tachyon · GPU Accelerated DataFrame Library

tachyon is a fast, GPU-native DataFrame engine for analytics and machine-learning workloads. It uses a columnar execution model with runtime-compiled (JIT) GPU kernels to deliver high throughput and optimized execution.

Features

• JIT-compiled GPU kernels for dynamic, hardware-specialized execution.
• Half-precision numeric support (f16, bf16) for reduced memory footprint and improved performance.
• Columnar DataFrame engine inspired by Apache Arrow.
• Vectorized math operations, filtering, and expression evaluation.

Setup, Build, Test

git clone https://github.com/neocrafttech/tachyon.git
cd tachyon
./bolt.sh setup
./bolt.sh build
./bolt.sh test

Sample Usage

Parse and inspect an expression

use compute::data_type::DataType;
use compute::error::ErrorMode;
use compute::expr::SchemaContext;
use compute::parser::parse_scheme_expr;

let expr = parse_scheme_expr("(+, a, 10)")?;
let schema = SchemaContext::new()
    .with_column("a", DataType::I32)
    .with_error_mode(ErrorMode::Tachyon);

assert_eq!(expr.infer_type(&schema)?, DataType::I32);

Evaluate on GPU

use std::sync::Arc;

use compute::column::{Column, VecArray};
use compute::data_type::DataType;
use compute::error::ErrorMode;
use compute::evaluate::{Device, evaluate};
use compute::expr::Expr;
use compute::operator::Operator;

// Requires the `gpu` feature and a CUDA-capable environment.
let input = Column::<u64>::new(
    "a",
    Arc::new(VecArray { data: vec![1_i32, 2, 3], datatype: DataType::I32 }),
    None,
);
let expr = Expr::binary(Operator::Add, Expr::col("a"), Expr::i32(5));
let output = evaluate(Device::GPU, ErrorMode::Tachyon, &expr, &[input]).await?;

assert_eq!(output[0].data_as_slice::<i32>(), Some(&[6, 7, 8][..]));