0:03 Nvidia had a lot to announce in its 93
0:06 minute consumer electronic show keynote.
0:08 >> We would like to have this AI stay with
0:10 us our entire life and remember every
0:11 single conversation we've ever had with
0:14 it. Right? People ask where is the money
0:15 coming from. >> Yes,
0:15 >> Yes,
0:18 >> this last year was incredible.
0:22 This last year there's a slide coming.
0:23 This is what happens when you don't practice.
0:26 practice.
0:28 It's the first keynote of the year. I
0:29 hope it's your first keynote of the
0:30 year. Otherwise you going you have been
0:32 pretty pretty busy basic way of building
0:43 These models are also world class.
0:51 This never happens in Santa Clara.
0:53 I think my system's still down. But
0:56 that's okay. I I I've I uh I'll make it
0:58 up as I go. brought to you by the most
1:00 powerful tech company in the world and
1:02 the company in charge of the global
1:04 economy. So, it's not the strongest
1:06 start, but you know, Jensen, AI is a
1:08 little bit scary for a lot of people.
1:11 So, give us something cute, something
1:13 that helps us drop our guard, but maybe
1:15 if you think about it a little too much,
1:16 it's still terrifying. You got anything
1:17 like that?
1:18 >> With Brev, I can share access to my
1:20 Spark and Reachi. So, I'm going to share
1:26 >> Hey Richi, what's Potato up to?
1:29 He's on the couch.
1:30 >> I remember you don't like this.
1:32 >> I'll tell him to get off.
1:35 >> Aa la vista, baby.
1:37 >> That's right. Nvidia saw Dogtober and
1:48 >> This is This is Wow. It's super heavy.
1:50 You have to be a CEO in really good
1:59 Okay. All right. So, this thing is I'm
2:02 going to guess this is probably I don't
2:17 >> Have you GUYS HEARD THIS ONE?
2:18 >> You laugh that way because you know
2:19 you're wealthy. Is that
2:21 >> I'm very wealthy and then that's
2:23 >> that's odd. I That's how wealthy people laugh.
2:23 laugh.
2:26 >> I WOULDN'T HAVE GUESSED. COME ON. It
2:27 could have been.
2:28 >> Do another one. Do another one. Tell
2:29 another joke.
2:32 >> It's usually about two tons, but today
2:35 it's 2 and 1/2 tons because um when they
2:37 shipped it, they forgot to drain the
2:40 water out of it.
2:42 So, we we shipped a lot of water from California.
2:48 It's funny because there's a localized
2:50 water crisis all over parts of the
2:51 United States because data centers are
2:53 taking all the water and took a thousand
2:55 pounds of water and put it in a box and
2:56 shipped it from California to a desert
2:58 to put on stage. It's also at a time
3:00 when the concentration of nitrates, the
3:01 percentage of it is increasing in some
3:02 areas of data centers because the data
3:04 center usage of water and also there's
3:06 sediment buildup in some people's water
3:09 where they can't drink it anymore.
3:12 >> Can you hear it squealing?
3:19 Oh, you could do it. Wow.
3:21 >> Nvidia did though have other important
3:23 and good news. In case you were
3:26 wondering, Nvidia is still accelerating
3:28 everything Palunteer does.
3:31 >> Uh Palunteer, for example,
3:34 um their their entire AI and data
3:36 processing platform is being integrated,
3:38 accelerated by Nvidia today.
3:40 >> That's right.
3:41 What is it Palanteer does again? The
3:44 data mining firm Palunteer faces
3:46 backlash over its partnership with ICE.
3:47 The government agency paid the company
3:49 $800 million last year.
3:52 >> Our product is used on occasion to kill people.
3:53 people.
3:56 >> Ah, good. Killing people. The good news
3:58 is that Jensen Juan doesn't need people
3:59 where he's going.
4:00 >> That's your future. We're going to give
4:02 Yeah, you're going to be born inside
4:05 these inside these platforms. Pretty
4:06 amazing, right?
4:07 >> Before that, this video is brought to
4:10 you by Thermaltch and their V600 TGKS.
4:12 The V600 TG comes in multiple different
4:14 color options like their future dusk
4:17 colorway, a classic white or black,
4:19 yellow, and a light blue and red
4:22 combination. Thermaltch's V600TG uses an
4:24 angular glass front design with heavily
4:27 ventilated side intake and bottom fan
4:29 mount options, plus large holes and
4:31 ventilation at the back for more flow
4:33 and pressure customization. The case
4:35 makes for a good showcase PC with colors
4:37 difficult to find elsewhere while also
4:38 providing adequate cooling options
4:40 through the perforations around the
4:42 case. It's also compatible with some
4:44 back connect motherboard form factors.
4:45 Learn more at the link in the
4:47 description below. Nvidia's consumer
4:49 electronic show keynote had zero
4:50 consumer news in it. And actually, it
4:53 had almost zero news in it at all. There
4:55 was a good amount of news. They just
4:57 didn't put it there. And maybe that's
4:59 because as Nvidia CEO Jensen Juan has
5:01 said, the keynotes serve a different purpose.
5:02 purpose.
5:05 >> When when we give a keynote, everybody's
5:06 stock price goes up.
5:08 >> Nvidia's stock didn't move much during
5:09 the keynote, but Jensen did give the
5:12 audience a slathering of corpo slop
5:16 speak for 93 minutes as he moved around
5:19 on the stage to to prove to shareholders
5:20 that he's healthy enough to continue
5:21 being CEO.
5:23 >> You have to be a CEO in really good
5:26 shape to do this job. But it was 93
5:28 minutes for a good reason.
5:30 >> We have about 15 keynotes worth of
5:31 material to pack in here.
5:34 >> Please, please, God, no. Please don't
5:36 don't do that. Nvidia's keynote covered
5:38 a medley of the company's greatest hits
5:41 like AI, self-driving cars, robots, AI,
5:44 autonomous vehicles and machines, AI,
5:47 and of course, money. What that means is
5:50 some $10 trillion or so of the last
5:52 decade of computing is now being
5:54 modernized to this new way of doing
5:56 computing. A hundred trillion dollars of
5:59 industry several percent of which is R&D
6:02 budget is shifting over to artificial intelligence.
6:03 intelligence.
6:05 >> Almost 800 plus billion dollars.
6:06 >> Hundreds of billions of dollars. Well,
6:09 it turns out we have billions of dollars
6:11 of supercomputers in operation. These
6:13 are billions of dollars. Let's say a
6:15 gigawatt data center is $50 billion.
6:16 There's a couple hundred billion dollars
6:17 in VC funding.
6:19 >> Other than money, the keynotes few
6:22 newsworthy details included Vera Rubin
6:24 details on the company's upcoming
6:26 platform that includes, it says, six new
6:28 chips within one, as it calls it, AI
6:30 supercomputer. They also had gaming
6:33 news, but we'll come back to that kind
6:36 of like how they never did. Jensen spoke
6:38 at length about how Nvidia co-designed
6:40 the six new chips, the Reuben GPU, Vera
6:43 CPU, Envyink 6 switch, Connect X9
6:46 Superdick Bluefield 4 DPU, and Spectrum
6:48 6 Ethernet switch and mentioned how
6:50 these would lower AI costs and improve
6:53 performance. Jensen Han even invited his
6:54 robot friends to the stage from last
6:55 year as well.
7:06 Hey guys,
7:08 hurry up.
7:10 >> He then spent the ensuing 20 to 30
7:11 minutes doing the talking to an empty
7:14 chair routine while he weirdly addressed
7:17 the robots for everything instead of the
7:18 audience. I he was still using the
7:21 second person you but he would look over
7:23 there at the robots and the audience
7:24 seemed very confused about what to do.
7:27 But may maybe he's just preparing for
7:28 when he replaces the entire audience
7:31 with AI.
7:32 There no GeForce gamer in the room.
7:34 >> Jensen and the Nvidia pet robots
7:37 unveiled robots on parade with the sure
7:40 shot sure to make investors drop money
7:42 as they showed robots ranging from
7:44 surgical use cases to construction
7:46 robots from Caterpillars. Although they
7:47 didn't show any of the killing drones
7:50 that uh the friends of the military
7:51 industrial complex might use. So that's
7:54 right. Once again at the consumer
7:56 electronics show, Nvidia has covered
7:57 everything except for consumer
7:59 electronics. And this time, even in its
8:02 B2B and enterprise discussion, it didn't
8:04 really talk about much news at all.
8:06 Instead, Nvidia highlighted how Nvidia
8:08 technologies would make AI operate in
8:10 the physical world, which Nvidia called
8:13 physical AI. Nvidia showed actual and
8:14 simulated demonstrations with its robots
8:16 and autonomous vehicles working in a
8:18 physical environment. The company also
8:21 introduced Alpameo, a family of open
8:23 reasoning models for autonomous vehicle
8:24 development. Nvidia's press release
8:26 called it quote part of the sweeping
8:28 push to bring AI into every domain. End
8:30 quote. As is Nvidia tradition, of
8:31 course, it also celebrated its
8:33 partnership with all the B2B and
8:35 enterprise companies out there. For
8:37 example, uh Palunteer Palunteer. Palanteer.
8:38 Palanteer.
8:40 >> Nvidia, could you just give us
8:43 something, please? Anything like like a
8:46 drop of water in a desert. Need anything.
8:46 anything.
8:48 >> And although we're not announcing any
8:49 new GPUs today.
8:50 >> Okay. All right. So, there aren't any
8:52 new GPUs. It's fine. We didn't really
8:54 expect any. It's not like they have the
8:55 memory supply to ship for it right now
8:56 anyway because they've got to suck it
8:58 all into the new systems that they have
9:01 that pull somewhere around two terabytes
9:03 for uh server. So in this video with
9:06 Jacob Freeman formerly of EVGA, he's
9:08 talking about how Nvidia actually did
9:10 have consumer gaming news, just not a
9:12 single piece of it was in the consumer
9:14 electronics show keynote. And actually,
9:17 weirdly, the gaming side had more news
9:19 than the non-gaming side in terms of
9:21 like news meaning stuff that hasn't been
9:23 said before. So, for some reason, none
9:25 of this stuff from the video that
9:27 Jacob's in made it into the keynote at
9:29 the Consumer Electronic Show. Like,
9:30 literally none of it. It almost seems
9:32 like Nvidia thinks talking about gaming
9:35 reduces the seriousness of their company
9:37 to the investors they pander to now. We
9:39 just wonder what Jensen Juan thinks
9:41 about the PC gamers who built his
9:42 company to where it is now.
9:44 >> Like I said, nobody's as cute as you
9:44 guys are.
9:47 >> Around 1 hour into the keynote, Nvidia
9:49 finally shared its first actual news
9:52 about literally anything. This was on
9:54 its upcoming Vera Rubin solution. So
9:55 that's right, at the Consumer
9:57 Electronics Show, if you have your own
9:59 $50 billion data center, you too can be
10:02 a consumer of a Vera Rubin solution.
10:04 Gaze upon the thin siphoning away all of
10:06 the consumer hardware allocation. will
10:07 give them credit for building what looks
10:09 to be a fairly modular design where
10:11 Jensen Juan illustrates on stage how the
10:14 prior solution had he says 43 cables and
10:15 six tubes for cooling while still
10:17 relying on air for some components
10:20 stating that the new Vera Rubin boxes
10:21 instead which is the new architecture
10:23 following Grace Blackwell moved to zero
10:26 cables and two tubes for water in and
10:27 out. Looking at this render of the
10:29 server, Nvidia shows blade style
10:31 connectors for power at the front edge
10:32 of the two primary boards shown at the
10:35 back sliding into the server solution
10:37 with liquid cooling tubes integrating
10:38 more completely with water cooling
10:40 blocks used on top of the CPU and GPU
10:42 parts. Juan also claimed that it's
10:44 faster to deploy, so data centers have
10:46 one less obstacle between them and more
10:48 feckless expansion.
10:52 >> It takes 2 hours to assemble this.
10:55 If you're lucky, it takes two hours. And
10:56 of course, you're probably going to
10:58 assemble it wrong. You're going to have
11:00 to retest it, test it, reassemble it.
11:02 So, the assembly process is incredibly
11:05 complicated and it was understandable as
11:07 one of our first supercomputers that's
11:10 deconstructed in this way. This from 2
11:13 hours to 5 minutes. Janu says that the
11:15 new server solution is 100% liquid
11:17 cooled, up from a stated 80% on the
11:19 prior model. He also claimed, and we're
11:21 not sure how much of this was a joke
11:22 versus wasn't since it actually was
11:24 unclear, that it has half a ton of water
11:26 in a full rack of these Ver Rubin
11:29 servers. The Vera part of Ver Rubin is
11:32 the CPU, which Nvidia says will have 88
11:35 cores branded Olympus, 176 threads, a
11:37 1.8 terabyte per second and VLink
11:39 connection, a 1.5 TB capacity of system
11:41 memory with 1.2 terabytes per second of
11:43 bandwidth on LPDDR5X,
11:46 and will be a 227 billion transistor
11:47 solution for the Vera CPU alone. The
11:50 Reuben component is the GPU part with a
11:53 stated quote up to 288 GB of HPM4 and
11:56 quote per GPU with multiple GPUs per
11:58 configuration possible. Nvidia states a
12:00 bandwidth of up to 22 tabytes per
12:01 second. Nvidia also makes a bunch of
12:03 claims about performance of NVFP4
12:05 inference and training stating 5x and
12:08 3.5x blackwell for each with the 22
12:10 tabby per second HPM4 bandwidth getting
12:12 a major line item here. Currently, the
12:13 memory suppliers have shifted towards
12:15 more HBM manufacturing to keep up with
12:18 the data center GPUs like this. HBM is
12:20 expensive and costs more wafer area when
12:22 factoring in things like yield losses,
12:24 contributing more to the memory crisis.
12:25 Really, it's totally tonedeaf.
12:28 Announcing things with terabytes upon
12:29 terabytes of memory in them for like
12:31 millions of dollars to a bunch of
12:33 consumers who can't afford things and
12:34 can't get anything with memory in it.
12:37 It's kind of like announcing that you're
12:39 the only guy with water in a place that
12:42 has a huge drought. So, we we shipped a
12:49 >> Can you hear it squealing?
12:51 >> Okay, look. But he didn't It's not like
12:53 he in case you were wondering who cares.
12:54 It's the tech billionaires and leaders
12:55 of other trillionaire and soon to be
12:57 trillionaire companies all lining up
12:59 with their handout to Jensen. Open AAI
13:02 CEO and who who just recruited the
13:04 Grinch to steal Christmas in Whoville,
13:05 Sam Alman, had this to say. Quote, "The
13:07 NVIDIA Rubin platform helps us keep
13:09 scaling this progress so advanced
13:11 intelligence benefits everyone." End
13:14 quote. Elon Musk said, "Quote, green
13:17 heart emoji, confetti emoji, rocket ship
13:20 emoji, robot emoji. Nvidia Rubin will be
13:22 a rocket engine for AI. Reuben will
13:24 remind the world that Nvidia is the gold
13:27 standard. Uh, see prior set of same
13:28 emojis." End quote.
13:29 >> I say the stupidest things that cannot
13:30 possibly be true.
13:32 >> So, the executive pickaxe trade
13:34 continues. And once again, it's your
13:35 data that they're mining.
13:36 >> There's so many companies that would
13:38 like to build. They're sitting on gold
13:40 mines. Gold mine. Gold mine. It's a gold
13:42 mine. Gold mine. Gold mine. It does this
13:44 repeatedly, token after token after
13:46 token. And obviously, if you have a long
13:49 conversation with that AI over time,
13:50 that memory, that context memory is
13:52 going to grow tremendously. Not to
13:54 mention the models are growing, the
13:56 number of turns that we're using, the AI
13:58 are are increasing. Even Sai Nadella,
14:00 CEO of Microsoft and guy who is recently
14:02 excited that they have warehouses full
14:04 of GPUs they can't plug into data
14:06 centers because they don't have the grid
14:08 capacity yet was excited for Ver Rubin
14:10 presumably so he can add it to his
14:13 smalike mountain of hoarded silicon
14:15 components. Nvidia also talked about its
14:17 new connect X9 spectrum X super nick
14:20 which is 1x short of requiring a VPN to
14:22 read about. The Nick is advertised as
14:24 running 800 Gbit per second Ethernet.
14:26 The company wrote on its blog quote in
14:28 the Ver Ruben NVL 72 rack scale
14:30 architecture each compute tray contains
14:32 four connectx9 superdick boards
14:34 delivering 1.6 6 terabts per second of
14:36 network bandwidth per Ruben GPU. This
14:38 ensures GPUs can participate fully in
14:40 expert dispatch, collective operations,
14:41 and synchronization without becoming
14:43 bottlenecked at the network edge. End
14:45 quote. Further noting that connectx9 has
14:47 these security capabilities quote data
14:49 and transit encryption acceleration for
14:52 IP security or IPS and platform security
14:55 protocol PSP to secure GPU toGPU
14:57 communications data at rest encryption
14:59 acceleration to secure storage platforms
15:02 secure boot firmware authentication and
15:04 device attestation this is part of
15:06 Nvidia's dubbed spectrum X Ethernet
15:08 scaleout architecture that's the quote
15:10 using bluefield 4 DPUs for handling
15:12 quote networking storage security and
15:14 control services and quote across
15:17 Reuben. Reuben consists of NBLink, the
15:20 CPU, the GPU, Bluefield DPUs for network
15:23 storage security needs, the Supernick
15:25 and the Ethernet switch. The end result
15:28 is what Nvidia refers to as new chips,
15:29 six of them to combine into its
15:31 solution. That's a lot of supporting
15:33 silicon to enable the GPU and the CPU to
15:35 run inference and training tasks. Using
15:37 all of these extra silicon components in
15:39 the server means that the GPU and CPU
15:41 can remain dedicated entirely to
15:43 so-called AI processing workloads. For
15:45 specs, Nvidia published a table
15:47 comparing Vera to Grace for the CPU.
15:50 Nvidia shows a move to 2 megabytes of L2
15:52 cache per core, 162 megabytes unified
15:55 L3, up from 114 megabytes, an increase
15:57 to 1.2 terabytes per second memory
16:00 bandwidth from 512 GB per second. a move
16:03 to 1.5 terabytes of LP DDR5X capacity
16:05 from 480 gigabytes maximally previously
16:08 faster NVLink solutions and a move to
16:11 PCIe Gen 6 and CXL 3.1 from PCIe Gen 5
16:13 previously. This table references what
16:15 they call spatial multi-threading for
16:17 Vera which we weren't familiar with
16:19 before. Nvidia defines this as quote a
16:21 new type of multi-threading that runs
16:22 two hardware threads per core by
16:24 physically partitioning resources
16:26 instead of time slicing enabling a
16:27 runtime trade-off between performance
16:30 and efficiency. This approach increases
16:32 throughput and virtual CPU density while
16:34 maintaining predictable performance and
16:36 strong isolation, a critical requirement
16:38 for multi-tenant AI factories. End
16:40 quote. And just to get ahead of the
16:41 marketing that they keep
16:43 putting in here, AI factories means data
16:46 centers. That is what that is. They're
16:47 trying to brand it as a factory to make
16:50 it seem like some kind of approachable
16:52 bluecollar thing presumably so they can
16:55 go ask the government for more money and
16:58 better regulations or something. uh but
16:59 it is a data center. That's what that
17:01 means. For memory, Nvidia notes that
17:04 it's using SOCAM or small outline
17:06 compression attached memory modules.
17:08 These look something like this pictured
17:10 in a serve the home article from
17:12 previously. This means that the memory
17:14 isn't BGA soldered to the board and is
17:16 instead attached to a stick. It's just
17:19 so cam with pins instead of a DDR5 style
17:22 stick. NVIDIA mounts the LPDDR5X to
17:24 these types of SOCAM sticks for the
17:26 server allowing modularity for capacity
17:28 per configuration while also giving some
17:30 level of replaceability if a chip on a
17:32 stick happens to go bad. Now, we
17:34 frequently see comments posted where
17:36 people ask if there's any potential
17:37 secondhand market in the future for all
17:39 these data center server components
17:41 where I think generally people are
17:42 wondering, okay, once all these things
17:44 get retired in like 3 to 5 years or
17:47 less, does it end up on the consumer
17:49 market where I can just buy sticks of
17:51 DDR5 for pennies on the dollar because
17:52 they're dumping hundreds of terabytes of
17:54 memory onto the market? And the answer
17:56 for things like this is no, because this
17:58 has no use in current desktop type
18:01 standard computers. uh it may have a
18:04 secondhand use for other types of data
18:07 centers uh or startups or something but
18:09 there's effectively zero use for any of
18:12 this in consumer uh there may be a use
18:15 also in say China where in Shenzhen you
18:17 might find a guy at a shop who will
18:19 desolder all the LPDDR5X from abandoned
18:21 modules although they seem worth more so
18:22 modules but and then put it on to
18:24 something else maybe there's a use case
18:26 there uh generally speaking though no
18:27 this is not something that you can just
18:30 stick into a an X87 70 motherboard or
18:32 something. Nvidia's block diagram shows
18:34 the layout of Vera Rubin. The CPU
18:37 connection to LPDDR5X sits at the top
18:39 going out to the SOCM sticks at up to
18:41 1.5 terabytes of capacity. The CPU
18:44 connects to PCIe Gen 6 depicted on both
18:46 sides of the block with NVLink COC at
18:48 1.8 terabytes per second going to the
18:50 two attached Reuben GPUs for the one
18:53 CPU. These are depicted with 288 GB of
18:56 HPM4 each. And we covered this in our
18:59 video entitled Nvidia. What the
19:00 question mark where we talk about how
19:04 HPM actually requires more wafer area
19:06 and allocation than a like for like
19:08 capacity of something like say DDR
19:11 normal DRAMM DDR5 or something uh or
19:14 VRAM if you want to take that comparison
19:16 the reason for this is a combination of
19:18 yield losses where HPM has more yield
19:20 losses because you've got a more complex
19:22 set of things vertically stacked if
19:23 anything goes bad in there you might
19:24 have to throw out the entire chip so
19:27 you're throwing a lot more silicon uh to
19:30 other factors such as requiring separate
19:33 uh IO or control silicon solutions and
19:34 interposer silicon solutions although
19:36 the interposers would come from TSMC not
19:38 from the memory factories fabs uh but we
19:41 talk about that in our separate piece
19:44 Nvidia WTF. So anyway point being at 288
19:48 GB of HBM per GPU listed as a max
19:51 capacity that is more than just 288 GB
19:54 of memory per GPU offset from the
19:56 consumer market. It would not be a
19:58 onetoone loss for consumer. It would be
20:00 greater than that. Nvidia also noted
20:03 that Reuben is a 336 billion transistor
20:05 chip for the full die solution with two
20:07 compute dies connected via fabric
20:09 centrally rather than using a single
20:10 monolithic chip. Just like with
20:12 Blackwell, Nvidia continues to leverage
20:14 ARM for multiple parts of its systems.
20:17 One example being Bluefield 4 running 64
20:19 ARM neoverse v2 compute solutions. The
20:22 Bluefield 4 component alone has a listed
20:24 memory capacity of 128 GB. though even
20:26 beyond the CPU and the GPU, these
20:28 additional processors that are shipping
20:31 with it are consuming huge amounts of
20:32 memory. The previous generation was
20:35 listed at 32 GB. As for the Ethernet
20:38 Spectrum 6 switch, Nvidia notes a 102.4
20:41 terab per second solution at 512 by 200
20:44 Gbit per second ports. Nvidia
20:45 illustrates performance with what it
20:47 calls a quote expert dispatch benchmark
20:49 end quote where completion time is
20:51 measured in milliseconds. The NVLink
20:53 switch tray is its own entire chassis
20:56 depicted here with NVL link six
20:57 switches, spine connectors, and a
21:00 claimed 3.6 terabyte per second per GPU
21:02 all to all solution. Also fully liquid
21:04 cooled and increasing water demands.
21:06 Nvidia spent some time towards the
21:08 bottom talking about power consumption
21:10 in its article. Seeing as data centers
21:11 are currently in the process of a
21:13 ruinous takeover of the grid in the
21:14 United States, this seems like something
21:16 they would want to talk about. Nvidia
21:17 notes that it's attempting to improve
21:19 efficiency, but of course, obviously,
21:21 and they don't say this, it is still
21:22 ultimately just ramping power
21:23 consumption beyond what we actually have
21:25 capacity for. The company showed power
21:27 smoothing and GPU power draw in this
21:29 chart measured in at megawws, which
21:30 really tells you everything you need to
21:32 know. Now, you might be wondering what
21:35 all this means for Nvidia.
21:38 Maybe you're not. I It's But just just
21:40 let let me let me get there. Nvidia has
21:42 published this helpful image showing the
21:44 circle of what all this means. There's
21:46 just one thing wrong with it. It was a
21:47 small error they made. One second. Let
21:49 me just There we go. That's more
21:52 accurate. The circle of profit. With all
21:55 of that out of the way, we got some
21:58 actual news in there. Look, I tried Doug
21:59 threw a lot of stuff they wrote because
22:01 it wasn't in the keynote. I can tell you
22:03 that cuz I was falling asleep watching
22:04 the keynote. But with all that out of
22:07 the way, let's go to the gaming news.
22:09 So, they announced a half-step iteration
22:12 on DLSS with DLSS version 4.5 that
22:14 introduces a new transformer model, an
22:15 updated transformer model to the one
22:18 that we tested last year. And they also
22:20 are introducing a new dynamic multiframe
22:24 generation that now goes up to 6x from
22:26 4x previously. There's some other stuff
22:28 in here too, like some RTX Remix news.
22:30 Nvidia started its presentation by
22:33 providing some yaxis devoid charts that
22:34 report growth in the PC gaming segment.
22:37 The first graph is units sold, which is
22:40 unknown at 2019 to 2024, referencing
22:42 Steam, Gartner, and Nvidia firstparty
22:44 results. The company reports a 14%
22:47 decline in PC adoption over the period
22:50 versus a 51% claimed increase in gaming
22:52 PC adoption. As for what constitutes a
22:55 gaming PC, Nvidia says it's anything
22:57 with a discrete GPU installed in it.
22:58 This would also mean that office
23:00 workstation PCs, work from home PCs
23:02 bought during COVID, and even local AI
23:05 processing desktops with the DGPU would
23:07 be classified as gaming PCs. Since this
23:10 chart crosses the CO explosion in the PC
23:12 market in 2020 through 2022, the growth
23:14 in DGPU equipped machines is
23:16 unsurprising. It's masking whatever
23:18 happened in the last year or so. What it
23:20 doesn't show is 2025, that data is
23:22 probably not fully in yet to be fair,
23:24 but the projections that are out now
23:25 from some of the analyst firms in the
23:28 industry are forecasting a reduction in
23:32 PC purchasing just in general for 2026.
23:35 Uh, which is also not a surprise. Nvidia
23:37 also noted that its percent total of
23:40 Steam install base has increased for
23:42 Blackwell versus prior generations. The
23:44 X-axis is months after launch and the
23:47 Yaxis is who the knows? But you
23:48 could calculate this one manually if you
23:50 cared. We'll start with the DLSS 4.5
23:53 changes. Nvidia's DLSS 4.5 introduces
23:55 two major changes. An updated
23:57 transformer upscaler model and dynamic
24:00 multiframe generation up to 6x now for a
24:02 semi-adaptive target frame rate
24:03 solution. Meaning frame generation would
24:04 be reduced in scenarios where you're
24:06 closer to your target frame rate and it
24:08 would be increased in scenarios where
24:10 you're further from it. The LSS 4.5's
24:12 second generation Transformer model,
24:13 which is the new one for the upscaler,
24:15 is available basically immediately to
24:18 all RTX GPU users. So that goes back to
24:20 the 20 series. It excludes only GTX
24:23 cards, including GTX 16 series cards
24:26 with dynamic MFG 6X reserved for 50
24:28 series users who wish to enslify their
24:30 gaming. Adaptive multiframe generation
24:32 is pretty interesting. They're calling
24:34 it dynamic. Uh, but we talked about Lost
24:36 the Scalings adaptive multiframe
24:38 generation previously. That's the $7
24:40 tool you can buy on Steam. Uh we
24:41 benchmarked some of those solutions and
24:43 did image quality analysis. But what
24:47 they do is a fractional multiplier uh as
24:50 the the minimum sort of divisible amount
24:53 where you can have say a 1.6x frame
24:55 generation or something like that in
24:57 order to hit a target frame rate that is
24:59 defined in the software. In this
25:02 situation, Nvidia can only accept whole
25:05 numbers. They told us that uh 1 through
25:07 6x or just no frame generation at all
25:10 are all possible values, but there's no
25:12 fractional value in between. And Nvidia
25:14 also noted that it can run on either a
25:16 manually set frame rate target, just a
25:18 hard number, or you can set it to
25:20 operate based on a refresh rate for the
25:21 screen target. As for the updated
25:23 transformer model, Nvidia's first party
25:24 demo makes it appear as if some issues
25:26 that we found previously have been
25:29 resolved. In our early 2025 DLSS testing
25:31 for DLSS4 at the time, which was done
25:32 independently of Nvidia's background
25:33 pressure that we've detailed in the
25:35 past, we found that the new DLSS
25:37 transform model sometimes had issues
25:39 with ghosting and trailing edges on
25:41 objects or UI elements. Overall, we did
25:43 think it was an improvement on the CNN
25:45 model, but it still had some areas to
25:47 improve. The new transformer model in
25:50 4.5, according to Nvidia's first-party
25:51 capture they distributed, should resolve
25:53 at least some of the ghosting, trailing
25:56 edge, and UI element detail issues, such
25:59 as text on the screen. We haven't yet
26:01 tested this, but Nvidia's first party
26:02 video demonstrating the feature does
26:04 look to improve upon issues that we
26:06 showed when we tested DLSS4 last year.
26:09 Nvidia's demo with Oblivion Remastered
26:11 helps show the ghosting issue in DLSS4
26:14 versus 4.5 with 4.5 looking a lot better
26:16 from the updated Transformer model.
26:18 Nvidia says that its updates will
26:20 primarily target performance and
26:21 ultraformance modes with super
26:22 resolution for image quality
26:25 improvements. Nvidia also anticipates
26:26 improvement in other DLSS quality
26:28 settings, but says that these will be
26:30 the most affected. There are override
26:32 settings in the NV app utility for these
26:34 features. And beyond these
26:36 announcements, Nvidia's game modding
26:38 utility RTX Remix also received updates.
26:40 This tool has been used to patch old
26:42 games like Halfife 2 and Portal in order
26:45 to maybe ironically do things like
26:47 leverage more VRAM today. It's also been
26:50 used for swapping APIs and modernizing
26:52 the sort of core of the games that the
26:54 modders are working on and generally
26:56 adapting the games forward with mods.
26:57 And we're big fans of modding games
26:59 here. It's actually kind of where I got
27:02 a lot of the start for GN was covering
27:04 game mods for things like Oblivion,
27:07 Skyrim, so forth. And um so RTX Remix
27:09 has always been kind of interesting uh
27:10 despite some of the other issues that
27:11 we've had with Nvidia over the years,
27:13 but this has gotten an update. And the
27:15 updates to Remix include what they're
27:17 calling Remix Logic, which allows for
27:19 game event detection to create triggers.
27:21 Anyone who used map editors back in the
27:23 day will be familiar with simple trigger
27:25 systems. So something like opening a
27:27 door or encountering a specific enemy
27:29 will give modders the opportunity to
27:31 easily insert game events tied to those.
27:33 Nvidia gave the example of a screen
27:35 pulse in Halfife 2 when enemies are
27:37 near. They also showed some simple
27:38 trigger response workflows with the
27:40 update. In theory, this should make it a
27:42 lot easier for people to mod the game
27:44 without a whole lot of programming uh
27:46 knowledge in the background. Nvidia also
27:49 discussed more of its in-game AI agent
27:50 solutions. The only thing we'll point
27:53 out is they're recommending 7 GB of VRAM
27:55 or more for using these features as it
27:57 executes locally. Nvidia's mid-range
27:59 cards only have 8 GB of VRAM in some
28:01 cases. So, they're really sort of just
28:03 making everyone's point uh for them.
28:05 Nvidia also had some GeForce Now
28:07 updates. We covered some of them in our
28:09 recent video about Nvidia. The one we
28:10 covered, they didn't talk about today,
28:12 which is the price going up effectively
28:15 because if you're over 100 hours of play
28:17 time now, you have to pay for more hours
28:19 in 15 hour chunks for their game
28:23 streaming service through the cloud. And
28:25 uh as we talked about previously, that
28:27 just means you're kind of sacrificing
28:28 ownership of the device, but now you're
28:30 also going to be giving away more and
28:32 more money for just playing games. And
28:35 although 100 hours a month sounds like
28:37 maybe a lot of time, one, it's it's
28:40 really not that much. But two, uh if
28:43 you're idling in the games or maybe more
28:45 common if you have friends, family,
28:47 roommates, whatever, who may share uh
28:50 typically a system with you to play on,
28:52 then moving all of that to a single user
28:54 account on GeForce Now would mean that
28:56 you blow through those hours really fast
28:58 if it's more than one user. um which you
29:00 could of course have with just a normal
29:03 computer if you have a a kid or a
29:04 significant other or someone who might
29:06 use the same device. So anyway, they
29:07 didn't talk about that. What they did
29:09 talk about was technically the 5080 tier
29:11 now supports some more games. So that's
29:14 the news on that. Um we think
29:15 cloud-based subscription computer
29:16 services will lead to the death of
29:19 hardware ownership and already covered
29:21 how awful we think GeForce Now is for
29:22 consumers in a recent piece. So let's
29:24 move on to the next part which is beyond
29:26 all this. Nvidia had some G-Sync Pulsar
29:29 news and LLM performance improvement
29:31 claims on its GPUs. Namely, Nvidia
29:33 claims optimizations in Comfy UI to run
29:35 some models with significantly reduced
29:38 memory. One example has the 87 GB
29:39 requirement for a BF16 workload being
29:42 brought down to 26 with NVFP4. We're not
29:44 really experts in LM workloads or
29:47 processing or how linearly this can or
29:49 can't be compared. That's the chart they
29:50 made. So, we can't really comment much
29:52 on these claims beyond just presenting
29:53 them and that's what they are. All
29:54 right, that's it for this one. Thanks
29:56 for watching. As always, that's the
29:58 recap of Nvidia's keynote. I mean, we
29:59 did a lot more than they had in their
30:01 keynote here. Uh, it's the articles
30:02 they've published and their supporting
30:05 materials elsewhere. I look, they don't
30:06 put the gaming stuff or the consumer
30:08 stuff in the consumer electronics show
30:10 keynote. They haven't really focused on
30:11 that for a while. So, this in a sense
30:13 isn't new. It's just now it's like
30:15 particularly bad because they had gaming
30:18 in consumer news and actually more of it
30:22 than what was in the 92 minutes of like
30:24 rambling walking around on the stage
30:27 like he stayed up too late the night
30:30 before Jensen Juan addressing
30:33 a crowd that half the time didn't seem
30:34 to understand if they were supposed to
30:36 clap or if the robots he was talking to
30:37 were supposed to clap. Like that's what
30:40 we got. So
30:43 you're welcome. Uh, don't bother
30:45 watching it. That's what we're here for
30:47 because I can tell you it was incredibly
30:49 painful. They didn't even have chat
30:50 turned on. At least Intel had chat
30:52 turned on so I could jump in and say,
30:56 "Wait, what is that? It does B70. Don't
30:58 do this to us." And if you want to learn
31:00 about that, check out our Intel keynote
31:01 coverage. I don't know if it's up or not
31:03 yet, but it will be soon if it's not.
31:06 Subscribe for more store.gamesaxis.net.
31:09 support us directly and uh the I guess
31:10 torture that we just went through
31:12 watching these keynotes. You can help us
31:14 out if you find this useful or
31:16 patreon.com/gamers nexus. We'll see you
