tag: inference
articles: 4 · beats: 2
latest: October 20, 2026
---
Inference
4 Crashtech articles on Inference, spanning Ai Infrastructure and How AI Actually Works, published in October 2026. Every piece is full-text HTML with sources, structured data and an authored FAQ.
They span Ai Infrastructure (2), How AI Actually Works (2). Ai Infrastructure How AI Actually Works
Deepgram's Trainium Port Shows Why Inference Cost Depends on the Whole Stack
Deepgram and AWS report 11–70% more offline transcription per dollar on one Trainium2 setup. The kernels, runtime and workload explain the narrow result.
Why LLM Serving Teams Are Turning KV Cache Transfer Into a First-Class Design Choice
Two new vLLM transfer techniques target first-token latency by overlapping movement and reusing multi-turn context. See where each helps and what it costs.
NVIDIA Dory Makes Inference Compute a Knob. What Does That Buy?
NVIDIA Dory 1.0 reuses transformer layers at inference. Its adjustable loop count trades more computation for task performance, not free capability.
The KV Cache: How GQA and DeepSeek MLA Reduce Inference Memory
KV-cache size depends on layers, heads, precision and context. GQA shares key-value heads; MLA caches a learned latent plus positional data.
Questions we answer about Inference
- Did Trainium make every Deepgram workload 11–70% cheaper?
- What had to change to run Nova-3 on Trainium2?
- Is this directly comparable with live speech recognition?
- What is KV cache in LLM inference?
- Are the reported vLLM speedups guaranteed in production?
- When should a team consider these approaches?
- What is NVIDIA Dory 1.0?
- Does increasing Dory's loop count make it a larger model?
- Are the model card's benchmark numbers independent results?
- Is a KV cache always 80% of GPU memory?
Covered alongside
Frequently asked questions
What does Crashtech publish about Inference?
4 articles tagged Inference, the most recent published October 20, 2026. They span Ai Infrastructure (2), How AI Actually Works (2). Each carries numbered sources, an authored FAQ and full structured data.
What questions about Inference does Crashtech answer directly?
10 questions have a dedicated answer page under this tag, including “Did Trainium make every Deepgram workload 11–70% cheaper?”. Each answer is authored prose from the article it belongs to, not a generated summary.
Can AI assistants read Crashtech's Inference coverage?
Yes. Crashtech serves full static HTML to every crawler, allows all major AI user agents in robots.txt, and publishes an llms.txt manifest plus a full-text corpus, so assistants can retrieve and cite these articles directly.