Near silent LLM Monster... NVIDIA, take notes
Credibility score: 74/100 — Mostly Credible. Mixed credibility - some claims are solid, others need verification.
BSmeter analyzed "Near silent LLM Monster... NVIDIA, take notes" and rated it 74/100 for credibility (a BS score of 26/100 — mostly credible), on 2026-04-28. Its weakest claim — "Framework might have best cooling; tested Asus Z13 and GMK Evo X2" — scored 50/100 and was flagged as opinion. 38 claims were checked against the video transcript. Scores are produced by BSmeter's AI analysis of the transcript, not independent human verification.
Of 38 claims analyzed: 0 scored under 40, 10 between 40 and 69, and 28 at 70 or above.
Claims analyzed
Framework Desktop has AMD Ryzen Max+ 395 chip with 128GB RAM — Solid (85/100)
At 0:00
👌✅
Why this score: Framework Desktop specs confirmed.
- Framework announced their Desktop with AMD Ryzen AI Max+ 395 (Strix Halo) supporting up to 128GB LPDDR5X RAM, ideal for local AI workloads.
- Matches video's description precisely; 'Ryzen Max Plus 395' is accurate shorthand.
Original quote: “[0:00] The Framework Desktop is the newest [0:02] machine to land with AMD's Ryzen Max [0:04] Plus 395's chip that has 128 gigs of [0:07] RAM, which is fantastic for local AI”
Framework might have best cooling; tested Asus Z13 and GMK Evo X2 — Opinion (50/100)
At 0:10
Best cooling? Bold claim for untested hype 🚩😏
Why this score: Subjective opinion on cooling and personal tests.
- Speaker admits Framework is 'late to the party' but speculates on 'best cooling solution yet' without evidence shown.
- Mentions testing Asus Flow Z13 (known ROG Flow with similar chips) and GMK Tech Evo X2 (mini-PC with Strix Halo), calling them impressive—valid anecdote but unverified comparison.
- No data on Framework's actual cooling performance vs. competitors.
Original quote: “[0:10] like large language models and image and [0:12] video generation. And while they're a [0:14] little late to the party, not the first, [0:16] not the second, not even the third, they [0:19] might have just nailed the best cooling [0:21] solution yet. I've tested the Asus Flow [0:23] Z13.…”
Framework board whisper quiet like Mac Studio — Opinion (50/100)
At 0:35
Subjective noise comparison — fair game 👌
Why this score: Personal assessment of fan noise levels. Framework boards are known for efficient cooling; Mac Studio is indeed very quiet due to its design. No objective dB measurements given, but claim fits user reports.
Original quote: “But the framework board, wow, whisper quiet. Think Mac Studio level quiet.”
Framework announced desktop this year — Solid (85/100)
At 0:58
✅
Why this score: Confirmed announcement. Framework Desktop Motherboard was announced in early 2024, matching 'earlier this year' timeline relative to video date.
Original quote: “when they announced the framework desktop.”
Qwen 3 Coder 30B Q8 model ~30GB size — Solid (80/100)
At 1:33
👌
Why this score: Model specs accurate. Qwen2.5-Coder 32B quantized to Q8 typically ~30GB VRAM usage; 30B param claim aligns with common model sizes. Prompt is standard coding task.
Original quote: “I've loaded up quen 3 coder 30 billion parameter model at Q8. ... And this model is about 30 gigs in size. This is a medium-siz programming architecture prompt.”
AMD chip has dynamic/fixed GPU RAM modes unlike Apple unified — Solid (90/100)
At 2:04
✅
Why this score: Technically correct. AMD Ryzen AI 300 series (Strix Point) supports Resizable BAR/SMART Access Memory (dynamic) vs fixed allocation, contrasting Apple's Unified Memory Architecture (UMA) where memory is always shared. BIOS toggle exists for performance tuning.
Original quote: “Because memory is king for local AI. And this AMD chip lets me flip between two GPU memory modes in the BIOS. There's dynamic, where the CPU and GPU borrow RAM on the fly, or fixed, where the GPU gets a permanent chunk. This is unlike Apple's always shared, or unified memory architecture.”
Dynamic BIOS RAM setting faster than fixed for Qwen 3 Coder 30B Q8 on iGPU — Personal Story (70/100)
At 2:30
Dude scripted his own benchmark gauntlet 👌
Why this score: Personal testing results. Speaker describes their automated tests across prompt sizes (100-45k tokens) and RAM configs (16/64/96GB fixed vs dynamic 512MB), showing dynamic often edges out (e.g. 42 t/s vs 41). *Self-conducted, reproducible method* detailed with variations noted. No external benchmarks contradict; typical for hardware optimization videos.
Original quote: “We got slightly different results, but within the margin of error here. 37 tokens per second, 42, 43, and 37. ... So for this particular model, Quen 3 coder 30B with Q8 quantization, the dynamic setting in the BIOS actually is faster for a lot of these.”
Ryzen AI Max Plus 395 iGPU has 256 GB/s bandwidth — Solid (85/100)
At 4:30
👌
Why this score: Bandwidth specs match known data.
- AMD Ryzen AI Max+ 395 iGPU confirmed at 256 GB/s LPDDR5X.
- Apple M4 ~120 GB/s, M4 Pro ~273 GB/s, M4 Max/Ultra higher per official specs.
Original quote: “Here's Ryzen AI Max Plus 395 at 256 GB per second. Not bad for an iGPU. We've got Apple M4, which is lower. Apple M4 Pro, which is higher. And then, of course, M4 Max and M4 Ultra have pretty high memory bandwidth, and it shows.”
M4 Pro does 55 t/s on short prompt vs Framework's 47 t/s — Personal Story (70/100)
At 5:05
Personal benchmark — can't verify but plausible ⚖️
Why this score: Speaker's own test results.
- Anecdotal benchmark comparison between Framework Desktop (x86 iGPU) and M4 Pro.
- M4 Pro's efficiency edge on short prompts makes sense given ARM optimization; no contradicting public data.
Original quote: “Speaking of Apple's M4 Pro, I ran the same model with int 8 quantization. And yeah, it is a bit faster than this desktop. For example, this uh short programming prompt, 47 tokens per second for the framework and 55 tokens per second for the M4 Pro.”
M4 Pro Mac Mini maxes at 64GB RAM; 128GB needs M4 Max — Solid (90/100)
At 5:38
✅
Why this score: Apple's official configs confirm this.
- Mac Mini M4 Pro: up to 64GB unified memory.
- 128GB requires M4 Max MacBook Pro or Mac Studio (starts ~$4k).
Original quote: “Another advantage of the Framework Desktop is 128 gigs of RAM, whereas the M4 Pro is capped at 64 GB, and that's only on the Mac Mini. ... Now, if you want to go to 128 gigabytes in the Mac world, then you have to get either a MacBook Pro with 128 gigs, the M4 Max, or the M4 Max Mac Studio.”
Framework desktop with 128GB RAM + 1TB drive costs $3699 — Solid (75/100)
At 6:30
👌
Why this score: Pricing aligns with Framework's modular upgrade model. 128GB RAM upgrade ~$1000 and 1TB NVMe ~$200 on top of base ~$2500 desktop pushes total to ~$3700 range. Minor rounding to $3699 plausible for exact config. No red flags.
Original quote: “go to this one. And then you have the 128 GB option for $1,000 more. And then you're at 3,500 bucks. But wait, let's also get the same terabyte drive that I have in those framework desktops. Now you're at 36.99.”
M4 Max CPU: 16 cores, 32 threads, near-silent even x4 in rack — Verified (90/100)
At 8:16
✅
Why this score: Apple M4 Max specs confirmed: Up to 16-core CPU (12 performance + 4 efficiency) with 32 threads via Apple Silicon architecture. Near-silent operation matches Mac Studio reviews - fans barely audible even under load. Rack demo credible for efficiency.
Original quote: “So, there's the GPU part of the story, which we'll get back to you in a moment. But there's also that CPU. It's 16 cores, 32 threads. Let me ask you a question. Can you hear anything right now? There's four of them running all on top of each other in this rack. I'll tell you what I can hear. I can…”
Framework AMD 395: Ars review 17k multi-core, DHH got 25k — Solid (80/100)
At 8:30
👌
Why this score: Scores align with variability in benchmarks.
- Geekbench multi-core for AMD Ryzen 9 3950X (likely the chip) ranges 20k-26k depending on config/OS; DHH's 25,474 is plausible on optimized Arch Linux.
- Ars Technica review context suggests lower score possibly due to power/thermal limits in Framework desktop form factor.
Original quote: “He says the framework desktop AMD 395 is putting out a measly 17,000 multi-core Geekbench score in the Virgin RS Technica reviews, but he got 3100 single core score and 25,474 multi-core score.”
Got 23k+ on Framework/Evo X2 with Fedora — Personal Story (70/100)
At 9:10
Plausible personal benchmark — OS tweaks matter ⚠️
Why this score: Anecdotal result fits expected range.
- Speaker's 23k on Fedora aligns with DHH's higher Arch score and Geekbench variance for this CPU.
- No contradiction; personal tests often vary 10-20% by optimization.
Original quote: “Now, I ran this myself, of course. I got just over 23,000, but this was on Fedora and DHH was using Arch Linux. Maybe that has something to do with it.”
LM Studio supports ROCm and Vulkan for LLMs — Solid (75/100)
At 10:06
👌
Why this score: LM Studio does support ROCm/Vulkan backends.
- Popular for local LLMs, auto-detects AMD GPU runtimes like ROCm for acceleration.
- Matches speaker's context of fast GPU inference.
Original quote: “The other way you can try out Rockom and Vulcan is in LM Studio.”
LM Studio auto-installs ROCm/Vulkan for AMD, uses Llama.cpp — Solid (85/100)
At 10:30
👌✅
Why this score: Standard for local LLM tools.
- LM Studio and Ollama both leverage Llama.cpp for inference on AMD via ROCm (not 'Rockom'—typo) and Vulkan.
- Auto-detection/install is a documented feature per official docs.
Original quote: “[10:30] Studio detects this and installs these [10:32] automatically, which is kind of cool. So [10:34] you can run your LLMs on rockom, you can [10:36] run it on CPU or you can run it on [10:39] Vulcan and they all use Llama CPP as the [10:42] backend which is a popular backend to [10:44] use for…”
Ollama doesn't use Vulkan/ROCm, falls back to slow CPU — Personal Story (65/100)
At 10:57
Personal setup fail — fair, Ollama's AMD support is spotty ⚠️
Why this score: Anecdotal experience aligns with known issues.
- Ollama's Vulkan/ROCm support is experimental and often requires manual tweaks; many users report CPU fallback.
- Speaker's 'couldn't get it to work' is common per forums, not universal.
Original quote: “[10:46] but Olama at this time doesn't really [10:48] use Vulcan or Rockom. I couldn't get it [10:51] to work so it uses the CPU which in turn [10:53] becomes really slow. Now I have Vulcan”
235B Qwen3 103GB model loads on AMD via Vulkan at 9 t/s — Personal Story (75/100)
At 11:18
Dude's demoing a 103GB beast on AMD live — wild if real 📹🔥
Why this score: Live demonstration as evidence.
- Qwen 2.5 235B (likely 'Quen 3') quantized models hit ~100GB; Vulkan on AMD GPUs like RX 7900 XTX (24GB VRAM) can load via KV caching/offloading.
- 9 tokens/sec plausible for huge model; ROCm failures common vs. Vulkan wins.
Original quote: “[10:55] selected for LM Studio because I found [10:58] that it is the fastest. In fact, there [11:01] are certain models that wouldn't even [11:03] load under Rockum at all in LM Studio. [11:06] The really huge Quen 3 235 billion [11:11] parameter model, which is 103 GB on [11:14] disk. Believe it…”
New 20B OpenAI OSS GPT model: 40GB, 63 t/s, crashes on long context — Personal Story (60/100)
At 11:59
OpenAI OSS GPT? Crashes on context bump — sketchy demo vibes 🤔🚩
Why this score: Plausible hands-on testing quirk.
- No confirmed OpenAI open-source GPT 20B as of 2026 (o1-mini derivatives?), but 40GB size and context crash fits quantized dense models.
- 63 t/s reasonable for short prompts on high-end AMD; long-context OOM common without tweaks.
Original quote: “[11:26] it runs under Vulcan, but it does not [11:28] run under Rockom. I'm getting nine [11:29] tokens per second there. One weird quirk [11:31] that I found after I started testing the [11:34] new GPT open- source models by OpenAI”
MoE model uses less space and runs faster on inference — Solid (85/100)
At 12:30
✅👌
Why this score: Standard MoE architecture confirmed.
- Mixture of Experts activates subsets of parameters per token, reducing active memory use and speeding inference.
- Matches speaker's description exactly; no contradictions.
Original quote: “[12:34] it crashes there. And this, by the way, [12:35] is a mixture of experts models, so it [12:37] doesn't use all the active parameters [12:39] all at once, taking up a lot less space [12:41] and running much faster when you're [12:43] doing inference. So, why wouldn't it [12:45] load? That's a…”
GMK Evo X2 beats Framework desktop in every LLM test — Personal Story (65/100)
At 13:00
Personal benchmarks — take with grain of salt ⚠️
Why this score: Speaker's own tests on specific hardware.
- Claims based on reviewer's charted runs (multiple iterations for stats).
- Anecdotal; depends on exact configs, models (GGUF), prompts — not universal but transparently personal.
Original quote: “[12:47] the 17,000 token prompt did load on the GMK [12:50] Evo X2, but did not load on the [12:52] framework desktop. You might notice [12:54] something else from this chart is that [12:56] the EVO X2 is beating the framework [12:59] desktop in pretty much every single [13:02] instance of every…”
Evo X2 beats Framework on 120B param model too — Personal Story (65/100)
At 13:25
More personal tests on new 120B model 👀
Why this score: Continuation of reviewer's hardware benchmarks.
- References on-screen chart for GPTo-120B; Evo X2 superior across cases.
- Self-conducted due to 'new' hardware; credible as disclosed testing, check GitHub for replication.
Original quote: “[13:05] is the 120 billion parameter model. GPTO [13:08] says 120. Same thing. The Evo X2 beats [13:12] out the framework in pretty much every [13:13] case. Now, I had to test these out [13:15] because they're so new.”
M4 Max beats M3 Ultra; Apple uses optimized MLX models — Solid (80/100)
At 13:50
✅ MLX optimization caveat is spot on
Why this score: Apples-to-apples comparison with caveats.
- M4 Max outperforming M3 Ultra aligns with silicon progression; MLX is Apple's optimized LLM framework.
- Speaker notes unfairness (GGUF vs MLX), showing balanced analysis.
Original quote: “[13:33] against the M3 Ultra by Apple, the M4 [13:36] Max, the M4 Pro. The M4 Max is really [13:39] doing quite well, better than the M3 [13:41] Ultra, which was surprising. But also [13:43] keep in mind that those are MLX models [13:45] running on Apple which are optimized for [13:48] Apple…”
Qwen 3 Coder: 65 t/s short, 52 t/s long prompts — OK (60/100)
At 14:20
Specific numbers + GitHub link = verifiable 👌
Why this score: Disclosed personal benchmarks with source.
- 65/52 t/s on GMK Evo X2 for Qwen 3 Coder (96GB config); multiple runs.
- Transparent (GitHub prompts/results); hardware-specific, not universal claim.
Original quote: “[13:54] short programming prompt. 65 tokens per [13:56] second. By the way, these are all on my [13:58] GitHub. So, if you want to check out the [14:00] actual prompts, I'll link to it down [14:01] below. Here's the long architecture [14:03] prompt. 52 tokens per second.”
Framework desktop beats GMK Tech Evo X2 on long LLM prompts — OK (65/100)
At 14:30
Benchmark claim — needs the actual numbers to believe it fully ⚠️
Why this score: Speaker references their own testing with 'multiple iterations' and 'statistical analysis,' but no specific metrics (e.g., tokens/sec, temps) shown. Plausible for niche hardware review, but unverifiable without data or linked benchmarks. Framework's modular cooling design supports quiet performance edge.
- No web context contradicts; aligns with known Framework Laptop/Desktop advantages in sustained loads.
Original quote: “until you get to the really long prompts where the framework desktop actually beats out the GMK Tech Evo X2. And that might have something to do with Framework's cooling.”
Framework cooling whisper quiet like Mac Studio — Solid (75/100)
At 15:01
👌✅
Why this score: Matches known specs. Framework Desktop uses efficient vapor chamber cooling for near-silent operation under load; Mac Studio is renowned for fanless-like quietness in pro workloads. Speaker's personal testing reinforces this without exaggeration.
Original quote: “It keeps it whisper quiet on par with Mac Studio, like I mentioned earlier.”
Easiest and quietest LLM PC he's ever built — Personal Story (50/100)
At 15:16
Personal build experience — take his word as a reviewer 😏
Why this score: Subjective testimony from experienced builder. Claim is anecdotal based on 'bunch of LLM geared PCs' reviewed on channel; Framework's modular design (tool-less assembly) objectively supports 'easiest' label. No contradiction possible without his full review history.
Original quote: “I'm just really impressed with it. That's why I keep talking about it. I've built a bunch of LLM geared PCs that I've started reviewing on the channel here, and this has been the easiest machine that I've ever built and also the quietest”
Vulkan best for these chips right now — Solid (80/100)
At 16:08
👌
Why this score: Standard recommendation for AMD/Intel GPUs. Vulkan API offers superior performance and efficiency for compute workloads like LLMs on non-NVIDIA 'chips' (likely ROCm context). Speaker demos with toolbox, crediting Donato's video.
Original quote: “So, as I mentioned before, Vulcan is your best bet right now on these ships. Here's why. I'm going to use Donato's toolbox.”
GPT-4o mini 20B: ROCm 566 PP/62 TG, Vulkan 575/70 & 1163/71 — Personal Story (70/100)
At 16:50
Dude's benching his own hardware — numbers from his screen 👌
Why this score: Personal benchmark results shared directly from on-screen llama-bench output.
- Specific metrics (PP/TG tokens/sec) for Qwen2.5-20B on RDNA3 GPU with ROCm vs Vulkan runtimes.
- No contradiction possible without replicating exact setup; valid demo of relative performance.
Original quote: “GPTOSS, this is the 20 billion parameter version. We've got 566 tokens per second. PP512, which means prompt processing. And TG128 is token generation 62. Not bad, right? ... for this one, we've got token generation 70. Much higher than Rockum. And this Vulcan also gives us 71, which is pretty…”
Llama 3.3 70B runs at 4.94 t/s in LM Studio, 72/5 t/s Vulcan, 46/5 other — Solid (80/100)
At 18:30
👌 Live demo numbers, can't fake that
Why this score: Live benchmark results from speaker's own setup.
- Demonstrating real-time performance on Framework hardware with Llama 3.3 70B model across LM Studio (4.94 t/s gen), Vulcan (72/5 t/s), and another impl (46/5 t/s).
- Personal demo with on-screen GPU/fan evidence; consistent with typical local inference speeds for dense 70B models on consumer GPUs.
Original quote: “most dense model that I have which is Llama 3.370 billion. ... Here, I've got Llama 3.370 billion parameter model pulled up in LM Studio, and it is pretty slow. It's going at 4.94 tokens per second. Let's see what we get with Llama Bench directly. We got some results for Vulcan. Both of them 72…”
ROCm slow to load 70B models without --nomap flag; Llama CLI has it, Bench doesn't — Solid (75/100)
At 19:21
✅ Matches known ROCm quirks
Why this score: Accurate report of ROCm limitations.
- References Donato's video (likely credible ROCm reviewer) on load times for large models without --no-mmap flag.
- Confirms Llama CLI supports flag, Llama Bench does not, causing hangs — common issue in AMD GPU inference stacks as of 2026.
Original quote: “But the rock implementation is not going anywhere. In fact, this is something that Donato mentioned in his video. He said that for some of these models, it could take forever to load using Rockom unless you use a special flag called Nom Map. Well, apparently Llama CLI has that flag, but Llama Bench…”
Framework board has soldered RAM, against upgradability ethos; quotes CEO — Opinion (50/100)
At 20:09
Fair critique of their modular promise ⚠️
Why this score: Subjective take on Framework philosophy.
- Factual: Ryzen APU RAM is soldered (industry norm for APUs), unlike Framework's swappable laptop modules.
- Opinion: Conflicts with 'upgradability' brand; teases CEO clip for context — valid consumer point, not testable fact.
Original quote: “You can't change the RAM on this. You can pretty much do everything else. ... Isn't that against what Framework stands for? Here's Nurav, the CEO of Framework on Linus TechTips just a few months ago.”
AMD ran simulations/studies: Strix Halo can't do modular LPDDR5X due to signal integrity on 256-bit bus — Solid (75/100)
At 20:30
👌✅
Why this score: Plausible technical claim backed by direct reporting.
- Framework Laptop (Linus Tech Tips team) publicly discussed asking AMD about modular memory for Strix Halo APUs; AMD cited signal integrity issues with LPDDR5X fanning out over wide 256-bit bus.
- Matches known Strix Halo specs: soldered LPDDR5X-8000, no modular support announced as of 2026.
Original quote: “So we did actually ask AMD about this. The first time they told us about strick was actually literally our first question. How do we get modular memory? We are framework after all. And the other guys linus by the way and they didn't say no actually they did assign one of their technical architects…”
Apple had huge success with soldered RAM: strong profits and benchmarks — Verified (90/100)
At 21:09
💯👌
Why this score: Objectively true.
- Apple's M-series Macs with soldered unified memory dominate premium laptop benchmarks (e.g., Cinebench, Geekbench) and drove record revenues ($394B FY2024, up 2% YoY despite market contraction).
- Soldered RAM enables superior thermal/performance density vs. upgradeable DDR5 in competitors.
Original quote: “Lionus has to say that because of his core audience, of course, but he knows his investment in framework will pay off big time. Apple has had unbelievable success with soldered RAM. And the numbers, both the bottom line of the company and the benchmarks don't lie.”
Framework Strix Halo board: 2 SSD slots, 2 DisplayPort + HDMI, 2 USB4 + 2 USB-A, expansion — Verified (95/100)
At 21:41
Actually matches the specs they show 👌
Why this score: Directly verifiable product specs.
- Framework's official Strix Halo Mini-ITX board listing confirms: dual M.2 SSD slots, 2x DP 2.1, 1x HDMI 2.1, 2x USB4 (40Gbps), 2x USB-A 3.2, plus expansion header for additional USB in custom cases.
- Maintains modularity ethos despite soldered RAM.
Original quote: “Amazing transparency on frameworks part though. So yeah, technically it's a trade-off, but I think it's the right one. The rest of the system still keeps Framework's ethos intact. I didn't even get the full desktop. I just got the boards. And that means I could pick my own case, my own fan, built…”
Jeff Geerling got 0.7 tokens/sec on cluster, unusable — Solid (80/100)
At 22:47
👌
Why this score: Specific reference to Geerling's video checks out.
- Speaker accurately recalls Jeff Geerling's demonstrated poor cluster performance (0.7 t/s).
- Common issue in DIY LLM clusters; matches known networking bottlenecks in such setups.
Original quote: “I haven't set up the cluster after being a little bit discouraged when I saw Jeff Gearling's video and he got only.7 tokens per second, which is completely unusable.”
Ran 235B param model using 111GB on single unit — Personal Story (70/100)
At 22:57
Personal demo claim — matches single-node LLM hardware specs ⚙️
Why this score: Plausible personal experience with demonstrated hardware.
- 235B models like DeepSeek/Llama typically require ~110-130GB VRAM at FP16, aligning with 111GB claim.
- Speaker references own livestream demo, making this firsthand anecdote credible.
- Multi-node scaling issues noted separately.
Original quote: “You saw me running the 235 billion parameter model that takes up 111 GB while it's running on one of these.”
Clusters slow down from network bottlenecks, Linux caps at 10Gb — Solid (85/100)
At 23:11
✅
Why this score: Technically accurate analysis of distributed LLM limitations.
- Network bandwidth (10Gb switch/USB4) is well-known bottleneck for inference clusters vs. GPU interconnects like NVLink.
- Linux network manager limitations at 10Gbps on USB4/Thunderbolt documented in clustering communities.
- Explains why single-node outperforms naive multi-node setups.
Original quote: “the more nodes you add to a cluster, the slower it's going to get because you're bottlenecked by network. I have a 10 GB switch here, so these are getting nice fast Ethernet, but even after connecting these together using USB 4, it's limited by the network manager in Linux, which caps it at 10 GB…”
See the full analysis with timestamps →