Hey, Dan. Hey, Alex. Good to see you. Today we're building a gaming PC that's [music] going to blow everything else out of the water, but we're not
gaming on it. So, all the gamers are going to leave some comments, I'm sure. We're not gaming on this thing? We're doing AI. >> Okay. We're doing AI
with that. >> With this. The reason I'm a little confused by you saying that because it's not just like
any 5090, it's a lightning 5090. This thing is like what? 5 grand? Yeah. >> Yeah. This is not a cheapy and it's a big card. Of course, we have the
king of AI cards. It's not a very fancy box, but this is the RTX Pro 6000, which Micro Center carries, of course. But, I have a feeling that the that
GPU is going to blow this thing out of the water as far as speed. And I'm not sure. We'll test this
out. All right, this is a regular 5090, liquid cooled, also MSI. And this is [music] the box compared to that one. Yeah, big difference. So, we'll
test this against the 5090 and the RTX Pro 6000 coming up later in the video. But, for now, Dan, we're going to build this computer out. >> Yeah,
let's do it. And actually, I just received something that might be very [music] appealing for what we're
doing here. New bundle just came in. We're going to take advantage of this thing. >> So, that's really relevant because for like the last month or
two, we haven't had any bundle that does 64 gigs of memory. So, if you were going to get that 64 gigs, you were spending an extra 700 bucks on top of
everything else you were doing. And yeah, I am going to do a little gaming on this, but it's mostly for
AI. Tell me where we should start. >> I think we should start with your case because >> Ooh, we have a special case. He showed me this case a few
videos ago. You might remember that. [music] We're going to have it today. Oh, yeah.
Ah, I see it. I see it. >> [laughter] >> There we go. Nice. I've been a Doom fan since the first Doom and have been following Doom all along every
step of the way and this is a beautiful [music] case. All right, what else do we need? A cart because I don't think we should have to carry this
around. >> Ah, the mother wall. Look at all those motherboards. This place always impresses me with [music]
the amount of stuff they have here. Do you need help with that? Nah, I need to get my workout in. We're not doing push-ups today. Thank God.
Cameraman. >> had my caffeine yet. >> [laughter] >> That's my excuse. >> All right, let's go get the next part. I assume you have a power supply
already, right? >> Are you saying that's not in the budget? Woah. The reason I say that is because your graphics
card is going to need a little bit more than what a 1200 or a 1300 watt would give you. >> Uh-oh, what do we need? >> You need to have two 16-pin
connectors. Is this a computer? What is this? >> This is like a like a Windows tablet. That's the only way we can get the actual software to run on
it, but then people can just play around with the lighting however they want. Pretty neat. Actually, let's
just do the cooler first. Okay. We need to get a fan cooler. So, we have 420 on the case. I was thinking 620 on the cooler. >> Okay. I mean >> a
little pattern going. Yeah, it's a fun cooler made by Trix, Trikes, whatever. I I always forget their pronunciation. Uh they're the guys who kind of
brought out that panoramic display to the mainstream, and now this is a pretty powerful air cooler that
also has a really big screen on it. >> Woah. >> [laughter] >> This is going to be a bright machine. Can I display anything I want on that? >> You
could play that entire like E3 demo for Doom. So, [laughter] why not? Nice. All right, let's do that. Thanks for doing the videos, by the way. Yeah,
it's been fun. It's a lot Yeah, I know. I have a fun time every time. >> All right, what do we got? >>
Super Flower, for those who don't know, they are the OEM for EVGA's uh previous like higher-end lineup, and EVGA carries a lot of respect as a brand
because they put out really good product. So, 500 bucks, it's [music] Cybernetics A on noise level, which means this thing's going to be pretty quiet.
And it's titanium, which the Seasonic, to be fair, is also titanium, but it's 100 less watts as
far as its rating goes, and it's on sale for $429. Wow. Yeah, so no-brainer here. All right, let's do it. Let me verify what RAM it is real quick that
they just cuz that bundle just literally just started today. >> Okay, it's my lucky day I guess. Yeah. Oh my gosh, these RAM prices are still insane.
I don't know what that is. Wow, that's [music] an entire screen on a cooler. I always find the
coolest stuff in the store. It's just very dangerous to come in here for me. I think I'm going to do some shopping today, too, Alex. You're not
allowed. You're an employee. Oh, man. Yeah, he's right. He's right. [laughter] But check that out. Really decent discount on the processor. It's
almost [music] 50% off. Motherboard is at price, but all five of these components, two of which we're not
using, 999.99. That is crazy. So, always check the website to see what kind of deals they have on bundles that are new. And I should mention that the
RAM on the bundle is CL30, which is very nice latency. It's like the sweet spot for AM5. And is that RAM with RGB on the RAM? >> So, you'll have some
lighting in there. We're not going to go super tacky with it. >> Okay, very nice. Yeah. Oh my gosh,
that's the current price of just the RAM? Well, actually it's on sale for 999.99, but in your case you're getting it for 507 something. >> No way.
Okay. Okay, is that the motherboard? Here's our motherboard, X870E in all its glory. So, you're going to get your Gen 5 on the PCI slot for the
graphics card and on the SSD should you choose to do a Gen 5 SSD, followed by our processor, the 9800X3D.
Some of you guys may say, "Well, the 9850X3D is out there and it's faster, but like all it is a 9800X3D that runs at 400 MHz faster on it." So,
practically the same processor and some of these, if you win the silicon lottery, you can be even better than that. >> Silicon lottery? What's What's
that? All right, what do we got? [music] We got everything in the cart. Uh I think we're missing storage.
Yes. So, right now we are pre-tax 1890 [music] basically. Ooh, look at that. Up to 14,900 MB per second. Perfect. I really like the packaging on
these. Like the it's just very slick. >> It looks like prices might be going down. This is 329 on the sticker. I think that just might be because
Crucial's out of the consumer market. Okay. And then boom. Uh that's our entire machine. And then the
pre-tax is 2110. So it's This is the most expensive gaming machine I've ever built. Just a couple more fans cuz of the cooling is going to be
important factor in this build. >> I just really like the way that these Arctic Pro fans perform. They have very high static pressure. They're PWM and
they're not super loud. >> And they're not expensive. No, no. If you're going for like a really high
performance fan for like a radiator Uh-oh. Ignore that. If you're going for like a very high performance fan for a radiator, these are the ones to
get. They have a static pressure of over six. You got it already. Ready to go. Yeah. Well, this is like two builds now. I'm going to have to put
something in that box, too. So now it's time to talk about the 5090 Lightning, why it's special and why is
it in such a huge box. Dan? Well I'd be more than happy to talk about it, but Alex, I think you should open it up and show them first. >> Okay, let's
open it up. >> Oh. This is the box that matters. >> actual box. >> Yeah. Whoa. >> [laughter] >> Oh my gosh. Built to be perfect. >> three boxes. The
card box, accessory box, and then there's a box for the vertical mount. >> All right, so this is the
actual box. Ooh. Wow. Okay. 32 GB DDR7. And the different thing about this is that it's 1,000 W. Unlike a typical 5090, which has up to 600 W. And why
I said that this might be faster than the RTX Pro 6000 is because the Pro 6000 also is limited to 600 W. All right, let's take a look. I got to take
the sleeve off. >> Oh yeah, yeah. No, it was the right side up. Right side up. Believe me. You will
you will regret it if you do the other way. Okay. Now, does it Oh. There you go. Oh my gosh. No way. Woah. That is a thing of beauty. I did not get
the reference, but comment down below if you did. And I know somebody that's going to be jealous. 48 out of 1,300. This is the card, and this is the
cooler that's attached to it. The liquid cooler. And this card is actually smaller than the Gigabyte
one that I have. Insane. Okay, let's build this thing out and then we'll test it. All right, then. Thanks a lot again. Appreciate it. >> you. I like
the bundles that this guy finds. Come see him at the Micro Center in Rockville. Yeah, please do. Why not? We can have a fun time. Every time people
watch the channel and they come here, the first thing they say to me is, "You got nice hair." And I'm
like, "Ah, you watch Alex." Also check out the Austin store, which is going to be opening up pretty soon. Keep an eye out for that. You may see me
there, too. So. So, here it is on my desk. Case number 27 out of 1,666, and GPU number 48 out of 1,300. Yeah, both limited edition. And I finally have
a computer that matches this keyboard. All right, enough showing off. Sorry about that. >> [laughter]
The thing is, everybody else who's tested this card has tested it for games. That's it. Nobody pushed it on AI. And this card has a 1,000 W power
cap. 1,000 W. So, the question I want answered is what does it actually do for AI? I have my hunches. This is going to be for local LLMs, for video
generation, the workloads I actually care about. Let's find out. So, here's the lineup. The regular RTX
5090, 32 gigs of VRAM, 600 W power cap. The 5090 Lightning Z that we just built. Same VRAM, 32 gigs, but a 1,000 W power cap. I already mentioned that
a few times, I think. Hopefully, it'll pay off. And right here, I have the very quiet RTX Pro 6000 Black Well with 96 gigs of VRAM. This is the
workstation edition, which is also capped at 600 watts. Now my lab is all set up, but I have all three of
these machines on one circuit, and because this is United States, I had to run my tests one machine at a time so I don't trip the 1500 max circuit.
Yeah, you can laugh at me in European languages. All right, first up Qwen 3 4 billion parameters. Now the BF16 version. Okay, so it's a little bit
heavier. I know it's a small model, okay folks, chill out. Yeah, I always start out with it just to get a
baseline, so let's see what we get. Single stream prefill. Single stream meaning like you're chatting with it basically. The regular 5090 gets 27,000
tokens per second. By the way, prefill is the first stage. It's the prompt processing. That's how fast a GPU can crunch those numbers, not generate
text. We'll get into that momentarily. Here comes the 5090 lightning, 29,414. That's about 8% faster.
Okay. Not nothing, but check this out. >> [laughter] >> The RTX Pro 6000 gets 49,520. Almost twice as fast as either 5090. But here's the thing. Now
we're going to switch and talk about decode, and this is the token generation part. The regular 5090 gets 160. The lightning gets 162. And the Pro
6000? 151. So in this case, the most expensive card is the slowest card if you're just talking to it.
And on the lightning, power draw was only 553 watts. So the card has a 1000 watt budget, and the chip isn't even asking for 600, which means all the
headroom is just sitting there wasted on a 4 billion parameter model. Okay, let's try a bigger model. Qwen 3.5 9 billion. By the way, I'm looking over
here because that's where my charts are. I don't have space on desk anymore, all right? Too much
stuff here. P50 and P95 and max. P50 is the median, basically where the GPU sits most of the time as far as power usage. P95 is the top 5% of samples,
that's the high end, and max is the single peak that we reached. So, when I say the card is pulling 670 watts, that's sustained. That's where it
actually lives, not just a one-time occurrence. We're seeing a little bit of an improvement in
performance here on prefill over the regular 5090, but not that much. And this 600-watt card, the 6000, is clearly doing better. I'm more curious
about the power usage here and whether we're using all the power that we can in this GPU. So, what's going on with the power? Well, the lightning here
on P50 is pulling 670 watts sustained. The regular 5090 is obviously throttled at 599, that's because
this is a cap. So, the lightning is doing about 70 watts of compute that the regular 5090 cannot, literally. And that translates directly to about 7
to 12% more throughput here that we see in the prefill. This is actually good. This is the first lightning win for us. So, I'm thinking we got a
bigger model, there's more demand on the GPU compute, so we need more of models, more parameters, right?
Well, then I tried 14 billion BF16, 14 billion parameters, and BF16 is the quantization or the lack of quantization. It's 28 gigs of weights. Notice I
said 28 gigs of just weights, these two GPUs only have 32. Well, guess what? It didn't fit. I'm running the vLLM, so the model plus vLLM's CUDA graphs
spilled over the 32 gigs of VRAM available. So, I had to tweak a few flags, had to fall back to
eager mode, which basically kills kernel efficiency. Both 5090s collapse at concurrency of eight. Concurrency is how many requests we're sending at
the same time, which is something that pro cards are really good at handling. These uh GeForce cards, they handle it, but not as well. So, you'll see
that our 2 5090 are almost the same with the Lightning one being a little bit ahead. And here's the
Pro 6000. >> [laughter] >> You can see that at eight concurrent requests, the regular 5090 dropped from 8700 tokens per second all the way down to
about 3300. Lightning about the same, and Pro 6000 didn't even flinch. 12,362 tokens per second because it's not running out of memory. And that, my
friends, is why I love this GPU, the 6000. Even though it doesn't come in a pretty box. So, 14B BF16 on
a 32 gig card is just kind of sad. And the wattage debate doesn't even matter if you can't even fit the whole model in. Now, of course, there's
different quantizations we can try, so don't get discouraged. You can still run a 14 billion parameter model on the 32 gigabyte GPUs. You just have to
use quantized versions. And that's really not what I'm testing, can it run or not. I'm testing pushing it
to see how far the Lightning can go. So, I wasn't satisfied yet. I needed a workload that could actually saturate the Lightning. To do that, I went
back to the 9 billion parameters where CUDA graphs worked, and I just cranked the concurrency. So, this is the prefill versus concurrency chart again.
At one request, they're pretty much tied, both 5090s at 15,000 or so. 16 requests, we got a little
bit of a difference there. The Lightning starts pulling ahead. 32, still pulling ahead. 64, still pulling ahead, about the same. And then, at
concurrency 96, that means there's 96 concurrent requests going on at the same time to the GPU. The regular 5090 just falls off. It dropped to about
10,900. Why? Well, because it tried to throttle clocks to stay under 600 watts. The Lightning at 96
concurrent, 13,150. That's 20% more juice that we just squeezed. Uh >> [sighs] >> I should be careful with juice around these GPUs. This is a workload
where we actually see a payoff for the 1,000 watts that we just paid for. Is it that much of a payoff though? Cuz here's what's bugging me. Across
every LLM test that I ran, the lightning still never crossed 695 watts. The chip has a 1,000 watt
budget. Why isn't it asking for more? So, correct me if I'm wrong in the comments, I'm sure somebody will. My theory's that LLMs are mostly waiting on
memory. But, perhaps image and video generation will give us a different result. In image and video, every diffusion step is dense matrix math. So, I
installed ComfyUI on all three and ran WaN 2.2 5 billion video generation. All right, the regular
5090 finished that in 113 seconds. It pegged at 600 watts for half the run. So, we squeezed one more watt out of it. 600 is max. The lightning came in
at 109 seconds, a little bit faster, with 639 watts P50. And the Pro 6000, 120 seconds, the slowest of the three on a single video. Basically the same
story as chat. Lightning is performing better, but only slightly. And it still topped out at 652
watts. So, I went and checked persistent mode, power limits, clock limits. I re-ran the whole thing, and I got the same number. The chip just doesn't
want more than 650 watts on diffusion. And then I was like, "Oh man, what did I just pay $5,000 for? And can this chip even do 1,000 watts? Or is it
just a spec they put on paper?" So, I wrote the dumbest possible test. A few lines of PyTorch, just
torch matrix multiplication on a 16 by 16 BF16 tensor in a tight loop. No memory pressure, no KV cache, just pure tensor core hammer. I ran it. All
right, we're doing matrix multiplications. This is all the power we're getting out of it, which is way more than we were getting before. Okay, we're
over 1,000 watts here. 1,139. For the first time I'm seeing this during GPU burn. P50 power went to 760
watts. P95 hit 1,000 and the max 1,018 watts. Briefly over the cap. Still didn't blow any fuses though. And the top 20 samples were all at or above
999 watts. So, the chip just parked right at the cap and sat there. Sustained throughput 257 BF16 teraflops, 716 FP8 teraflops. That's what a 1,000
watt cap was built for. So, here's the actual answer. The Lightning 1,000 watts is real and the chip
will draw every watt of it. But, inference and video generation have built-in pauses. There's memory reads, there's attention masking, there's things
that prevent tensor cores from staying saturated. Those workloads plateau around 650 watts. The Lightning's extra headroom only opens up when you're
training or hammering pure matrix multiplications. So, should you buy Lightning? Well, if you can
still find one and if you're training models or you're running sustained compute and if you really want to look cool, cuz it is a pretty cool looking
card, then yeah, go for it. If you're just running inference or generating images, you're paying for headroom that runs about 20° cooler, which might
be worth it on its own. It's also way quieter than a regular 5090. And at the end of the day, this
build right here is a really cool gaming rig that I only could have ever dreamed of before. And it also happens to do AI surprisingly well. But, if I
were doing AI and only AI, no games, no spectacle, just pure throughput, RTX Pro 6000 all the way. Which one would you have bought? Let me know in the
comments down below. I actually did a more detailed comparison of the Pro 6000 versus the 5090 and
what the 96 GB actually unlocks in this video right over here. Go ahead and watch it. Thanks for watching this one. [music] I'll see you next time. >> [music]