mtone · HackerTrans

vor 26 Tagendiscuss

It was trained from scratch by Cohere. They're the only Canadian AI lab - I'm glad they're releasing open weights and I wish them luck catching up!

mtone

vor 26 Tagendiscuss

I am using the `voipmonitor/vllm:lucifer` docker from the RTX6K discord community discussed at the same link the other commenter posted. It is based around this PR https://github.com/vllm-project/vllm/pull/43477

mtone

vor 26 Tagendiscuss

Here's a DeepSeek-V4-Flash benchmark on 2X RTX Pro 6000:

  - Prefill: ~10K tok/s
  - Decode: 190 | 375 | 980 tok/s (for 1 | 4 | 16 concurrent requests)
  - GPU power draw during benchmark: Average: 585W | Max: 849W | Limit: 1200W with undervolt. Idle PC is 125W.

I've asked it to calculate the following considering a realistic blend of cached prompts and decode for agentic dev scenario.

Electricity-only (@ USD $0.08/kWh)

  Usage          | IN price  | OUT price | Monthly cost
  Concurrency=1  | $0.040/M  | $0.080/M  | $8.65 to $38.88 (5% to 100% active)
  Concurrency=4  | $0.024/M  | $0.044/M  | up to $48.67 (cheaper per token but higher power draw)

Total cost of ownership over 3 years is electricity + USD $20K (pre-hike pricing). In a production scenario, how much would I have to charge my users to break even, aiming for 4 concurrent requests 24/7?

A) Breakeven API pricing (est. 2B IN + 1B OUT throughput/month):

                        IN price    OUT price
  Self-hosted           $0.121/M    $0.363/M
  OpenRouter (budget)   $0.098/M    $0.196/M
  OpenRouter (DeepSeek) $0.140/M    $0.280/M

B) Breakeven subscription (users active ~1.5h/day):

    1 user: $563/mo (oh, hai)
    25 users: $23/mo
    100 users: $6/mo

mtone

vor 2 Monatendiscuss

You're forgetting a critical factor: concurrency. If a given hardware serves a single request at 150 tokens/s, it can also serve 20-30 requests at 100 tokens/s. Suddenly your $5K becomes $100K/month, enough to recoup the cost of the hardware in a year or so.

The reason it works: each time you read the model (memory bound) to calculate the next token, you can also update multiple requests (compute bound) while at it. It's also much more energy-efficient per token.

[1] https://aimultiple.com/gpu-benchmark

mtone

vor 3 Monatendiscuss

Thanks!

There's a heading control to include rotation in link: https://earth.google.com/web/@3.63731074,-23.1618975,-2690.7...

mtone

vor 5 Monatendiscuss

> Does privacy of Netflix ratings matter? The issue is not “Does the average Netflix subscriber care about the privacy of his movie viewing history?,” but “Are there any Netflix subscribers whose privacy can be compromised by analyzing the Netflix Prize dataset?”

Well said.

mtone

vor 5 Monatendiscuss

For this type of repetitive application I think it's common to "fine-tune" a model trained on your business problem to reach higher quality/reliability metrics. That might not be possible with this chip.

mtone

vor 6 Monatendiscuss

Good question. That was for my MS account/licenses and some Azure stuff. I use Google Authenticator for most things.

Thanks for the link, I'll take a look. I might just move it to a secondary device first.

mtone

vor 6 Monatendiscuss

Just looked - Microsoft Authenticator doesn't appear to work. I might be able to get off of it but it will take some prep. My banks are supported so that's good.

mtone

vor 7 Monatendiscuss

> if you could exclude all of the R&D and training costs

LLMs have a short shelf-life. They don't know anything past the day they're trained. It's possible to feed or fine-tune them a bit of updated data but its world knowledge and views are firmly stuck in the past. It's not just news - they'll also trip up on new syntax introduced in the latest version of a programming language.

They could save on R&D but I expect training costs will be recurring regardless of advancements in capability.

mtone

vor 9 Monatendiscuss

Recently llama.cpp made a few common parameters default (-ngl 999, -fa on) so it got simpler: --model and --context-size and --jinja generally does it to start.

We end up fiddling with other parameters because it provides better performance for a particular setup so it's well worth it. One example is the recent --n-cpu-moe switch to offload experts to CPU while filling all available VRAM that can give a 50% boost on models like gpt-oss-120b.

After tasting this, not using it is a no-go. Meanwhile on Ollama there's an open issue asking for this: https://github.com/ollama/ollama/issues/11772

Finally, llama-swap separately provides the auto-loading/unloading feature for multiple models.

mtone

vor 10 Monatendiscuss

Do you really need a H200 for this? Seems like something a consumer GPU could do. Smaller models might be ideal [0] as they don't require extensive world knowledge and are much more cost efficient/faster.

Why can't you build this today?

[0]: https://arxiv.org/pdf/2506.02153 Small Language Models are the Future of Agentic AI (Nvidia)

mtone

vor 4 Jahrendiscuss

I close a door or put a box in its way. No app.

comments