d-Matrix's Squadron platform scales to 16 racks and 1,024 Corsair XPUs, marking a move into integrated AI infrastructure.

Nvidia plans to invest in d-Matrix, an AI inference-chip startup that already plans to connect its Raptor chips to Nvidia racks. Terms weren't disclosed.

Nvidia plans to invest in AI chip company d-Matrix, integrating its inference XPUs into Nvidia’s NVLink Fusion and MGX rack ecosystem.

d-Matrixは、AIのワークロードでは最終的に推論が支配的になるという信念のもとに設立された。同社は学習用ハードウェアの転用ではなく、一からアーキテクチャを構築した。シェスによれば、推論向けのチップと学習向けのチップには、アーキテクチャ上の中核的な違いがある。学習は計算（コンピュート）の問題である。だが推論は、単なる計算の問題ではないという。推論は計算とメモリー（記憶）の問題なのだ。 そして、そのメモリー部分が遅延（レイテンシー）を増やす。 LLMでは、最初の「処理」段階はしばしばプリフィル（prefill＝事前充填）と呼ばれる。モデルがプロンプト（入力文）を取り込み、文脈を構築し、関

Rayon uses empty sums as a proxy for zero values while it's splitting parallel work, but that doesn't really make sense for `DMatrix` (dynamically sized?) if they can't know what size zero-matrix will be addable to the rest of the actual matrices. Note that it's the `nalgebra` code raising the panic. Maybe `.reduce_with(add)` would work better for you?

I have following function: pub fn calculate_std_dev(matrices: &[DMatrix<f64>]) -> Option<DMatrix<f64>> { if matrices.is_empty() || matrices.iter().any(|m| m.shape() != matrices[0].shape()) { return None; } let len = matrices.len() as f64; let mean = matrices.iter().sum::<DMatrix<f64>>() / len; Some( matrices // .iter() never panics .par_iter() // always panic with Cannot compute `sum` of empty iterator. .map(|x| x.zip_map(&mean, |a, b| (a - b).powf(2.0))) .sum::<DMatrix<f64>>() .map(|x| (x / len).sqrt()), ) } when using iter it works pretty fast, but still 50% slower than C++ eigen, I think maybe using parallel computing can be faster, so I choose rayon and used `par_iter()` , but it panics at the beginning sum, can somebody explain it to me? thanks

Hi there! I am working with **nalgebra**. When accessing matrix elements one needs to extract a submatrix. To this end, **nalgebra** provides functions like `view` and `view_mut`, for example let A = DMatrix::from_fn(6, 8, |i, j| (i * 10 + j) as f64); let subA = A.view((2, 2), (4, 4)); But, it would be more convenient to use slice-like access let subA = get_block(&A, 1..3, 1..3); get_block_mut(&A, 1..4, 0..3).copy_from(B); or even better let subA = A[1..3, 2..5]; A[1..4, 0..3] = B; I implemented the function `get_block` wrapping the call `view` as fn get_block<T, R, C, S1>(mat : &Matrix<T, R, C, S1>, yrange : std::ops::Range<usize>, xrange : std::ops::Range<usize>) -> Matrix<T, Dyn, Dyn, ViewStorage<'_, T, Dyn, Dyn, <S1 as RawStorage<T, R, C>>::RStride, <S1 as RawStorage<T, R, C>>::CStride>> where C : Dim, R : Dim, S1 : RawStorage<T, R, C>, { let start = (yrange.start, xrange.start); let shape = (yrange.end - yrange.start, xrange.end - xrange.start); mat.view(start, shape) } The function looks ugly, because I need to specify explicitly the huge resulting type. I would like to know if there is a better solution?

Building the current crop of artificial intelligence chatbots has relied on specialized computer chips pioneered by Nvidia, which dominates the market and made itself the poster child of the AI boom.

Share your videos with friends, family, and the world