AMD's 25,000 GPU: Instinct MI210 Tear-Down ft. Level1Techs
video description
Date: 2022-11-10
Related videos
Comments and reviews: 15
anon763
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.
reply
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.
reply
Teth47
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.
reply
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.
reply
Pardalis
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?
reply
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?
reply
gamersnexus
sitrep, we need to get the filming done. here's a wristwrap for each of you. If you see Wendell, try to catch him and bring him back here. Be careful, though, don't let him engage you in shoptalk, that could be the end of your day. call for help and keep track of him until backup arrives. we laid a trail of ECC-ram to try to trap him in the server room, but I'm not sure if he'll fall for that again.
Good hunting and good luck, we're gonna need it.
reply
sitrep, we need to get the filming done. here's a wristwrap for each of you. If you see Wendell, try to catch him and bring him back here. Be careful, though, don't let him engage you in shoptalk, that could be the end of your day. call for help and keep track of him until backup arrives. we laid a trail of ECC-ram to try to trap him in the server room, but I'm not sure if he'll fall for that again.
Good hunting and good luck, we're gonna need it.
reply
Jaden
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.
reply
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.
reply
soi
11:52 those aren't bridges, they are current sensing resistors with very low resistance due to the high current being measured. If you look, there are traces from the pads going to 2 small resistors centered right next to it, classic current sensing layout, it'll go from there to a differential amplifier and/or ADC which is most likely going to be right under it on the opposite side of the PCB.
reply
11:52 those aren't bridges, they are current sensing resistors with very low resistance due to the high current being measured. If you look, there are traces from the pads going to 2 small resistors centered right next to it, classic current sensing layout, it'll go from there to a differential amplifier and/or ADC which is most likely going to be right under it on the opposite side of the PCB.
reply
ProjectPhysX
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.
reply
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.
reply
LeNOYO
We need a consumer version of this card, half the memory if fine.
Vega was a beast and is still good at computing, even gaming.
If aldebaran is made from vega's blueprints, then it must the king of computing.
Even if hbm is hard to integrate on a substrate, thoses power saving may allow more power budget for the gpu.
I hope hbm will make a comeback on consumer cards.
reply
We need a consumer version of this card, half the memory if fine.
Vega was a beast and is still good at computing, even gaming.
If aldebaran is made from vega's blueprints, then it must the king of computing.
Even if hbm is hard to integrate on a substrate, thoses power saving may allow more power budget for the gpu.
I hope hbm will make a comeback on consumer cards.
reply
Robert
Other wonderful facts about the 1920-vintage design. The FETs are Hot Switched to reduce reliability and maximize switching losses. And when any 1 or the FETs shorts, the GPU voltage goes from 1 V to 12 V, frying the GPU. Just think of the clock speed (until the GPU melts).
reply
Other wonderful facts about the 1920-vintage design. The FETs are Hot Switched to reduce reliability and maximize switching losses. And when any 1 or the FETs shorts, the GPU voltage goes from 1 V to 12 V, frying the GPU. Just think of the clock speed (until the GPU melts).
reply
Michael
Why is there so many pins on the sli finger? Its probably not sli lol idk what to call it.. but when i think about the 2080ti and 3080 ect its less than half of the pins and thats sharing memory and lots of stuff so my question is what is that connecter for.
reply
Why is there so many pins on the sli finger? Its probably not sli lol idk what to call it.. but when i think about the 2080ti and 3080 ect its less than half of the pins and thats sharing memory and lots of stuff so my question is what is that connecter for.
reply
DrJay
get louis rossman to also join you two for one of these and i think it might cause a tear in something, some kinda continuum. Im not sure if the hardware would rejoice that it is in good hands or be terrified that it will 100% be taken apart lol
reply
get louis rossman to also join you two for one of these and i think it might cause a tear in something, some kinda continuum. Im not sure if the hardware would rejoice that it is in good hands or be terrified that it will 100% be taken apart lol
reply
evolution
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.
reply
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.
reply
Speelydan
Get some high-res images of this thing to Buildzoid to have him do a PCB analysis. I'm 100% not a professional user at all - heck, I'm barely a PC gamer at this point - but I'd be suuuuper interested in Buildzoid's view on this thing.
reply
Get some high-res images of this thing to Buildzoid to have him do a PCB analysis. I'm 100% not a professional user at all - heck, I'm barely a PC gamer at this point - but I'd be suuuuper interested in Buildzoid's view on this thing.
reply
Ayylien
I liked the HBM architecture, it just shows how AMD is ahead of things but flops on the consumer side of things. I remember when bulldozer came out and everyone was making fun of more cores. Well look at modern processors now.
reply
I liked the HBM architecture, it just shows how AMD is ahead of things but flops on the consumer side of things. I remember when bulldozer came out and everyone was making fun of more cores. Well look at modern processors now.
reply
Digital_Jedi
These things and the A/H100 cards are so insanely fast its not even funny. You're talking about petaflops of performance in a single node potentially, meaning each node would outperform a lot of late 90s supercomputers.
reply
These things and the A/H100 cards are so insanely fast its not even funny. You're talking about petaflops of performance in a single node potentially, meaning each node would outperform a lot of late 90s supercomputers.
reply
Add a review, comment
Other channel videos
