Friday, May 6th 2022

NVIDIA H100 SXM Hopper GPU Pictured Up Close

ServeTheHome, a tech media outlet focused on everything server/enterprise, posted an exclusive set of photos of NVIDIA's latest H100 "Hopper" accelerator. Being the fastest GPU NVIDIA ever created, H100 is made on TSMC's 4 nm manufacturing process and features over 80 billion transistors on an 814 mm² CoWoS package designed by TSMC. Complementing the massive die, we have 80 GB of HBM3 memory that sits close to the die. Pictured below, we have an SXM5 H100 module packed with VRM and power regulation. Given that the rated TDP for this GPU is 700 Watts, power regulation is a serious concern and NVIDIA managed to keep it in check.

On the back of the card, we see one short and one longer mezzanine connector that acts as a power delivery connector, different from the previous A100 GPU layout. This board model is labeled PG520 and is very close to the official renders that NVIDIA supplied us with on launch day.
Source: ServeTheHome
Add your own comment

37 Comments on NVIDIA H100 SXM Hopper GPU Pictured Up Close

#1
Tigger
I'm the only one
80GB of memory, yikes
Posted on Reply
#2
X71200
Yeah well, this is targeted towards HPC and they're using HBM3, so there's that...
Posted on Reply
#3
Fluffmeister
The previous gen A100 also had 80GB variants too, so it isn't unprecedented.
Posted on Reply
#4
CallandorWoT
FluffmeisterThe previous gen A100 also had 80GB variants too, so it isn't unprecedented.
these things are only built for like supercomputers and large servers correct? or am I remembering wrong?
Posted on Reply
#5
BiggieShady
This thing looks like it's for slotting into cloud gaming GeForce NOW rack mounts
Posted on Reply
#6
ARF
TSMC N4 node is probably just a marketing name for TSMC N5+ process?
Posted on Reply
#7
ArcanisGK507
I still don't understand how this device is connected... I don't even see the pins for the electrical source...
Posted on Reply
#8
ARF
ArcanisGK507I still don't understand how this device is connected... I don't even see the pins for the electrical source...
On the back of the card, we see one short and one longer mezzanine connector that acts as a power delivery connector
The quality of the photos is quite low, there is focus loss all over the image - blur...
Posted on Reply
#9
Lenne
S.T.A.R.S.
Is Hopper like Volta, that it's aimed completely for professional use?
ArcanisGK507I still don't understand how this device is connected... I don't even see the pins for the electrical source...
Maybe it's wireless? :laugh:
Posted on Reply
#10
X71200
These are snatched in like MXM, you can see traces (banded parts as well) alongside of screw holes, they don't plug in a computer case of the gaming form... so yeah.
Posted on Reply
#11
KainXS
ArcanisGK507I still don't understand how this device is connected... I don't even see the pins for the electrical source...
They're covered with the molex caps
Posted on Reply
#12
Lenne
S.T.A.R.S.
X71200These are snatched in like MXM, you can see traces (banded parts as well) alongside of screw holes, they don't plug in a computer case of the gaming form... so yeah.
Yeah, those black covers just are over them.
Posted on Reply
#14
Frank_100
That's not a video card.

Video cards are huge, have at least three fans and lots of LEDs.

Like the RTX 4090.
Posted on Reply
#15
ARF
Frank_100That's not a video card.
It is not, it is a calculations accelerator. The graphics cards are also calculations accelerators but they have video signal outputs, so you can see the results of the calculations directly visually on your screen.
Frank_100Video cards are huge, have at least three fans and lots of LEDs.
Not necessarily. The PCB with HBM memory can be very small and if the TDP is in check, then there would be no need for large aluminum heatsinks and fans.

Look at the Radeon R9 Nano:


AMD Radeon R9 Nano CrossFire Review | TechPowerUp
Posted on Reply
#16
Lenne
S.T.A.R.S.
Frank_100That's not a video card.

Video cards are huge, have at least three fans and lots of LEDs.

Like the RTX 4090.
We practically don't know anything about RTX 4090 yet. But yeah, it's a GPU, not a video card indeed.
Posted on Reply
#17
ARF
LenneWe practically don't know anything about RTX 4090 yet. But yeah, it's a GPU, not a video card indeed.
Well, if you follow the leaks, there are some specs:


NVIDIA GeForce RTX 4090 with AD102 GPU rumored to ship with 100 TFLOPS of FP32 compute power - VideoCardz.com


NVIDIA AD102 'Ada Lovelace' & AMD Navi 31 'RDNA 3' GPUs To Be Power Chugging Beasts, Over 400W Rumored (wccftech.com)



AMD RDNA3 GPUs get new toned down specifications, Navi 31 with up to 12288 Stream Processors - VideoCardz.com
Posted on Reply
#18
InVasMani
That's a lot of Tabasco sauce. It's amazing how densely packed this it front and back. Gotta hand it to the engineers on this one.
Posted on Reply
#19
80251
@700 Watts TDP it doesn't need any cooling? Is it supposed to operate in a container of LN2?
Posted on Reply
#20
DeathtoGnomes
80251@700 Watts TDP it doesn't need any cooling? Is it supposed to operate in a container of LN2?
the cards looks smaller than the tanks used for ln2.
Posted on Reply
#21
Dr. Dro
80251@700 Watts TDP it doesn't need any cooling? Is it supposed to operate in a container of LN2?
These are intended to install on an SXM socket and their intended usage is for HPC/data crunching :)
ARFWell, if you follow the leaks
They contradict themselves all the time... I wouldn't take anything posted on wccf or videocardz in a pre-launch season seriously or even consider it anything beyond daydreaming, these sites take a nosedive in quality, they'll post any rubbish sent to them on DMs or Discord for clicks :nutkick:
LenneIs Hopper like Volta, that it's aimed completely for professional use?
Yeah, though GV100 still had a 3D engine and was released to consumer channel even if through a meme of a limited release $3000 SKU.

This is much closer to GA100 or AMD's Arcturus and Aldebaran in nature as it's purely a compute card with no graphics output capabilities :oops:

Ada and Hopper will split NVIDIA's microarchitectures into two different ones geared at gaming and compute the same way RDNA and CDNA split, RDNA was an all new GPU meant to render graphics and little more, while CDNA took root in Vega (which was just as much of an insane accelerator as it was a bad gaming GPU), and developed from there.
Posted on Reply
#22
Mussels
Freshwater Moderator
Only 80GB?

This must be for the common folk
Posted on Reply
#23
InVasMani
The more you buy the more you save. Will that be cash or credit?
Posted on Reply
#24
PapaTaipei
Can nvidia release a card for broad consumers without rt cores? Would save up space for rasterization. Maybe a card marketed for competitive gaming.
Posted on Reply
#25
mashie
ArcanisGK507I still don't understand how this device is connected... I don't even see the pins for the electrical source...
The connectors are covered in electrical tape in the second picture.

This is how a server for the previous SXM4 based generation of these look like, you can fit 8 of these GPU's in the sockets at the bottom (not the PCIe slots):


Posted on Reply
Add your own comment
Jun 26th, 2022 00:55 EDT change timezone

New Forum Posts

Popular Reviews

Controversial News Posts