Paco Muñoz ha sido galardonado con uno de los premios "por sus contribuciones en el campo de la arq. de computadores y su alta relevancia internacional, que representan un paso muy importante en la investigación del diseño hardware" Muy merecido, @pakiko24

Charla en la FIUM: «Microarchitecture (Why no FUSS?) for Performance, energy and I forgot the security», por Biswabandan Panda um.es/web/informatic…

Elon Musk Buys Thousands of GPUs for Twitter's Generative AI Project tomshardware.com/news/elon-musk…

El fundador de la empresa Semidynamics, Roger Espasa, impartirá una charla en la Facultad de Informática sobre planificación de tareas en arquitecturas actuales. El seminario, que tendrá lugar el martes 28 de marzo a las 16:30 en el Salón de Actos. um.es/web/informatic…

Gordon Moore, Intel Co-Founder, Dies at 94 He famously forecast in 1965 that the number of transistors on an integrated circuit would double every year – a prediction that later came to be known as Moore’s Law intel.com/content/www/us… #HPC

Today, we lost a visionary. Gordon Moore, thank you for everything.

Want to learn more about how processors are fabricated? Check out this interesting article appeared in Xataka (in Spanish): xataka.com/componentes/as…

@mmc77gt Wow, impressive result! 👏🏻

@AMPE_UMU I reach a 37x speeup for the matrix multiply. Using FMA intrinsics and OMP on my 16 threads AMD 5800h.

Our Flexagon accelererator takes advandatage of both a unified merge-reduction network and a specific memory hierarchy to efficiently execute all the SpMSpM (sparse-sparse Matrix multiplication) dataflows. Check out our lighting talk: youtu.be/ozkY_d29QKM

For my machine learning / AI folks, the industry is begging you to use FP8. More than 30x improvements in FLOPS/Joule versus FP64, and more than 2x versus BF16. 3 exponent 2 mantisa format when? #ISSCC

Lisa Su is presenting performance improvements as 2x every 1.2 years, but with efficiency improving at only 2.2 years. @LisaSu says we could get to #Zettascale by ~2034ish if the industry continues this pace, but power would be ~500MW. "That's probably too much, but this is the #1 challenge facing the industry." IO scaling is becoming a massive limitation. Memory power continues to scale faster than any other component of SOC power. @AMD #ISSCC

.@AMD is a bit more conservative than @intel about the time frame for #Zettascale @LisaSu at #ISSCC2023: o #Supercomputers 2x perf every 1.2 years o Extrapolation implies Zetta by 2034 o Systems aren't small, systems aren't cheap! #HPC #AI via @IanCutress

Curious picture taken from blog.adobe.com/en/publish/202…

Brief history of vector architectures: redlineperf.com/vectors-how-th…

We've invited Roger Espasa (CEO of Semidynamics) to give a talk at @informaticaumu next March 28. We'll give you the details when the date is closer. Mark it in your calendar! Meanwhile, here is an interesting interview with Roger in Xataca xataka.com/p/237173

The Untapped Potential of the PS3’s CELL Processor And How Naughty Dog Tamed The Beast gamingbolt.com/the-untapped-p…

@AMPE_UMU PS3 cell CPU, a powerful cpu that combine 1 PowerPC cpu and 8 specific cores SPE. The SPE core was designed for single precision float Data. This video is in Spanish and very well explain. youtu.be/AF4Cd2ho9pk

One of the architectural styles we mentioned in today's class: "The 'Itanic'—Intel's ill-fated Itanium processor—finally sinks" networkworld.com/article/362845…

In today's class, we have learned about energy consumption. How much does it cost to train a model as large as GPT-3 (used by ChatGPT)? This interesting article talks about the huge energy costs of training DL models: numenta.com/blog/2022/05/2…