• Home
  • Technology
  • Gaming
  • Entertainment
  • World & Business
  • Science
  • Sports
  • AI
HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
    • HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI
    AI

    Extropic Unveils Z1T Models for Sparse Hardware

    Extropic's official announcement of its first Z1T family of models.

    B(
    GV
    EX
    20 Sources, 26d ago, first seen 26d ago

    TLDR

    Extropic posted on X that it is introducing Z1T, described as the company's first family of transformer-like models built for sparse probabilistic hardware such as Z1. The post states the models achieve up to 140x energy efficiency gains over GPUs and reveal a new scaling law for sparse transformers. The company directs readers to its blog post for further details and includes a video attachment showing a dark red smoky background.

    Combined views

    631.9K

    20 Sources, first seen 26d ago

    Combined views

    631.9K

    20 Sources, first seen 26d ago

    4.7K likes
    4.7K likes
    217 comments
    1.6K saves
    381 reposts
    217 comments
    1.6K saves
    381 reposts

    Sentiment

    Positive81.1%18.9%Negative

    Summary

    Sentiment

    Positive81.1%18.9%Negative

    Many accounts welcomed Extropic’s Z1T for its energy efficiency gains and deep tech substance, while negative replies called the hardware impractical and its roadmap overly optimistic.

    Based on 57 sentiment-bearing replies from 53 accounts across 8 conversations.

    Summary

    Many accounts welcomed Extropic’s Z1T for its energy efficiency gains and deep tech substance, while negative replies called the hardware impractical and its roadmap overly optimistic.

    Based on 57 sentiment-bearing replies from 53 accounts across 8 conversations.

    Today's Rank

    —

    Not ranked yet

    Today's Rank

    —

    Not ranked yet

    20 Sources

    @extropicIntroducing Z1T: Our first family of transformer-like models made for sparse probabilistic hardware like Z1 Achieving up to 140x energy efficiency gains over GPUs and revealing a new scaling law for sparse transformers Read the blog: https://extropic.ai/writing/z1t
    @beffjezosWake up babe, a new Transformer for Thermoputers just dropped
    @alexocheemaCrazy stuff. This caught my eye in the technical report... Disaggregated inference between Thermodynamic Sampling Units (TSUs) and XPUs / FPGAs. As I understand it, the TSU is extremely energy efficient at the sparse neural computations this new sparse transformer they've built does so running all of those on the TSU and everything else on an XPU or FPGA is much more energy efficient. It's also faster.
    @MTSliveExtropic founder @beffjezos reveals how their chips achieve 140X the energy efficiency of GPUs by computing with the jitter of electrons: "This is not even a digital computer. We can even encode continuous numbers with a single pbit. If you average it, you can encode a number in the average, and the more samples from the average you get, the higher the effective precision." "Our pbits use about 10,000 times less transistors than if you were to emulate them digitally. It's because we actually use transistors completely differently. They're not in deterministic mode. They're not doing digital computations." "We're using the jitter of electrons that occurs in transistors to do the actual computation for us." "The layers that we port on Z1, we get about 1,000X energy efficiency gain. Because we don't port all the layers and some of the layers still have to run on an FPGA or XPU, then that processor is eating up most of the energy of the total workload. So 100X for the total workload in our case across an FPGA and Z1." @extropic
    @GillVerdRT @alexocheema: Crazy stuff. This caught my eye in the technical report... Disaggregated inference between Thermodynamic Sampling Units (…
    @firesidealphaBeff Jezos (Verdon) says Z1T proves you can build arbitrarily scalable models on a new hardware substrate, going from MNIST a year ago to GPT-2 today "It just shows it's possible to create arbitrarily scalable models on a new hardware substrate." "We released our first results about a year ago and it was like MNIST, which is the very early days, it's like the 80s of machine learning, and today we have models that are GPT-2 level that can run on our hardware." "And so algorithmic progress is blazing fast, and who knows how far we could push this." "This is not the end, like people might figure out ways to use our hardware to eat even more of the layers." "And look, if we can get to 1000X greater energy efficiency for inference sooner rather than later, that changes the whole world. That has a multi-trillion dollar impact on the economy."
    @gregosuriExtropic’s Z1T is one of the more interesting AI compute experiments I’ve seen recently. The next 1000x may come from changing the relationship between models and hardware entirely. Bullish.

    20 Sources

    @extropicIntroducing Z1T: Our first family of transformer-like models made for sparse probabilistic hardware like Z1 Achieving up to 140x energy efficiency gains over GPUs and revealing a new scaling law for sparse transformers Read the blog: https://extropic.ai/writing/z1t
    @beffjezosWake up babe, a new Transformer for Thermoputers just dropped
    @alexocheemaCrazy stuff. This caught my eye in the technical report... Disaggregated inference between Thermodynamic Sampling Units (TSUs) and XPUs / FPGAs. As I understand it, the TSU is extremely energy efficient at the sparse neural computations this new sparse transformer they've built does so running all of those on the TSU and everything else on an XPU or FPGA is much more energy efficient. It's also faster.
    @MTSliveExtropic founder @beffjezos reveals how their chips achieve 140X the energy efficiency of GPUs by computing with the jitter of electrons: "This is not even a digital computer. We can even encode continuous numbers with a single pbit. If you average it, you can encode a number in the average, and the more samples from the average you get, the higher the effective precision." "Our pbits use about 10,000 times less transistors than if you were to emulate them digitally. It's because we actually use transistors completely differently. They're not in deterministic mode. They're not doing digital computations." "We're using the jitter of electrons that occurs in transistors to do the actual computation for us." "The layers that we port on Z1, we get about 1,000X energy efficiency gain. Because we don't port all the layers and some of the layers still have to run on an FPGA or XPU, then that processor is eating up most of the energy of the total workload. So 100X for the total workload in our case across an FPGA and Z1." @extropic
    @GillVerdRT @alexocheema: Crazy stuff. This caught my eye in the technical report... Disaggregated inference between Thermodynamic Sampling Units (…
    @firesidealphaBeff Jezos (Verdon) says Z1T proves you can build arbitrarily scalable models on a new hardware substrate, going from MNIST a year ago to GPT-2 today "It just shows it's possible to create arbitrarily scalable models on a new hardware substrate." "We released our first results about a year ago and it was like MNIST, which is the very early days, it's like the 80s of machine learning, and today we have models that are GPT-2 level that can run on our hardware." "And so algorithmic progress is blazing fast, and who knows how far we could push this." "This is not the end, like people might figure out ways to use our hardware to eat even more of the layers." "And look, if we can get to 1000X greater energy efficiency for inference sooner rather than later, that changes the whole world. That has a multi-trillion dollar impact on the economy."
    @gregosuriExtropic’s Z1T is one of the more interesting AI compute experiments I’ve seen recently. The next 1000x may come from changing the relationship between models and hardware entirely. Bullish.