Gemma 4 Deployed on AMD MI300X via AMD Developer Cloud for $1.99/Hour

·24 views

A developer has published a step-by-step guide for deploying Google's Gemma 4 E2B language model on an AMD Instinct MI300X GPU accessed through AMD Developer Cloud, which runs on DigitalOcean infrastructure, at $1.99 per hour on-demand. The setup uses a vLLM OpenAI-compatible server running on ROCm, managed entirely via SSH and the DigitalOcean v2 API from a workstation with no local AMD GPU. A custom Python MCP server exposes 21 tools to handle droplet management, deployment, logging, and benchmarking without requiring any local ROCm installation. Performance benchmarks recorded 305 tokens per second on a single stream and over 10,000 tokens per second across 64 concurrent streams, translating to roughly 2,713 tokens per dollar-hour. The full project code is publicly available on GitHub, and all GPU interactions are routed through SSH to avoid any dependency on local AMD hardware drivers.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Random UUIDv4 Primary Keys Can Silently Degrade Database Write Performance

A backend engineer reported a 340% spike in database write latency after their table passed 50 million rows, with CPU usage normal and storage I/O saturated. The cause was using random UUIDv4 values as primary keys, which were chosen to avoid enumeration attacks and simplify distributed ID generation. Relational databases store primary keys in sorted B-Tree indexes, so random keys land in arbitrary positions, forcing cold page reads, frequent page splits, and bloated, half-empty index pages. The problem is worse in MySQL's InnoDB, where the primary key also determines physical row order, fragmenting the entire table. The author benchmarks UUIDv4 against the time-ordered UUIDv7 standard, which places a timestamp first so new keys append near the end of the index.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why Logo File Formats Matter for Print and Web, and What to Request

A common problem arises when clients send low-resolution JPGs with white backgrounds that look poor on dark or high-density screens and in large-format print. The article argues that most logo issues stem from missing file types rather than design flaws. It recommends requesting a vector master file such as AI or SVG, print-ready PDF or EPS files, and transparent PNGs in several sizes. It also suggests color variants, multiple layouts, and color codes in HEX, RGB, CMYK, and Pantone. The core advice is that every other format can be exported from a vector source, while small raster images cannot be scaled up well.

0 comments Read more at DEV Community

ProgrammingHacker News ·

Blender 5.3 Released With Gaussian Splat Support

Blender 5.3 has been released, according to the official Blender website. The update adds support for Gaussian Splats, a technique for representing and rendering 3D scenes. The announcement drew attention on Hacker News, where it received 16 points and one comment. The source material provides only the release announcement and discussion metadata, so further details about the feature's capabilities are not available here.

0 comments Read more at Hacker News

ProgrammingHacker News ·

Google announces Gemini 4 Carbon is now available

Google has made Gemini 4 Carbon available, according to a post on its official blog. The item was shared on Hacker News, where it drew 55 points and 22 comments. The source material does not include further details about the model's capabilities, pricing, or intended uses.

0 comments Read more at Hacker News