Builds / NVIDIA L4 24 GB
NVIDIA L4 24 GB (72 W): Qwen3.8 27B stock llama.cpp 14.9 tok/s vs DSpine drafter 47.3 vs tqLLM 67.0 at 32K context
Last updated:
- GPU(s)
- NVIDIA L4 24 GB
Amazon · eBay (used) · Newegg - VRAM / unified memory (GB)
- 24
- CPU
- not stated
- Motherboard
- not stated
- RAM
- not stated
- PSU
- not stated
- Other
- 23,034 MiB usable with ECC on, 72 W limit (median draw 71.4-71.9 W in every row, peak 78.03 W); cloud notebook session
- Reported total cost
- not stated
- Power
- 72 W card limit
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| Qwen3.8 27B | UD-IQ4_XS (Unsloth, 14.25 GB), q8_0 KV | llama.cpp (stock) | 14.9 | 729 (prompt length not stated) | context: 32K context row, one request📏 Median of 3 greedy runs, answers capped at 256 tokens; 12.98 tok/s at 64K, 10.11 at 128K; TTFT 33.04 s at 32K, 82.36 s at 64K, 227.73 s at 128K |
| Qwen3.8 27B | UD-IQ4_XS (Unsloth, 14.25 GB) + Tencent DSpine-b16 drafter (1.26 GB) | llama.cpp fork (Ajay9o9 dspine branch) | 47.25 | 718 (prompt length not stated) | ⚡ spec-decode (DSpine (n-max 15))context: 32K context row, one request📏 Median of 3 greedy runs; 27.09 at 64K (n-max 15), 21.18 at 128K (n-max 7); prefill within 1.5% of stock |
| Qwen3.8 27B | TQ3_4S TurboQuant (14.5 GB, different weights) + DFlash2 drafter (0.79 GB) | tqLLM 1.4.4 (engine 9a9e7e5) | 66.97 | 691 (prompt length not stated) | ⚡ spec-decode (DFlash2)context: 32K context row, one request📏 Median of 3 greedy runs; 61.59 at 64K; 20.94 at 128K because 3.09 GB of KV plus the LM head and embeddings moved to host memory on this 23,034 MiB card |
Prompt processing = prefill speed (how fast the prompt is read) as reported by the source, with the prompt length it was measured at; blank = the source does not report it. est. TTFT @8k is calculated from the measured prefill speed (8,192 ÷ prefill tok/s), shown only when prefill was measured on a prompt of 512+ tokens. It is not a measurement and ignores load time and caching. measured TTFT is the source's own time-to-first-token. Why prefill matters for coding assistants.
First-party run by vramcalculator.com (published 10 Nov 2026, updated 14 Nov). Three stacks on one card: stock and DSpine share the same weights; tqLLM uses its own TurboQuant file, so that row compares stack against stack, not bytes against bytes. Three runs of one prompt per size, not a quality test. The DSpine and tqLLM 3090 figures in the same article are the authors' own claims and are not listed here.
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.
Speculative-decoding rows on this page are not like-for-like with plain decode. See how much MTP and DFlash really speed things up (measured pairs).