Builds / RTX A6000 48 GB
RTX A6000 48 GB: Ripple-65B Preview (dense 65B coder) Q4_K_M on llama.cpp at 14.3-16.3 tok/s, prompts up to 256K on one card
Last updated:
- GPU(s)
- RTX A6000 48 GB
Amazon · eBay (used) · Newegg - VRAM / unified memory (GB)
- 48
- CPU
- not stated
- Motherboard
- not stated
- RAM
- not stated
- PSU
- not stated
- Other
- InferenceKey Cloud host, RTX A6000 (300 W, CC 8.6), all 80 layers on the GPU; llama.cpp b12167 prebuilt ubuntu-cuda-13.4-x64, -ngl 99 -fa on --jinja
- Reported total cost
- not stated
- Power
- not stated
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| Ripple 65B Preview | Q4_K_M (AtomicChat GGUF, 38.8 GB) | llama.cpp b12167 llama-server (CUDA) | 14.4 | 540 @ 25,930-token prompt est. TTFT @8k ≈ 15.2 s (calculated from measured prefill) | context: 32K ctx, f16 KV, one request at a time📏 Streamed decode on a 174-token coding answer (merge_intervals); 14.3-16.3 tok/s across the tests, 14.5-15.5 per agent turn; 13.5-14.3 with 128K q8_0 or 256K q4_0 KV; peak VRAM 40,593 MiB at 32K |
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.
Owner-run deploy test by InferenceKey (published 8 Nov 2026). Ripple-65B Preview is Ricursive's dense 65B coder (Apache 2.0, weights on Hugging Face 4 Nov 2026), runnable in llama.cpp since b12167. A needle at token 237,532 was found with 256K ctx and q4_0 KV (43,440 MiB peak, 760 s). AtomicChat's card lists Q4_K_M at 93.0% same-top-token vs BF16. Same article: a real RTX 3090 with the 2-bit file is a separate RigBase build.
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.