File size: 953 Bytes
fd4d6aa
984576f
 
 
fd4d6aa
 
984576f
 
 
fd4d6aa
 
984576f
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
---
title: VKIE  VIDRAFT Kernel Inference Engine
emoji: 🚄
colorFrom: purple
colorTo: indigo
sdk: static
pinned: true
license: apache-2.0
short_description: VKIE (비키)  VKAE accelerates, VKUE saves, VKIE maximizes
---

# 🚄 VKIE (비키) — VIDRAFT Kernel Inference Engine

The unified control tower for VIDRAFT's inference line, across three axes:

- 🏎️ **VKAE** — GPU **acceleration** (the sports car): ~9× faster single-stream on a B200.
- 🚗 **VKUE** — GPU **savings** (the compact car): 34.7B on a FREE CPU, no GPU.
- 🚄 **VKIE** — accel + savings = **max serving** (the train): 18,057 tok/s on one B200.

Every number is measured; every demo is live (they run on their real hardware, embedded or linked here). The "after" figures come from VIDRAFT's optimized serving — results are public, engine internals are proprietary.

Links: VKAE · VKUE · GPU-vs-CPU · CPU-only · FREE CPU · Image (T4) · Model · Blog.