Qwen 3.8 27B at 67 tok/s on Two $78 Nvidia Tesla P100s
This video shows how two inexpensive 16 GB Tesla P100 cards can run a Qwen 3.8 27B Quant 6 model at roughly 67 tokens per second by combining a community llama.cpp build, careful hardware setup, and speculative decoding. It explains the VRAM and quantization tradeoff, fixes for firmware, power, and cooling problems, and the limits of this budget local-LLM configuration.
Junk GPU Lab12 minTranscript found
Quick learning frame
Read this before watching.
A local runtime lesson is about fit: model, quantization, hardware, endpoint, latency, privacy, tool integration, and task limits.
New playlist item from Junk GPU Lab; queued for transcript-backed review, topic mapping, and a practical learning artifact.
Skill you build: The ability to design and tune a low-cost local LLM system by balancing model quantization, aggregate VRAM, hardware constraints, cooling, power, and inference acceleration.
Watch for the shift from claim to mechanism. The learning value is the point where the transcript reveals a repeatable action, tool boundary, context move, review habit, or artifact.
Concept diagram
Where this video fits.
01Task
02Hardware
03Model/quantization
04Runtime endpoint
05Agent tool loop
06Benchmark task
07Fallback
Deep lesson
Turn this video into working knowledge.
1,917 cleaned transcript words reviewed across 558 timed caption segments.
Thesis
Qwen 3.8 27B at 67 tok/s on Two $78 Nvidia Tesla P100s teaches a practical local model/runtime move: This video shows how two inexpensive 16 GB Tesla P100 cards can run a Qwen 3.8 27B Quant 6 model at roughly 67 tokens per second by combining a community llama.cpp build, careful hardware setup, and speculative decoding. It explains the VRAM and quantization tradeoff, fixes for firmware, power, and cooling problems, and the limits of this budget local-LLM configuration.
The goal is not to remember the video. The goal is to extract the operating principle, tie it to timestamped evidence, test how far the claim transfers, and make something reusable.
0:56
Size VRAM First
“was like, holy cow. I had to know if that was real on my own machine. >> >> One problem, though. A model has to fit in your graphics card's memory, the VRAM. And if it doesn't fit,...”
A model that exceeds GPU memory spills into normal RAM and becomes much slower, while heavier quantization reduces memory use at a potential intelligence cost. The creator targeted Quant 6 for the 27B model, requiring about 32 GB of VRAM, and combined two 16 GB P100s to reach that capacity cheaply. For one local model, record its parameter count, desired quantization, estimated VRAM requirement, and the lowest acceptable generation speed before choosing hardware.
6:38
Speculate for Speed
“speculative decoding. The model has a small extra head that guesses the next four words, and then the full model checks all four guesses in one go instead of writing one word at a time. It's like the...”
Speculative decoding uses a small extra head to guess several upcoming tokens so the full model can verify them together. On this setup, enabling it raised generation speed from 29.7 to about 49 tokens per second initially, with predictable reasoning later reaching roughly 67 because more guesses were accepted. Benchmark the same prompt with speculative decoding disabled and enabled, then record speed and acceptance rate for both predictable reasoning and ordinary prose.
8:16
Tune the Whole Rig
“Oh, and a smaller mixture of experts model, Gemma 4, hit 79 on just one of these cards. Now, I'm not saying these beat a modern GPU. Without speculative decoding, a 3090 is still clearly faster. But turn...”
The working recipe combined graphics-mode firmware, restored ECC, 150 W power caps, a properly ducted fan shroud, a P100-specific llama.cpp fork, and four speculative guesses. The setup is presented as suitable for local LLM inference, not gaming or video generation, and it trades plug-and-play convenience for a much lower hardware price. Create a preflight checklist covering firmware backup, address-space mode, ECC, power delivery, temperature under stress, driver support, and inference-build compatibility.
01
Task
Start with this video's job: This video shows how two inexpensive 16 GB Tesla P100 cards can run a Qwen 3.8 27B Quant 6 model at roughly 67 tokens per second by combining a community llama.cpp build, careful hardware setup, and speculative decoding. It explains the VRAM and quantization tradeoff, fixes for firmware, power, and cooling problems, and the limits of this budget local-LLM configuration. Treat "Task" as the outcome you are trying to make visible, not a topic label. Anchor it to 0:56, where the video says: “was like, holy cow. I had to know if that was real on my own machine. >> >> One problem, though. A model has to fit in your graphics card's memory, the VRAM. And if it doesn't fit,...”
02
Hardware
Use "Hardware" to locate the part of the local model/runtime mechanism the video is demonstrating. Ask what changes in your real setup if this claim is true. Anchor it to 6:38, where the video says: “speculative decoding. The model has a small extra head that guesses the next four words, and then the full model checks all four guesses in one go instead of writing one word at a time. It's like the...”
03
Model/quantization
Turn "Model/quantization" into the reusable artifact for this lesson: A local model/runtime fit sheet with hardware constraints, model choice, endpoint setup, task benchmark, and fallback rule. This is where watching becomes something you can inspect and reuse.
04
Runtime endpoint
Use "Runtime endpoint" as the application surface. Decide whether the idea touches a browser flow, a local file, a model choice, a source document, a UI, or a review step.
05
Agent tool loop
Use "Agent tool loop" to prove the lesson. The evidence should connect back to the video title, transcript anchors, and a concrete output, not a generic best-practice claim.
06
Benchmark task
Use "Benchmark task" to carry the idea forward: save the prompt, checklist, diagram, or operating rule that would make the next agent run better.
07
Fallback
Connect "Fallback" to Qwen 3.8 27B at 67 tok/s on Two $78 Nvidia Tesla P100s by naming the claim, the evidence, and the artifact it should produce.
Example
Source-backed artifact packet
Convert the video into a scoped artifact request that includes the transcript claim, mechanism, acceptance criteria, and proof. The output should be a local model/runtime fit sheet with hardware constraints, model choice, endpoint setup, task benchmark, and fallback rule..
Example
Local model/runtime proof brief
Separate what the speaker claims, what the demo actually proves, and what still needs outside verification before you adopt the local model/runtime pattern.
Example
Teach-back module
Transform the lesson into a definition, a Task -> Hardware -> Model/quantization -> Runtime endpoint -> Agent tool loop -> Benchmark task -> Fallback diagram, one misconception, one practice exercise, and a check-for-understanding question.
Do not learn it wrong
Treating the title as the lesson without checking what the transcript actually says.
using a local model like ChatGPT
ignoring latency/context limits
no benchmark task
Letting the lesson drift into local-model ideology.
Letting the lesson drift into hardware specs without workflow fit.
Letting the lesson drift into benchmarks unrelated to the actual task.
Do not count this as learned until these are true.
01
State the transcript-backed claim in your own words: This video shows how two inexpensive 16 GB Tesla P100 cards can run a Qwen 3.8 27B Quant 6 model at roughly 67 tokens per second by combining a community llama.cpp build, careful hardware setup, and speculative decoding. It explains the VRAM and quantization tradeoff, fixes for firmware, power, and cooling problems, and the limits of this budget local-LLM configuration.
02
Explain the practical stakes without hype: New playlist item from Junk GPU Lab; queued for transcript-backed review, topic mapping, and a practical learning artifact.
03
Map the idea onto the Task -> Hardware -> Model/quantization -> Runtime endpoint -> Agent tool loop -> Benchmark task -> Fallback sequence and name the weakest link.
04
Produce the artifact and include the evidence that proves it: A local model/runtime fit sheet with hardware constraints, model choice, endpoint setup, task benchmark, and fallback rule.
Put it into practice
Give this grounded prompt to Codex or Claude after watching.
You are helping me turn one specific YouTube video into real, durable learning.
Source video:
- Title: Qwen 3.8 27B at 67 tok/s on Two $78 Nvidia Tesla P100s
- URL: https://www.youtube.com/watch?v=JvLUoJYUjDM
- Topic: Creative Automation
- My current learning frame: Plan a budget local-LLM build by calculating its VRAM target, documenting firmware, power, and cooling safeguards, and defining a repeatable benchmark with speculative decoding on and off.
- Why this matters: New playlist item from Junk GPU Lab; queued for transcript-backed review, topic mapping, and a practical learning artifact.
Transcript anchors from this exact video:
- 0:56 / Evidence 1: "was like, holy cow. I had to know if that was real on my own machine. >> >> One problem, though. A model has to fit in your graphics card's memory, the VRAM. And if it doesn't fit,..."
- 2:40 / Evidence 2: "that's basically a card Fred Flintstone used. Look at the specs. 16 GB of high-bandwidth memory each, the fast kind, stacked right next to the chip. $78 a pop. Two of them make 32 GB, exactly what I..."
- 4:58 / Evidence 3: "setting is called GPU mode graphics. Two tips. Back up the original firmware first, always. And the flash quietly switches off the memory error checking, ECC. So, turn it back on afterwards. After that, the demand shrinks to..."
- 6:38 / Evidence 4: "speculative decoding. The model has a small extra head that guesses the next four words, and then the full model checks all four guesses in one go instead of writing one word at a time. It's like the..."
- 8:16 / Evidence 5: "Oh, and a smaller mixture of experts model, Gemma 4, hit 79 on just one of these cards. Now, I'm not saying these beat a modern GPU. Without speculative decoding, a 3090 is still clearly faster. But turn..."
- 9:59 / Evidence 6: "adapter cables for these cards, and a ducted fan shroud. The cards. Flash both to graphics mode with NVFlash and back up the firmware first. The system. Ubuntu 2404 and the Nvidia 580 driver, the last one that..."
Video-aware target:
- Prompt lane: Local model/runtime
- Mechanism to extract: Identify why the local setup works or fails for this specific agent task, not whether local models are generally good.
- Artifact to produce: A local model/runtime fit sheet with hardware constraints, model choice, endpoint setup, task benchmark, and fallback rule.
- Artifact must include: hardware; runtime; model/quantization; endpoint; agent integration; benchmark/fallback
Your task:
1. Use the transcript anchors above as the primary source packet. If you add outside context, label it clearly as outside context and keep it secondary.
2. Create a source-check table with columns: timestamp, claim, transcript support, what the demo proves, confidence, and what still needs verification.
3. Extract the actual teachable mechanism from the video: Identify why the local setup works or fails for this specific agent task, not whether local models are generally good. Do not invent claims that are not supported by the title, lesson frame, or transcript anchors.
4. Build a reusable learning artifact: A local model/runtime fit sheet with hardware constraints, model choice, endpoint setup, task benchmark, and fallback rule.
5. Include:
- a plain-English definition of the core idea
- a diagram or structured model using this sequence: Task -> Hardware -> Model/quantization -> Runtime endpoint -> Agent tool loop -> Benchmark task -> Fallback
- answers to these source questions: What machine/runtime is shown? | What task exposes the model limit? | What setup change improves the loop?
- 3 concrete examples that apply the video idea to real agentic work, such as Ollama or LM Studio coding endpoint; MLX Apple Silicon runner; DGX-backed Hermes session
- 2 failure modes the video helps prevent, chosen from the transcript evidence and these likely risks: using a local model like ChatGPT; ignoring latency/context limits; no benchmark task
- a checklist for the next real workflow, focused on: task fit, runtime setup, latency/context, tool loop, fallback
- one practical exercise with a clear done signal: Choose one real coding task and specify the pass/fail benchmark for a local model.
6. Add a "learning transfer" section: what changes in my workflow tomorrow if I actually learned this?
7. Add a "source check" section that cites which transcript anchor supports each major takeaway.
Quality bar:
- Make this specific to "Qwen 3.8 27B at 67 tok/s on Two $78 Nvidia Tesla P100s", not a generic Creative Automation essay.
- Tie each harness element to a transcript anchor that names a tool, state boundary, permission, model behavior, or verification step.
- Prefer operational examples, failure modes, and reusable artifacts over broad definitions.
- Call out uncertainty instead of smoothing over weak evidence.
- Avoid these generic drifts: local-model ideology; hardware specs without workflow fit; benchmarks unrelated to the actual task.
- If evidence is weak or missing, stop and say what transcript segment or timestamp needs review instead of guessing.
- Finish with a concise artifact I could paste into my learning app.
Misconceptions
What to stop believing.
Creative AI removes the need for taste.
It increases the need for taste because output volume explodes.
The best prompt is enough.
References, critique, iteration, and post-production matter just as much.
Practice studio
Learning only counts when you make something.
01
Transcript evidence map
Separate what the video actually says from what you already believe about the topic.
3 source-backed takeaways with timestamps, confidence, and a transfer note.02
One useful artifact
Apply the video to a real workflow and produce a local model/runtime fit sheet with hardware constraints, model choice, endpoint setup, task benchmark, and fallback rule..
A reusable artifact with a done signal and one verification step.03
Local model/runtime teach-back card
Explain the local model/runtime mechanism to someone who has not watched the video yet.
A 90-second explanation, one diagram, one example, and one misconception to avoid.
Recall check
Answer first, then reveal — without rewatching.
Why did the creator choose two Tesla P100s for the Quant 6 version of the 27B model?
How does speculative decoding accelerate generation on this setup?
Which four changes does the creator summarize as essential to the successful P100 setup?
Source shelf
Use the video as a doorway, then verify with primary sources.
