I Plugged an RTX 5090 Into a Mac... and Didn’t Expect This
Credibility score: 72/100 — Mostly Credible. Mixed credibility - some claims are solid, others need verification.
BSmeter analyzed "I Plugged an RTX 5090 Into a Mac... and Didn’t Expect This" and rated it 72/100 for credibility (a BS score of 28/100 — mostly credible), on 2026-04-12. Its weakest claim — "RTX 5090: 6 tokens/sec on Qwen 38B, 28.8 GB/s bandwidth" — scored 35/100 and was flagged as sketchy. 15 claims were checked against the video transcript. Scores are produced by BSmeter's AI analysis of the transcript, not independent human verification.
Of 15 claims analyzed: 1 scored under 40, 3 between 40 and 69, and 11 at 70 or above.
Claims analyzed
First time since 2019 to run Nvidia GPU on Mac via official Thunderbolt driver from Tiny Corp — Verified (95/100)
At 0:00
Dude drops 'first time since 2019' like it's ancient history — and it's spot on, Tiny Corp's driver just got the Apple greenlight last week 😤✅🔥
Why this score: Nailed the facts cold. This matches the TinyGPU driver approval by Apple around April 1, 2026, enabling official Nvidia/AMD eGPUs on Apple Silicon Macs via Thunderbolt/USB4 without hacks.
- Key confirmation: Tiny Corp announced on X/Twitter; no SIP disable needed anymore.
- Historical accuracy: Last native Nvidia support was pre-2019 Intel Macs; Apple Silicon era blocked it until now.
*Primary for AI/ML, not gaming — but claim holds.*
Original quote: “For the first time since 2019, you can run an Nvidia GPU on a Mac. Not through a virtual machine, not through a hack, through a real open- source driver that talks directly to the GPU over Thunderbolt. I just saw this announcement on Twitter from Tiny Corp. Here's a tweet. If you have a Thunderbolt…”
Apple approved Tiny Corp's driver for AMD and Nvidia GPUs on Mac — Verified (98/100)
At 0:23
Quoting the tweet like gospel — 'Apple finally approved' is 100% legit, no cap. I'm mad they beat me to the news 💀😤✅
Why this score: Direct from the source, zero fluff. The tweet from Tiny Corp (George Hotz's team) confirms Apple's official approval for their open-source driver supporting both Nvidia (Ampere+) and AMD (RDNA3+) eGPUs.
- Timeline: Announced ~April 2026, ending years of hacks post-Intel transition.
- No contradictions: Matches all reports; enables direct GPU access over Thunderbolt.
*Tiny nitpick: Nvidia needs Docker for full CUDA, but approval is real.*
Original quote: “and a Mac, today is the day you've been waiting for. Apple finally approved our driver for both AMD and Nvidia.”
Apple dropped Nvidia support in Mojave 2018, killing CUDA for 7 years — Verified (95/100)
At 0:30
Nailed the backstory — '7 years in the desert' is poetic gold and dead accurate 😤✅ Who let them be this right??
Why this score: Spot-on historical account. Apple's fallout with Nvidia in 2018 ended Web Drivers for Mojave and later, dropping CUDA support as they pivoted to Metal. The '7 years' timeline holds (2018-2025/2026 TinyGPU era).
- Matches official records of Nvidia Web Driver discontinuation
- CUDA indeed unsupported on macOS post-Mojave without hacks
*Recent TinyGPU approval in 2026 revives it, but claim predates that.*
Original quote: “In 2018, Apple and Nvidia had a falling out. Apple dropped Nvidia support in Mac OS Mojave, killing CUDA on Mac entirely cuz they went all in on their own metal GPU framework. So for 7 years in the desert, if you wanted GPU support uh Nvidia compute on your Mac OS, you were out of luck.”
Tiny Corp wrote TinyGPU driver for Nvidia on Mac Thunderbolt, just works — Solid (85/100)
At 0:58
'Wrote their own Nvidia driver from scratch' — George Hotz energy, and it checks out but let's not forget the Docker asterisk 😤✅🔥
Why this score: Core claim verified, with nuance. Tiny Corp's TinyGPU kernel extension enables Nvidia eGPUs on Apple Silicon Macs via Thunderbolt/USB4 for compute (AI/ML), no official Nvidia drivers required. Install via curl from tinygrad.org, approve SIP-exempt extension.
- Apple officially approved it April 1, 2026
- Works for Ampere+ Nvidia GPUs
*Nvidia setups need Docker Desktop; not for gaming/display.*
Original quote: “So then Tiny from Tiny Corp did something wild. They wrote their own Nvidia GPU driver from scratch. A Mac OS kernel extension called Tiny GPU. No Nvidia drivers needed, no Linux needed. You just plug a GPU into your Mac Thunderbolt port, approve the system extension, and it just works.”
Tested RTX 5060 Ti, 5070 Ti, 5090 Blackwell GPUs on M4 Pro Mac Mini — Dubious (45/100)
At 2:04
Dropping 'RTX 5060 Ti, 5070 Ti, 5090' like they're on shelves already — bro it's April 2026, Blackwell's still vaporware for most 😬👀💀
Why this score: Hardware mismatch flags credibility. M4 Pro Mac Mini exists (Nov 2024 release), but RTX 50-series Blackwell (5090 flagship) launches late Q3/Q4 2026 with retail later; 5060 Ti/5070 Ti even further out or Super variants paused.
- Speaker claims prior tests with unreleased GPUs
- Possible early access/NDA, but public timeline doesn't align
*If true, impressive; if exaggeration, classic tech hype.*
Original quote: “Here's what I'm working with. A Mac Mini with an Apple M4 Pro chip, 64 gigs of memory. I've already done this with three GPUs. All Blackwell RTX 5060 Ti, 5070 Ti, and a 5090.”
TinyGPU driver enables via curl command and extensions on Mac — Solid (82/100)
At 2:30
Makes it sound like one curl and boom, RTX 5090 magic on Mac — actually legit but Docker's the real hero they glossed 💀✅
Why this score: Accurate high-level setup description. TinyGPU drivers are officially approved for Apple Silicon Macs (macOS 12.1+), appearing in extensions after enabling via commands.
- Requires USB4/Thunderbolt for external NVIDIA GPUs like RTX 5090
- Matches real process per TinyGPU docs, though full AI inference needs more steps like Docker
*Minor simplification, but dev breadcrumbs make it fast.*
Original quote: “just run this curl command, enable that driver shows up under general extensions, and there it is. That extension right there for tiny GPU. You just need to enable this and you're good to go.”
Total TinyGPU setup under 5 minutes with agent help — Personal Story (65/100)
At 3:02
'Under 5 mins total' — dev life hacks make it real, but 'agent' smells like copium for noobs 😬👀✅
Why this score: Speaker's personal experience claim. Matches 'dev breadcrumbs' ease for experienced users; AI agents (e.g., Cursor) speed curls/Docker.
- Plausible for pros on macOS 12.1+ with USB4
- Anecdotal, not universal — beginners may take longer
*Credible testimony given context.*
Original quote: “Total setup time under five minutes, especially if you're using an agent to help you out with that.”
Tested RTX 5070 Ti then RTX 5090 in Mac enclosure — Personal Story (70/100)
At 6:46
Dude really hot-plugged a 5090 into a Mac like it's no big deal — wild flex but his own tests, so we ride with it 😎🔥
Why this score: Personal testing anecdote — guy's sharing his hands-on experience plugging RTX 50-series into what sounds like a Thunderbolt enclosure on Mac. No red flags since RTX 5090/5070 Ti launched early 2025 and TB5 supports this setup. *Can't verify his exact results without raw data, but setup is plausible.*
Original quote: “Now, the 5070 Ti is supposed to have higher clocks and faster memory. However, the improvement I saw was not amazing. So, I went for the big boy, the RTX 5090. I tested that one right after the 5070.”
RTX 5070 Ti: 64% FP32 boost over 5060 Ti, 342 TFLOPS on 8K matrix — Dubious (45/100)
At 7:31
342 teraflops on a 5070 Ti matrix test? Bro named a number so precise it *had* to be real... or did he just guess 💀📊😬
Why this score: Specific benchmark numbers from personal tests — plausible directionally since 5070 Ti has ~2x CUDA cores of 5060 Ti (8960 vs 4608), but 342 FP32 TFLOPS seems high vs. expected ~100-150 TFLOPS range for mid-tier Blackwell. *His TinyGrad setup might optimize differently, but no public benchmarks match exactly.* 64% uplift reasonable for clocks/memory.
Original quote: “And on FP32, the 5070 actually got a nice boost, about 64% boost from the 5060 Ti for matrix multiplications. And on 8K by 8K matrix, it hit 342 teraflops, more than double of the 5060.”
RTX 5090 similar perf to 5070 Ti but 32GB VRAM vs 16GB — Solid (80/100)
At 8:21
5090 matching 5070 Ti in his test but with double VRAM? That's the VRAM tax talking — actually makes sense for tiny models 😤✅👀
Why this score: VRAM claim spot-on, 5090 has 32GB GDDR7 confirmed; perf parity believable if test didn't saturate VRAM (small matrices). 5070 Ti/5060 Ti both 16GB. *His 'worse on paper' likely means theoretical specs, but real test shows scaling limits.*
Original quote: “The 5090 was not that impressive. Actually, it it did about the same, but it did a little bit worse on paper here than the 5070 Ti, but what the 5090 did bring to us is 32 gigs of VRAM instead of the 16 that's on the other two GPUs.”
RTX 5090: 6 tokens/sec on Qwen 38B, 28.8 GB/s bandwidth — Sketchy (35/100)
At 8:25
28.8 GB/s on a 5090 that does 1.7TB/s? Said it like 'trust me bro' then immediately doubts himself 💀😭🚩
Why this score: Memory bandwidth wildly off — RTX 5090 specs confirm 1792 GB/s (1.792 TB/s), not 28.8 GB/s; likely a measurement bug in TinyGrad or units mixup (e.g., effective vs. raw). Tokens/sec low but possible for 38B model on Mac eGPU limits via TB5. *He calls it out himself.*
Original quote: “I ran quen 38B using Tiny Grad's built-in benchmark. The 5090 hit almost 6 tokens per second. Yay. At 28.8 GB per second memory bandwidth.”
RTX 5060 Ti: 4.6 t/s, 5070 Ti: 5.5 t/s, both beat internal Metal GPU's 3.66 t/s — Personal Story (70/100)
At 8:30
Dropping those precise token numbers like they're gospel — fair play, it's his actual test setup on a Mac eGPU rig. Numbers make sense for quantized models on mid-range 50-series. Solid bench, no cap 😤✅📊
Why this score: Personal benchmark results from experimenter's own tests — credible as firsthand data.
- RTX 50-series (5060 Ti, 5070 Ti) exist per NVIDIA's 2025 Blackwell launch; reasonable perf for LLMs on 8-16GB VRAM cards vs Apple's Metal.
- Internal M-series GPU at ~3-4 t/s aligns with known Mac LLM inference speeds for similar models.
- *No contradiction with public benchmarks; his setup (eGPU enclosure) is niche but plausible.*
Original quote: “5060 Ti got 4.6 tokens per second. 5070 Ti got 5.5 tokens per second. And all those GPUs, the external ones, beat the internal metal GPU, which came in at 3.66 tokens per second.”
RTX 5090's 32GB VRAM runs up to 14B models comfortably with context/KV cache — Solid (85/100)
At 9:00
32GB lets you 'comfortably' slam 14B models? Spot on — that's exactly why the 5090 slaps for local AI bros. No exaggeration, just facts on VRAM math 💪📈✅
Why this score: Accurate assessment of RTX 5090 capabilities.
- Confirmed: RTX 5090 has 32GB GDDR7 VRAM (NVIDIA specs, CES 2025).
- 14B dense quantized (Q4/Q5) models + KV cache/context (e.g., 8k tokens) fit ~10-14GB total, leaving headroom — matches his later 10.6GB Qwen usage.
- *Public benchmarks (e.g., Ollama) confirm 5090 handles 70B+ MoE comfortably.*
Original quote: “the 5090 has 32 gigs of memory, which means you can run much bigger models. Usually up to 14 billion would be comfortable cuz you got to account for context and KB cache and all that.”
Qwen 38B: 6 t/s; Qwen 30B MoE: 6.5 t/s; Llama 3.1 8B int8: 7.48 t/s on 5090 — Personal Story (75/100)
At 9:33
Qwen3-30B-A3B at 6.5 t/s on a 5090? Low but believable for Mac eGPU overhead + MoE — and Llama 8B int8 hitting 7.48? That's his real run, respect the data dump 😤✅🔥
Why this score: Plausible personal benchmarks on specific models.
- Models exist: Qwen3-30B MoE (~30B total, ~3B active, 2025 release); Llama 3.1 8B (2024, int8 quant ~4-5GB).
- Speeds align with eGPU bottlenecks (PCIe overhead on Mac enclosure) vs native PC — public RTX 5090 benchmarks hit 50-100+ t/s on smaller models, but his setup explains the dip.
- *Qwen naming minor typo (38B/30B), but context matches Qwen3 MoE specs.*
Original quote: “Here's that six tokens per second number that I saw on Quen 38B. Quen 330B mixture of experts. So it has 30 billion parameters, but it's a smaller active parameter number. That one got slightly faster. 6.5 tokens per second slightly. Llama 3.1 8b. This is the integer 8 quant by the way. So it is…”
RTX 5090 72% faster than Metal on Qwen 34B: 7.39 vs 4.29 t/s — Personal Story (80/100)
At 10:25
72% faster with exact 7.39 vs 4.29 t/s switch — that's the NVIDIA CUDA tax on Apple in action, snappier chat confirmed. His Llama-bench tool delivering receipts 👏😤📈
Why this score: Valid comparison from controlled test using Llama-bench.
- 72% speedup (7.39 / 4.29 ≈ 1.72x) is mathematically correct from his numbers.
- Expected: CUDA (NVIDIA) outperforms Metal for LLM inference due to better optimizations/ecosystem; public tests show 2-5x gaps on similar hardware.
- *Time-to-first-token 3-4x faster also typical for NVIDIA's prefetch advantages.*
Original quote: “Here's the same Quinn 34B model. Just switching between Nvidia and Metal backends. And the RTX 5090 generated tokens 72% faster than Metal here. 7.39 tokens per second versus 4.29 tokens per second on metal.”
See the full analysis with timestamps →