<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Blog on 0x0800</title><link>https://0x0800.com/blog/</link><description>Recent content in Blog on 0x0800</description><generator>Hugo</generator><language>en</language><lastBuildDate>Sat, 21 Nov 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://0x0800.com/blog/index.xml" rel="self" type="application/rss+xml"/><item><title>Intel D270 vs RTX PRO 6000: the 5.9x is real, on very narrow ground</title><link>https://0x0800.com/blog/intel-d270-5-9x/</link><pubDate>Sat, 21 Nov 2026 00:00:00 +0000</pubDate><guid>https://0x0800.com/blog/intel-d270-5-9x/</guid><description>&lt;p&gt;Intel says its D270 (350 W) delivers up to 5.9 times the throughput of Nvidia&amp;rsquo;s RTX PRO 6000 Blackwell Server Edition (600 W). I read the report and its notes: the figure is real, but it covers very narrow ground.&lt;/p&gt;&#10;&lt;p&gt;The 5.9x comes from double-precision (FP64) compute applied to option pricing: European Monte Carlo (5.9x), European Black-Scholes (5.6x), binomial tree (4.6x). On the American option the gap falls to 1.92x, and the four tests average 4.5x. Nvidia limits FP64 on its workstation cards to a 1:64 ratio, while the D270, built on the Xe3P architecture, runs it at full rate, with 64 FP64 FMAs per core. Intel is attacking where Nvidia left a gap.&lt;/p&gt;</description></item><item><title>A private AI on a €1,900 card: what the Intel Arc Pro B70 changes for regulated professions</title><link>https://0x0800.com/blog/ia-locale-b70-metiers/</link><pubDate>Fri, 23 Oct 2026 00:00:00 +0000</pubDate><guid>https://0x0800.com/blog/ia-locale-b70-metiers/</guid><description>&lt;p&gt;On a €1,900 (ex-VAT) graphics card, a 27-billion-parameter model answers at more than 100 tokens per second, without any document leaving the company. What makes it possible: our port of the TensorFold engine to the Intel Arc Pro B70 card, now merged into the official project.&lt;/p&gt;&#10;&lt;div class="kit-kf"&gt;&lt;div class="kit-kf-cell"&gt;&lt;div class="kit-label"&gt;Card price&lt;/div&gt;&lt;div class="kit-kf-value"&gt;€1,900 ex-VAT&lt;/div&gt;&lt;div class="kit-kf-detail"&gt;Intel Arc Pro B70, 32 GB, against €5,700 ex-VAT for an RTX 5090 with the same amount of memory&lt;/div&gt;&lt;/div&gt;&lt;div class="kit-kf-cell"&gt;&lt;div class="kit-label"&gt;One person&lt;/div&gt;&lt;div class="kit-kf-value"&gt;90 to 133 tok/s&lt;/div&gt;&lt;div class="kit-kf-detail"&gt;Qwen3.8-27B, depending on the text produced (conversation or code)&lt;/div&gt;&lt;/div&gt;&lt;div class="kit-kf-cell"&gt;&lt;div class="kit-label"&gt;Replayable answers&lt;/div&gt;&lt;div class="kit-kf-value"&gt;100 %&lt;/div&gt;&lt;div class="kit-kf-detail"&gt;an answer served alongside others equals the same answer served alone, on every try&lt;/div&gt;&lt;/div&gt;&lt;/div&gt;&#10;&#10;&lt;h2 id="the-real-obstacle-is-hosting"&gt;The real obstacle is hosting&lt;/h2&gt;&#10;&lt;p&gt;In an accounting firm, a notary&amp;rsquo;s office, a mutual insurer or an HR department, AI first runs into a simple question: where do the documents go? Client files, payslips or medical reports do not go to an external API, because professional secrecy and the GDPR forbid it.&lt;/p&gt;</description></item><item><title>TensorFold on Intel Arc Pro B70: from a Python port to a Zig engine</title><link>https://0x0800.com/blog/tensorfold-xe2-b70/</link><pubDate>Fri, 23 Oct 2026 00:00:00 +0000</pubDate><guid>https://0x0800.com/blog/tensorfold-xe2-b70/</guid><description>&lt;p&gt;TensorFold only knew Macs and NVIDIA GPUs. We ported it to the Intel Arc Pro B70 card (Xe2 architecture, 32 GB), first in Python through PyTorch, then in Zig directly on Level Zero, with no PyTorch and no Python at run time. The second port has been part of the official project since TensorFold 1.0.12. This text gives the path, the pitfalls and the measurements, including the ones that are less favourable to us.&lt;/p&gt;</description></item><item><title>DeepSeek-V4-Flash 0731 on 2× DGX Spark: Setup, Measurements, Tuning</title><link>https://0x0800.com/blog/dsv4-2x-spark-rapport/</link><pubDate>Tue, 04 Aug 2026 00:00:00 +0000</pubDate><guid>https://0x0800.com/blog/dsv4-2x-spark-rapport/</guid><description>&lt;div class="kit-kf"&gt;&lt;div class="kit-kf-cell"&gt;&lt;div class="kit-label"&gt;Retained quality&lt;/div&gt;&lt;div class="kit-kf-value"&gt;96/100&lt;/div&gt;&lt;div class="kit-kf-detail"&gt;judged by execution, before any throughput&lt;/div&gt;&lt;/div&gt;&lt;div class="kit-kf-cell"&gt;&lt;div class="kit-label"&gt;Single stream&lt;/div&gt;&lt;div class="kit-kf-value"&gt;78.55 tok/s&lt;/div&gt;&lt;div class="kit-kf-detail"&gt;code 512, at the ceiling imposed by k&lt;/div&gt;&lt;/div&gt;&lt;div class="kit-kf-cell"&gt;&lt;div class="kit-label"&gt;Sixteen streams&lt;/div&gt;&lt;div class="kit-kf-value"&gt;188.9 tok/s&lt;/div&gt;&lt;div class="kit-kf-detail"&gt;the only setting that produced a gain&lt;/div&gt;&lt;/div&gt;&lt;div class="kit-kf-cell"&gt;&lt;div class="kit-label"&gt;KV pool&lt;/div&gt;&lt;div class="kit-kf-value"&gt;1.79 M&lt;/div&gt;&lt;div class="kit-kf-detail"&gt;tokens, servable context 393,216&lt;/div&gt;&lt;/div&gt;&lt;div class="kit-kf-cell"&gt;&lt;div class="kit-label"&gt;Parameters ruled out&lt;/div&gt;&lt;div class="kit-kf-value"&gt;12 of 16&lt;/div&gt;&lt;div class="kit-kf-detail"&gt;the most minimal configuration is the optimum&lt;/div&gt;&lt;/div&gt;&lt;div class="kit-kf-cell"&gt;&lt;div class="kit-label"&gt;b12x cliff&lt;/div&gt;&lt;div class="kit-kf-value"&gt;&amp;#43;31 s&lt;/div&gt;&lt;div class="kit-kf-detail"&gt;of fixed cost from ~1850 tokens of context&lt;/div&gt;&lt;/div&gt;&lt;/div&gt;&#10;&#10;&lt;p&gt;Reproducible report. It describes how to set up the stack end to end, which numbers you get, what each parameter changes, and which combination gives the best performance/quality trade-off.&lt;/p&gt;</description></item><item><title>Serving DeepSeek-V4-Flash-0731 on Two DGX Spark (TP=2): The Full Recipe</title><link>https://0x0800.com/blog/dsv4-2x-spark-recette/</link><pubDate>Tue, 04 Aug 2026 00:00:00 +0000</pubDate><guid>https://0x0800.com/blog/dsv4-2x-spark-recette/</guid><description>&lt;div class="kit-kf"&gt;&lt;div class="kit-kf-cell"&gt;&lt;div class="kit-label"&gt;Correct code&lt;/div&gt;&lt;div class="kit-kf-value"&gt;96/100&lt;/div&gt;&lt;div class="kit-kf-detail"&gt;judged by execution, shipped config&lt;/div&gt;&lt;/div&gt;&lt;div class="kit-kf-cell"&gt;&lt;div class="kit-label"&gt;Single-stream throughput&lt;/div&gt;&lt;div class="kit-kf-value"&gt;78.55 tok/s&lt;/div&gt;&lt;div class="kit-kf-detail"&gt;code, 512 tokens, excluding TTFT&lt;/div&gt;&lt;/div&gt;&lt;div class="kit-kf-cell"&gt;&lt;div class="kit-label"&gt;Aggregate throughput&lt;/div&gt;&lt;div class="kit-kf-value"&gt;188.9 tok/s&lt;/div&gt;&lt;div class="kit-kf-detail"&gt;at sixteen streams, TTFT under 0.8 s&lt;/div&gt;&lt;/div&gt;&lt;div class="kit-kf-cell"&gt;&lt;div class="kit-label"&gt;KV cache&lt;/div&gt;&lt;div class="kit-kf-value"&gt;1.79 M&lt;/div&gt;&lt;div class="kit-kf-detail"&gt;tokens, read from the startup log&lt;/div&gt;&lt;/div&gt;&lt;div class="kit-kf-cell"&gt;&lt;div class="kit-label"&gt;Power&lt;/div&gt;&lt;div class="kit-kf-value"&gt;41 W&lt;/div&gt;&lt;div class="kit-kf-detail"&gt;GPU power per box at peak, no throttling&lt;/div&gt;&lt;/div&gt;&lt;div class="kit-kf-cell"&gt;&lt;div class="kit-label"&gt;RDMA latency&lt;/div&gt;&lt;div class="kit-kf-value"&gt;1.86 µs&lt;/div&gt;&lt;div class="kit-kf-detail"&gt;typical under load, 8-byte messages&lt;/div&gt;&lt;/div&gt;&lt;/div&gt;&#10;&#10;&lt;p&gt;This document describes what we did to serve DeepSeek-V4-Flash-0731 in TP=2 on two DGX Spark (GB10, SM121), and what we measured. It includes the pitfalls that cost us time, because those are the ones missing everywhere else.&lt;/p&gt;</description></item><item><title>Serving Laguna S 2.1 NVFP4 on a single NVIDIA DGX Spark (GB10): the complete runbook</title><link>https://0x0800.com/blog/laguna-s21-single-spark/</link><pubDate>Thu, 23 Jul 2026 00:00:00 +0000</pubDate><guid>https://0x0800.com/blog/laguna-s21-single-spark/</guid><description>&lt;p&gt;From the box killed by its own JIT compiler to a working 256K service. A 117.6B-parameter MoE (8.5B active), DFlash speculative decoding, native 256K context, all on a single unified-memory machine — no TP=2, no cross-node, no NCCL. Every command in this runbook is given verbatim.&lt;/p&gt;&#10;&lt;p&gt;Laguna fits on a single machine, where DeepSeek-V4 required two nodes and a coordination layer. The hard part is therefore no longer distribution but the &lt;strong&gt;box&amp;rsquo;s memory&lt;/strong&gt;, which an unbridled JIT exhausts in a few seconds. This runbook covers the hardware context, the three traps that kill the boot, the complete single-node launch and the benchmark results measured on our hardware.&lt;/p&gt;</description></item><item><title>Running vLLM on an NVIDIA DGX Spark (GB10): from git clone to NVFP4 inference</title><link>https://0x0800.com/blog/dgx-spark-vllm-nvfp4/</link><pubDate>Mon, 20 Jul 2026 00:00:00 +0000</pubDate><guid>https://0x0800.com/blog/dgx-spark-vllm-nvfp4/</guid><description>&lt;p&gt;The DGX Spark ships a GB10 GPU (Blackwell, SM121 compute architecture) on an aarch64 board. An architecture recent enough that several libraries of the CUDA ecosystem do not yet have a precompiled kernel for SM121. Three fixes that do not jump out of the official documentation, explained here from cloning the repository to launching an NVFP4-quantised model.&lt;/p&gt;&#10;&lt;p&gt;This article covers the full installation, from cloning the repository to the successful launch of an NVFP4-quantised model (Mistral-Small 119B), explaining why each step is necessary. All the commands are tested on our hardware; adapt the paths (&lt;code&gt;~/models/...&lt;/code&gt;, model names) to yours.&lt;/p&gt;</description></item></channel></rss>