It's just genuinely super fun to hangout with Wendell and see this kind of stuff! He's borrowing the server and these GPUs for benchmarking on his channel, so check him out to see the results! (FYI - we filmed this way before the AMD announcements, it was just a late upload, so the 7900 XTX wasn't known yet). Via Wendell, this is an AMD MI210 GPU (meaning that the shroud lists the MI200 series, maybe because it's an older shroud, but the GPU core itself is the MI210 -- so the silicon would be the MI210). Subscribe to Level1Techs: ruclips.net/user/Level1Techs (you can also find his benchmarks here!) Find Wendell's tour of our office on his channel: ruclips.net/video/tx43XvMIgis/видео.html Watch our other video with Wendell, featuring a server tear-down (the one that had these cards): ruclips.net/video/GrOsiCZ9MVE/видео.html
Wendell is a unique combination of an easy going, downright great guy who knows VERY, VERY well what he's talking about. Love his channel and his visits to people like yourself and Linus.
13:54 The speaker is labeled SPK1 on the PCB. Beyond the fact it looks like one it is labeled as one. Many components that can overheat and are not directly monitored by IPMI type systems have these sorts of alarms.
Holy crap! I couldn't nail it down for the longest time. But it finally hit me. Wendell looks and sounds almost exactly like the gadget inventor father on the movie "Gremlins". WOW man. That is crazy!
Basically, there are two (or more) materials with different CTEs (coefficient of thermal expansion), in this case the BGA substrate and the PCB. When the board heats up, the BGA and the board want to grow/shrink at different rates and different total amounts. The weakest point is the solder joint between the two substrates, and the stress caused by the different expansion rates can cause the joints to crack (and then the card doesn’t work anymore). This black goop tries to take some of the strain off the joints. All “fixed” Xbox 360s had this goop applied to them as far as I know.
Damn, the amount of SMD components is so much, this reminds me of "an architects dream is an engineer's nightmare". Probs to anyone who can actually solder anything like this.
Just correctly place the components and bake it with hot air. Tin will melt and connects the components together to the PCB. Did it many times at school. Its much easier then traditional soldering, because you don't need physical connection with your hand to the PCB which can result in touching / moving other components. The components on these kind of PCB's are off course done with machines. Machines have a higher precision and less prone to errors than humans, especially if its repetitive.
People do not solder these assemblies. The parts are placed by automated pick and place machines. The paste is put down with a stencil. the reflow is done in an IR/hot air oven with a controlled temperature profile. Success is by process, not chance.
Former SMT Machine Operator here, It was my job to operate the machines that make video cards and many other PCBAs for companies such as Google, Honeywell, Boeing, IBM, NASA, and VIASAT, to name a few. These are made on a automated line. First, non-populated blank boards called printed circuit boards (PCBs) are loaded onto a conveyer. Then, a stencil with the pattern of the solder pads is loaded. A squeegy goes across and spreads solder paste through the stencil onto the PCB. The pasted PCB continues down the line through a series of pick and place machines, that are basically high-speed, high-accuracy cap and chip machine guns that "shoot" the parts onto the solder pads, which the parts "stick" to through surface tension. The printed circuit board assemblies (PCBAs) are populated from the smallest parts first such as 0201 capacitors and resistors, to the largest parts such as chokes, transformers, and BGAs last. Only odd-shaped and large parts such as connectors are placed by hand. After being populated, the PCBAs then go through a oven with a specific temperature profile that gradually heats up the boards and solder paste, flows the solder, and cools it at a controlled rate. From here they move onto either hand-placed parts, seal/epoxy/coatings and finally into a automated tester, followed by being packaged and shipped to the customer. The lead-in time and lead-out times in the oven are so the PCBAs and the parts on them do not experience thermal shock, which could interfere with the function of the device or cause broken solder connections, amongst other issues.
@@cockatoo010 Yea I remember even way back AMD's always been able to put out some absolutely monster compute architectures, with the only thing holding them back being nvidia's dominance with CUDA, and how much longer it took for openCL to become mainstream enough to get serious traction. I guess when you find a niche that works though, you might as well exploit it, and AMD certainly has been doing that with these cards. RDNA on the flipside of it has been killing it for gaming performance/watt compared to nvidia, excited to see what the upcoming GPU release truly holds
We bought 10 Radeon VII back in the day for computational fluid dynamics, where you need a lot of memory and fast memory. From the theoretical 1024GB/s you get maybe 600GB/s in practice, but that is still way faster and cheaper than Nvidia cards at the time. Plus, the Radeon VII can do ~3.5 TFLOPs/s FP64 compute. None of the Nvidia cards can do that, except the ones for >$10k.
My favorite thing about these server cards is the EPS 12V 8-pin connector. Used in servers for many years, capable of delivering 384w per connector. The 40 series would have been fine if Nvidia just used 2 EPS 8-pins instead of the 12-pin.
@@Alex-zi1nb The EPS connector has 4 12V and 4 GND pins versus the 3 12V, 3 GND and 2 sense pins of the 8-pin PEG connector. I'm guessing that when the EPS PSU specification was designed, GPUs barely used more than 150w, but dual socket and higher servers (since EPS is part of the SSI mainboard spec) could pull far more power; and the different pinouts were to prevent the wrong (underpowered) connector from being used. Utilizing 5 12v rails spread two 8-pin EPS connectors (12V1, 12V2, 12V4, 12V5) and the 24-pin ATX connector (12V3), 1008 watts can be drawn through the motherboard (plus CPUs, DRAM, bus-powered PCIe cards, etc) alone.
The teardown we all waited for! That GPU has more decoupling caps behind the DIE than a typical "consumer" GPU on the whole board. What a beast of a card!
11:52 Precision current shunt resistors. With typical 2-wire shunts you're forced to measure not just the shunt, but also the resistance of the tabs and solder joints. They're significant down at the milliOhms level, and throw off accuracy a bit. The "4-wire" ones put all the load current through the large tabs. The small tabs let you measure voltage drop directly across the shunt, without all the other garbage in series with it.
Just wanted to agree and add that current sense resistors are already sensed 4-w through the PCB. This specific resistor has cutouts to help soldering (very low thermal impedance power lines and high thermal impedance sense lines) and to maintain the low inductance. (this is also why it's so wide and flat: very low inductance). I looked it up and this is a WSL3637 by Vishay Dale. 1L0 stands for 1 milliohm.
I believe the "bridges" are current sense resistors and the speaker is a speaker, because it says SP on the silkscreen (and it looks like a little piezo speaker).
I'm really glad to see these collaborations with L1Techs (Wendell) and you guys branching out in to some non-gaming things. Your perspective is always welcome!
Wendell and Steve, possibly my favorite duo on all of the internet! I have a bunch of upgrades to install in a PC this weekend and the GN mod mat with grounding bracelet is a must this time of year for me. Just thinking of walking across the room gets me all charged up with static. 😂
I find wearing the ESD strap around my ankle to be less of a hassle. I tend to knock things over when reaching for things with the ESD strap on my wrist. I once accidentally burned through the strap's wire with a soldering iron while cleaning flux. The wire got caught in one of the coils of the soldering iron stand, and I deservedly received much hazing from my coworkers. 😀
@@JJFX- 😆 exactly. I've been building computers since the 80's I think I used that thing like once. 😆 I guess since I never built a PC on a carpet or plastic floors. 😆😆😆
@@TheSocialGamer I didn't say I used ESD straps to build computers. I worked on radar when in the military and now repair sensitive electronics equipment.
5:18 "Look. There's central Park right there. The Bronx. Rossmans shop right there" "He's screaming at Apple" I laughed so hard 😂👌Shoutout to Louis Rossman, he's a great guy! Edit: Nice video! And an amazing piece of hardware. AMD rules!
Wendel and Steve are probably my favorite duo in tech, always know its going to get really geeky fun episode but be super interesting at the same time.
As a Vega owner, it's interesting to see how incredible it is at actual computer tasks..... And how often I don't even come near to any actual compute tasks heh
13:08 in theory 1638GB/s, but in practice you can get maybe 800-900GB/s max, similar to a 3080 Ti. I'm using these for computational fluid dynamics, an application where you need as much VRAM as possible. On a single server with 8 MI200 GPUs (512GB VRAM) I can pull off simulations with 10 billion grid points, 10x what NASA does on 27000 GPUs. The software makes the difference.
1638GB/s is max bandwith, not counting CRC, signal encoding and such. There's lots of stuff going on behind to ensure signal and data integrity and it reduces how much actuall data can be moved. Manufacturer's like to show us high numbers.
@@ProjectPhysX That I don't know. I only know they (manufacturers) like to advertise big numbers, especially cellular operators, to show how great their hardware/services are and I learned some on how signal synchronization, data transfers and such is maintained between PC hardware. Depending on the method used it can take significant portion of bandwitha and we're not talking about protocols yet. I'm not going further with it. I don't know enough to explain more. Just a little I learned at Uni.
The comment about "the next Stable Diffusion" got me thinking. Would it be feasible to add some sort of AI benchmark to future GPU reviews? With models like SD it's becoming practical for non-professionals to play around with AI on consumer gaming hardware, and it'd be interesting to see comparisons. It'd be neat to have a chart for something like iterations per second, for a standardized test with fixed dimensions, prompt, sampler, and seeds. Maybe the GPU reviews already have too much going on to fit that in though, especially since I'm not sure if there's anybody on the team familiar with that sort of thing.
@@tanmaypanadi1414 It'd definitely be tough. It's not polished, it's changing quickly, and it can be tricky to get working properly. For one thing I don't even know whether things like ROCm support for the 7900 GPUs would be available at the time GN would be doing a review, or if it'd arrive later.
That is a speaker and I know this because it has "SPK1" silkscreened next to it lol The component marked "1L0" may be a ultralow ohmic resistor used for current detection - it is in the right area for it.
That buzzer will scream like hell if GPU overheats, Mi25 have same and it's bloody loud. Also that 4 pins next to it? Fan header, and I believe it's also powered and controlled
True the chances of anything happening are pretty much zero, but with such an expensive card I would take the same precaution. Honestly an ESD strap is pretty much not needed in the consumer side of things, but when you are handling expensive or handling hundreds of electronics a day it becomes a must because statistically it will happen (like in manufacturing for example).
The big parts at the back are shunt resistors, used to measure current, probably by the multiphase controllers. It could sound an alarm if one of the phases faults, goes overcurrent, or overtemp with the buzzer on there.
@@stevewatson6839 You do know RUclips dosent show all comments to all people. it's just the luck of the draw. sometimes it shows up on desktop but not on mobile and vice-versa.
There's been mainstream GPUs with beepers too. They were usually just to remind the user if they forgot to plug the PCIe power in AFAIK. There's probably more functionality for the beeper on server cards though.
They are SMD current shunts. the two smaller pads are where they meassure voltage drop, ie current through the resistor. And the other thing is def a piezo beeper.
Most of us will never have a use case for some of these GPU's but its very interesting to see them. Like I'll never own a Lambo but I do like to see them😎
I haven’t taken apart a graphics card but I used to work at a sewing shop as a repairman and working on a $25,000 embroidery machine is a little stressful. Esd is definitely a must have for the pricy electronics
that looks almost identical to a lot of AMD epyc heatsinks i've seen in the industry, down to the size of the constant tension springs. Thanks for showing this y'all!
I used to wonder who that mysterious man hiding his face behind all those monitors were and how he knew so much. Now a decade later, I still love me some Wendell content, he truly is the GOAT!
There's so many caps on the back of the card that RUclips compression freaked out when the card flipped over 😂 Very good looking card and cooler design
This thing is so unfathomably nuts it's not even funny anymore. This thing will kill compute applications like you wouldn't believe. Awesome content as always GN and also thanks for bringing it by Wendell - awesome guy!
Wendell the god. Great content! On a side note, can you even call these GPUs anymore? I know even the manufacturers are doing it, but these don't have video outputs, and I don't figure they're meant for processing video, at least not in the way we usually think of that. Does the name stick just because the form factor and the general architecture (modularity, fast memory, many small cores...) is the same? Or is there a good reason not to call these something like Compute Units or whatever?
11:58 those are current shunt resistors, they're doing the current monitoring, and are expensive parts. If they're the good Vishay ones they're about $20 apiece, have a resistance of 0.001 ohms to an accuracy of 0.1% and don't drift a whole lot with temperature. A glance and a guess says they're also watching low and high side currents, which is interesting and probably necessary given multiple voltage sources on the card. It's important to know if current from the power port is leaving through the PCIe slot or vice-versa and it's easier to monitor there.
![Felix "Unfair" | [Stray Kids : SKZ-PLAYER]](http://i.ytimg.com/vi/Oswujxm2Ag0/mqdefault.jpg)
It's just genuinely super fun to hangout with Wendell and see this kind of stuff! He's borrowing the server and these GPUs for benchmarking on his channel, so check him out to see the results! (FYI - we filmed this way before the AMD announcements, it was just a late upload, so the 7900 XTX wasn't known yet). Via Wendell, this is an AMD MI210 GPU (meaning that the shroud lists the MI200 series, maybe because it's an older shroud, but the GPU core itself is the MI210 -- so the silicon would be the MI210).
Subscribe to Level1Techs: ruclips.net/user/Level1Techs (you can also find his benchmarks here!)
Find Wendell's tour of our office on his channel: ruclips.net/video/tx43XvMIgis/видео.html
Watch our other video with Wendell, featuring a server tear-down (the one that had these cards): ruclips.net/video/GrOsiCZ9MVE/видео.html
I’m interested in the toilet seats haha
Like these videos. Could you guys had text to explain some of the acronyms and terminology? Sometimes its hard to keep track lol
Wendell is a unique combination of an easy going, downright great guy who knows VERY, VERY well what he's talking about. Love his channel and his visits to people like yourself and Linus.
13:54 The speaker is labeled SPK1 on the PCB. Beyond the fact it looks like one it is labeled as one. Many components that can overheat and are not directly monitored by IPMI type systems have these sorts of alarms.
Holy crap! I couldn't nail it down for the longest time. But it finally hit me. Wendell looks and sounds almost exactly like the gadget inventor father on the movie "Gremlins". WOW man. That is crazy!
I like to think Wendell just shows up to the GN offices unannounced with various hardware saying, "Hey, check this out!"
I now will forever believe this
What a dream :)
Then as he leaves "back to you steve!"
Via server rack, this is a crucial step.
"BRUH! CHU GOT TO SEE DIS!"
Black goop is called “BGA underfill” and it’s for reducing thermal stress on the solder balls. Became a much bigger thing after the Xbox 360 RROD.
Interesting, go on ⛏️
Basically, there are two (or more) materials with different CTEs (coefficient of thermal expansion), in this case the BGA substrate and the PCB. When the board heats up, the BGA and the board want to grow/shrink at different rates and different total amounts. The weakest point is the solder joint between the two substrates, and the stress caused by the different expansion rates can cause the joints to crack (and then the card doesn’t work anymore). This black goop tries to take some of the strain off the joints. All “fixed” Xbox 360s had this goop applied to them as far as I know.
Thanks for the info!
Someone should have told Sony about it and use it on their fat PS3s
@@ragingraven7915 Shhhh! Don't enable those folks! Microsoft do something half-right and you want to spoil it? 🙂
Wendell is such a great guy! I always love seeing you two doing collaborations. Have a great day! o7
I always learn so much from Wendell when he visits!
@@GamersNexus and some of it's probably even about computers!
Steve, Wendell, Ian. and Gordon are like the 4 horsemen of the techocalypse. LOL! I love all of you guys.
@@shiftreport3229 HaHaHa! I love it! Take note, Steve! We need t-shirts for this now!
Used to like Wendell but then he ruined my image of him talking about politics on his channel.
Wendell is such a legend! Had a really good catch up with him last week in Vegas.
Wow, this is a piece of art.
@Bully Maguire 🅥 how ironic.
But can it run Crysis?
@@therealnoodledog6660 can the AI on it play Crysis
@@marvintpandroid2213 and would the AI run Crysis on max settings?
I love Wendell too, and he is a piece of art indeed, but can we not call him a "this"? 🤣
There's something about the videos with Steve and Wendell that I enjoy. They always seem to be having a good time, and messing with cool hardware.
Damn, the amount of SMD components is so much, this reminds me of "an architects dream is an engineer's nightmare". Probs to anyone who can actually solder anything like this.
Any decent PCBA can do this.
Just correctly place the components and bake it with hot air. Tin will melt and connects the components together to the PCB. Did it many times at school. Its much easier then traditional soldering, because you don't need physical connection with your hand to the PCB which can result in touching / moving other components. The components on these kind of PCB's are off course done with machines. Machines have a higher precision and less prone to errors than humans, especially if its repetitive.
People do not solder these assemblies. The parts are placed by automated pick and place machines. The paste is put down with a stencil. the reflow is done in an IR/hot air oven with a controlled temperature profile. Success is by process, not chance.
SMD components 😏
Former SMT Machine Operator here, It was my job to operate the machines that make video cards and many other PCBAs for companies such as Google, Honeywell, Boeing, IBM, NASA, and VIASAT, to name a few. These are made on a automated line. First, non-populated blank boards called printed circuit boards (PCBs) are loaded onto a conveyer. Then, a stencil with the pattern of the solder pads is loaded. A squeegy goes across and spreads solder paste through the stencil onto the PCB. The pasted PCB continues down the line through a series of pick and place machines, that are basically high-speed, high-accuracy cap and chip machine guns that "shoot" the parts onto the solder pads, which the parts "stick" to through surface tension. The printed circuit board assemblies (PCBAs) are populated from the smallest parts first such as 0201 capacitors and resistors, to the largest parts such as chokes, transformers, and BGAs last. Only odd-shaped and large parts such as connectors are placed by hand. After being populated, the PCBAs then go through a oven with a specific temperature profile that gradually heats up the boards and solder paste, flows the solder, and cools it at a controlled rate. From here they move onto either hand-placed parts, seal/epoxy/coatings and finally into a automated tester, followed by being packaged and shipped to the customer. The lead-in time and lead-out times in the oven are so the PCBAs and the parts on them do not experience thermal shock, which could interfere with the function of the device or cause broken solder connections, amongst other issues.
14:31 the great irony in the Radeon VII is that it would live on to be one of the best mining cards
GCN is insanely good at computing. Not so much at gaming, which is the reason why AMD based CDNA on GCN and made RDNA pretty much from scratch
Wasn't it also easily killed by it?
@@cockatoo010 Yea I remember even way back AMD's always been able to put out some absolutely monster compute architectures, with the only thing holding them back being nvidia's dominance with CUDA, and how much longer it took for openCL to become mainstream enough to get serious traction. I guess when you find a niche that works though, you might as well exploit it, and AMD certainly has been doing that with these cards. RDNA on the flipside of it has been killing it for gaming performance/watt compared to nvidia, excited to see what the upcoming GPU release truly holds
Typing this comment using a Radeon VII :)
We bought 10 Radeon VII back in the day for computational fluid dynamics, where you need a lot of memory and fast memory. From the theoretical 1024GB/s you get maybe 600GB/s in practice, but that is still way faster and cheaper than Nvidia cards at the time. Plus, the Radeon VII can do ~3.5 TFLOPs/s FP64 compute. None of the Nvidia cards can do that, except the ones for >$10k.
My favorite thing about these server cards is the EPS 12V 8-pin connector. Used in servers for many years, capable of delivering 384w per connector. The 40 series would have been fine if Nvidia just used 2 EPS 8-pins instead of the 12-pin.
if a single 8 pin can deliver 384 watts why tf have companies been adding 2 and 3 (and now 4090 issues)?!?!
@@Alex-zi1nb The EPS connector has 4 12V and 4 GND pins versus the 3 12V, 3 GND and 2 sense pins of the 8-pin PEG connector. I'm guessing that when the EPS PSU specification was designed, GPUs barely used more than 150w, but dual socket and higher servers (since EPS is part of the SSI mainboard spec) could pull far more power; and the different pinouts were to prevent the wrong (underpowered) connector from being used.
Utilizing 5 12v rails spread two 8-pin EPS connectors (12V1, 12V2, 12V4, 12V5) and the 24-pin ATX connector (12V3), 1008 watts can be drawn through the motherboard (plus CPUs, DRAM, bus-powered PCIe cards, etc) alone.
@@JMccovery How many layers of mobo you would have to dedicate to power planes entirely to be able to pass 1kW?
@@Mr.Leeroy Could probably be done on 12 layers, but would more than likely have 16+ layers, like the EVGA SR-3 Dark and Asus Dominus Extreme.
I cannot overstate how valuable of a resource Wendell's channel has been to me. You guys have the best content, love the collab
The teardown we all waited for! That GPU has more decoupling caps behind the DIE than a typical "consumer" GPU on the whole board. What a beast of a card!
You guys are my primary source of tech info (GN+LVL1). Great to see the 2 of you dissecting that compute monster.
Thank you Steve, Wendell, and the GN crew. I have had an extremely rough day and this really helped turn it around! You all are awesome!
I really wish we could get more clean-looking cards like that in the consumer market.
That won't ever happen because everyone knows RGB adds 300% fps /s
@@wpyoga The more RGB in ur setup = the more pro
Right.. A nice slim black matte block maybe
@@miguelpereira9859 rgb is useless as my shit
But.. Dragons, and eagle wings and flames and tribal tattoo looking things are what the children want.
I love the Wendell episodes because I love watching Steve have so much fun learning from and listening to Wendell.
Significant more to cook in consumer cards, but look just as nice. Just because this is small, you’re saying that.
Wendell is a highly demanded cameo celebrity now. Good for him!
Congrats, you got the most convincing bot comment I've ever seen.
Really impressive piece of tech and engineering. Thanks for showin!
Also, really love your collaborations.
Yay! More Wendell colabs. You two together present really well.
I love the collab videos with Wendell, brings out the best of you Steve. Take care
11:52 Precision current shunt resistors. With typical 2-wire shunts you're forced to measure not just the shunt, but also the resistance of the tabs and solder joints. They're significant down at the milliOhms level, and throw off accuracy a bit. The "4-wire" ones put all the load current through the large tabs. The small tabs let you measure voltage drop directly across the shunt, without all the other garbage in series with it.
Just wanted to agree and add that current sense resistors are already sensed 4-w through the PCB. This specific resistor has cutouts to help soldering (very low thermal impedance power lines and high thermal impedance sense lines) and to maintain the low inductance. (this is also why it's so wide and flat: very low inductance).
I looked it up and this is a WSL3637 by Vishay Dale. 1L0 stands for 1 milliohm.
Really loving every Wendell collaboration. He is like Anthony from LTT, extremely knowledgeable and super kind!
Yeah they both have awesome personalities!
may be the other way around?
Because Wendell's been around a sh1t lot longer..
I believe the "bridges" are current sense resistors and the speaker is a speaker, because it says SP on the silkscreen (and it looks like a little piezo speaker).
Looks just like the beap speakers on my z390 and z590 Dark mobos.
I'm really glad to see these collaborations with L1Techs (Wendell) and you guys branching out in to some non-gaming things. Your perspective is always welcome!
I love videos with Wendell! Wild to see the bleeding edge tech to come out
Wendell and Steve, possibly my favorite duo on all of the internet! I have a bunch of upgrades to install in a PC this weekend and the GN mod mat with grounding bracelet is a must this time of year for me. Just thinking of walking across the room gets me all charged up with static. 😂
I find wearing the ESD strap around my ankle to be less of a hassle. I tend to knock things over when reaching for things with the ESD strap on my wrist. I once accidentally burned through the strap's wire with a soldering iron while cleaning flux. The wire got caught in one of the coils of the soldering iron stand, and I deservedly received much hazing from my coworkers. 😀
As someone with a likewise problem, tah for the tip!
Where do you work that's making you use ESD straps?
@@JJFX- 😆 exactly. I've been building computers since the 80's I think I used that thing like once. 😆 I guess since I never built a PC on a carpet or plastic floors. 😆😆😆
You can get esd straps that have alligator clips on both ends, one to your ground and one to a lucky nipple 😉
@@TheSocialGamer I didn't say I used ESD straps to build computers. I worked on radar when in the military and now repair sensitive electronics equipment.
Gamers : Haha a $25K card is totally insane
Jensen Huang : Hold my leather jacket
Don't give the plank ideas. 🙂
Seriously. The Instinct cards are *bargains* compared to the Nvidia GPGPUs.
Gamers : Haha a $25K card is totally insane
Jensen Huang : Hold my 18 billion dollar wallet
5:18 "Look. There's central Park right there. The Bronx. Rossmans shop right there"
"He's screaming at Apple"
I laughed so hard 😂👌Shoutout to Louis Rossman, he's a great guy!
Edit: Nice video! And an amazing piece of hardware. AMD rules!
13:59
Wendell: "I've never seen a GPU with a beeper."
Maybe the GPU is a drug dealer.
Wendell: Why can't I have a 4 kilobit-wide memory bus on my video card?
Me: Hey Wendell, wanna see my old R9 Fury? 🤣
Wendel and Steve are probably my favorite duo in tech, always know its going to get really geeky fun episode but be super interesting at the same time.
As a Vega owner, it's interesting to see how incredible it is at actual computer tasks..... And how often I don't even come near to any actual compute tasks heh
Wendell should be there in as many videos as possible. He is a brainiac.
I love this kind of content. I'd never use such hardware myself but seeing how it's engineered is always a good time
Same here! This stuff is like the big brother of more average consumer tech. It's fun seeing what other markets get to play with. :D
13:08 in theory 1638GB/s, but in practice you can get maybe 800-900GB/s max, similar to a 3080 Ti.
I'm using these for computational fluid dynamics, an application where you need as much VRAM as possible. On a single server with 8 MI200 GPUs (512GB VRAM) I can pull off simulations with 10 billion grid points, 10x what NASA does on 27000 GPUs. The software makes the difference.
Houdini?
@@qyoinqyuri no, my own software, FluidX3D
1638GB/s is max bandwith, not counting CRC, signal encoding and such. There's lots of stuff going on behind to ensure signal and data integrity and it reduces how much actuall data can be moved. Manufacturer's like to show us high numbers.
@@JorgeForge yet Nvidia GPUs deliver 100% of the advertised bandwidth with coalesced access, full 1555 GB/s on the A100 40GB.
@@ProjectPhysX That I don't know. I only know they (manufacturers) like to advertise big numbers, especially cellular operators, to show how great their hardware/services are and I learned some on how signal synchronization, data transfers and such is maintained between PC hardware. Depending on the method used it can take significant portion of bandwitha and we're not talking about protocols yet. I'm not going further with it. I don't know enough to explain more. Just a little I learned at Uni.
WENDELL!!!!!! *back to you, Steve*
You and Wendell are great together
The comment about "the next Stable Diffusion" got me thinking. Would it be feasible to add some sort of AI benchmark to future GPU reviews? With models like SD it's becoming practical for non-professionals to play around with AI on consumer gaming hardware, and it'd be interesting to see comparisons. It'd be neat to have a chart for something like iterations per second, for a standardized test with fixed dimensions, prompt, sampler, and seeds.
Maybe the GPU reviews already have too much going on to fit that in though, especially since I'm not sure if there's anybody on the team familiar with that sort of thing.
That could be tough . but if someone made something like Puget bench that would be awesome.
@@tanmaypanadi1414 It'd definitely be tough. It's not polished, it's changing quickly, and it can be tricky to get working properly. For one thing I don't even know whether things like ROCm support for the 7900 GPUs would be available at the time GN would be doing a review, or if it'd arrive later.
yay Wendell everybody loves the Wends colabs and that thing is insane very interesting to see
That is a speaker and I know this because it has "SPK1" silkscreened next to it lol
The component marked "1L0" may be a ultralow ohmic resistor used for current detection - it is in the right area for it.
The R VII (which was pretty awesome for not gaming) was just an Instinct card anyway, this is its family.
amazing collab!
Perfect example of why the modmat needs magnetic foil under the fabric in some of the useful areas.
This is what I would imagine Batman would be like “look there disassembling my one of my computer parts”.
This! Nowt like a co-lab with Wendell to restore the jaded Steve and Audience!
We need an MI250X teardown. That’s the one with two gpu chiplets and eight stacks (128GB) of HBM2e
Wendell is great, I am so glad he kept doing RUclips after the whole TekSyndicate fiasco
Man you get to tear apart the coolest stuff.
Tear apart a quantum computer.
Ha when the esd strap bumped into all the tracked screws and they cut ahead... Pretty funny.
That buzzer will scream like hell if GPU overheats, Mi25 have same and it's bloody loud.
Also that 4 pins next to it? Fan header, and I believe it's also powered and controlled
Holy teraflops, batman!
That's a lesser dynamic duo.
Amount of bypass caps. Wow. It is more than even big FPGAs.
Wave soldering is for through hole components only. Surface mount components are reflowed in a reflow oven.
Omg hope we see some benchmarks
I actually remember a few GPUs that have beepers. I think they were Nvidia 8000 Series cards, when you didn't plug in the PCIe power they would beep.
Love the ESD strap. you can tell how serious and passionate he is just by him useing it
True the chances of anything happening are pretty much zero, but with such an expensive card I would take the same precaution. Honestly an ESD strap is pretty much not needed in the consumer side of things, but when you are handling expensive or handling hundreds of electronics a day it becomes a must because statistically it will happen (like in manufacturing for example).
Finally! Now I know how to put new thermal pads on my Crossfire MI210s so I can drop my temps in Blender by 1 degree! Thanks, Steve!!!
If Steve is like a mega-geek, this guy Wendell is a giga-geek. Fun to watch!
The moment he mentiond Louis Rossman, i knew i had to watch this. I don't even know how any of this works
Maybe you can send good quality PCB pics to AHO and Buildzoid will do PCB Breakdown.
Wendell, thanks for providing the equipment and visiting GN.
First time I've ever seen someone use an anti-static wristband
If there's two people I would like to spend a few hours in a room with my pc. It's theses two. Their passion just radiates from them
"bridges" are current shunt resistors, 4 terminal for kelvin connection. Probably a few milliohm at most.
"Capacitor City" is actually the most accurate way of putting it, damn
the speaker is labeled spk1, so not a humidity sensor
The big parts at the back are shunt resistors, used to measure current, probably by the multiphase controllers. It could sound an alarm if one of the phases faults, goes overcurrent, or overtemp with the buzzer on there.
Riiiiigghhtt. Thanks, but you are the umpteenth person to mention this. The comment even leads those about Crysis!
@@stevewatson6839 You do know RUclips dosent show all comments to all people. it's just the luck of the draw. sometimes it shows up on desktop but not on mobile and vice-versa.
There's been mainstream GPUs with beepers too. They were usually just to remind the user if they forgot to plug the PCIe power in AFAIK. There's probably more functionality for the beeper on server cards though.
One thing for sure about these cards is the cables don't melt
Dad and son sharing war stories of computers back their day 🥰👍
Buildzoid needs to make a PCB breakdown of this one
But can he put it on liquid nitrogen? ;-)
They are SMD current shunts. the two smaller pads are where they meassure voltage drop, ie current through the resistor. And the other thing is def a piezo beeper.
11:58 Those are shunt resistors. Really nice shunt resistors. Used to measure the current on the 12V connector.
Most of us will never have a use case for some of these GPU's but its very interesting to see them. Like I'll never own a Lambo but I do like to see them😎
Well, because these gpus are absolutely not targeted at the average consumer.
I see Wendell colab I click
Nothing more, nothing less
I haven’t taken apart a graphics card but I used to work at a sewing shop as a repairman and working on a $25,000 embroidery machine is a little stressful. Esd is definitely a must have for the pricy electronics
that looks almost identical to a lot of AMD epyc heatsinks i've seen in the industry, down to the size of the constant tension springs. Thanks for showing this y'all!
The metal ring around the CPU is to strengthen the substrate to facilitate direct-die heatsink contact.
@10:19 what a beautiful card. It has a Feng Shui and Karesansui(Zen Garden) look to it, with the placement of all the parts.
Idk if I’m alone but this video was not pushed to my subscriptions feed
Holy fuck, that is an incredible card
I used to wonder who that mysterious man hiding his face behind all those monitors were and how he knew so much. Now a decade later, I still love me some Wendell content, he truly is the GOAT!
I had a pair of these come in recently, and I was shocked to see that the shipping box clearly indicated what was inside
I bet you are in a constant state of shock.
@@ShainAndrews just surprised to see a box with a label that basically says "I'm worth $50,000 if you steal me"
Yeah because is common knowledge to everyone the price of those super cards, if someone steal one would probably sell it for less than a grand lmao
@@alelokox88 I can find out what it is and its value in thirty seconds from that info. Don't be that berk.
@@jshanks1001 Depends upon how it's shipped. There are couriers that deal with high value items.
There's so many caps on the back of the card that RUclips compression freaked out when the card flipped over 😂
Very good looking card and cooler design
@11:59 shunt resistors to sense voltage drop, then you can calculate current flowing through and also Watts.
This is the entry level one
One amazing GPU!!!
I'm having troubles to follow up the bandwidth of this whole memory stuff. How many full HD movies per second would that be?
Awesome collab. When is your new server video coming out?
There is Louis Rossman shop right there wait what's that he is screaming at Apple 🤣
This thing is so unfathomably nuts it's not even funny anymore. This thing will kill compute applications like you wouldn't believe. Awesome content as always GN and also thanks for bringing it by Wendell - awesome guy!
I would hope so for $25,000. That's in car loan territory.
I can't wait to see the teardown of the AMD Instinct MI300 that was just announced!
Tow of my favorite people on the same set! Wendell and Louis Rossmann!
"Collects Toilet Seats", that alone raised my interest level at least 65% (Very interesting hobby!)
Wendall, we miss you on the LvL1 show. Good to see you getting to travel and have fun with fellow enthusiasts. Take care of him Steve and GN staff.
Wendell the god. Great content!
On a side note, can you even call these GPUs anymore? I know even the manufacturers are doing it, but these don't have video outputs, and I don't figure they're meant for processing video, at least not in the way we usually think of that. Does the name stick just because the form factor and the general architecture (modularity, fast memory, many small cores...) is the same? Or is there a good reason not to call these something like Compute Units or whatever?
Godly Processing Units.
"Compute Units" is already taken. H(eadless)GPU might be a better way to put it.
Loved the shoutout to Louis Rossman lol
Always nice to see Tech Dad visiting!
14:37 the recording cameras lens depth of field makes Wendell look 10” tall and Steve looks like a lil child 👦 😅😂🤣
glad to see you Anti-Static Wrist Straps i never see them on you tube anymore
11:58 those are current shunt resistors, they're doing the current monitoring, and are expensive parts. If they're the good Vishay ones they're about $20 apiece, have a resistance of 0.001 ohms to an accuracy of 0.1% and don't drift a whole lot with temperature. A glance and a guess says they're also watching low and high side currents, which is interesting and probably necessary given multiple voltage sources on the card. It's important to know if current from the power port is leaving through the PCIe slot or vice-versa and it's easier to monitor there.