NVIDIA will manufacture H100 GPUs using TSMC 4-nm process.
Click here to post a comment for NVIDIA will manufacture H100 GPUs using TSMC 4-nm process. on our message forum
geogan
Opted to outsource? Haven't all NVidia GPUs been made at TSMC for years (apart from the pandemic Samsung emergency)?
vestibule
Well the rumour was 5nm so @4nm. Hey hey hey for performance and energy savings.
Good news.
Here's hoping that TSMC yields will be 100%.
user1
the current ampere gpus are manufactured on samsung's 8nm, so this would be a return to tsmc, presumably this wasn't expected
edit: I wouldn't exactly call the swtich to samsung an emergency, I Think it has more to do with total allocation available at tsmc, after nvidia likes BIG chips
Astyanax
the rumor for hopper has been both 4 and 5nm, lovelace is on 5n
fantaskarsef
https://i.kym-cdn.com/entries/icons/original/000/005/574/takemymoney.jpg
Kaarme
Hard to say if it's good for availability that AMD, Nvidia, and Intel GPUs are all manufactured by TSMC. Last year just AMD was made there, Nvidia by Samsung, yet you still couldn't buy anything for a good long while, and when you could, things were so scarce that the prices were in the stratosphere. But then again, it seems like Nvidia may continue to sell the 3000 series (from Samsung fabrication plants) even after the 4000 is out. AMD might also offer the 7nm 6000 series at the same time with the 7000 series. Not very exciting, but it would increase production numbers, assuming other components, like GDDR, won't form another bottleneck.
JamesSneed
@Hilbert Hagedoorn Should note this doesn't mean they will use 4nm for the next gen graphics cards. That's what I assume people are going to read into this over here.
JamesSneed
Rumor might still be true. It could be 4nm is used for the h100 and 5nm for everything else. The h100 is so different it will have its own masks etc so it could be designed on a different node. I would not read too much into this yet as its still all up to Nvidia and what they chose to use.
tunejunky
spot on.
this is for Nvidia's most profitable product line and this is the ultimate expression of the monolithic die.
4n "ain't cheap" and regardless of what many may think, as a monolithic die you're subject to a lower yield in every case against a MCM so it will be very interesting to see Hopper and the latest Instinct battle it out especially in regards to pricing. i suspect that Nvidia's marketing is going to be far narrower than AMD's as AMD can add or drop chiplets (up to the socket size) as they like.
as for us, the great unwashed masses, i rather suspect 5n as the production cost is significantly lower but the yield should be in a similar ballpark. TSMC has massive capacity @ 5n of various flavors.
to put it another way, TSMC has more capacity @ 5n now than they did with 7n when Ryzen launched
so i don't think there will be any production logjams.
mainly because if there were it would be making news right about now.
Maddness
Wasn't Hopper supposed to be MCM and not monolithic. That's what everything I have read was pointing to.
vestibule
@tunejunky
Uh huh. Informative post.
tunejunky
nope they went with the upgraded NV Switch (900Gb/sec bi-directional)
https://www.nvidia.com/en-us/technologies/hopper-architecture
but that doesn't mean that the uArch won't appear at a later date.
Nvidia has seen what TSMC is doing with the M1 Ultra and AMD's Infinity Fabric.
my sneaking suspicion is the interconnect of the M1 Ultra is exactly what Nvidia's designers want.
whether they are able to license it is another Q as AMD isn't going to let them use Infinity Fabric.
Denial
Like 90% of rumors about hopper were wrong.
They released more info about their approach to MCM and it sounds much different than their whitepaper from 2017. Instead of splitting up actual GPU modules up, it's more like having dedicated MCM parts for specific workloads/flexible cache/bandwidth setups for workloads that require it.
http://3s81si1s5ygj3mzby34dq6qf-wpengine.netdna-ssl.com/wp-content/uploads/2022/01/nvidia-copa-ai-hpc-options.jpg
https://www.nextplatform.com/2022/01/06/nvidia-research-plots-a-course-to-multiple-multichip-gpu-engines/
I guess they don't feel like MCM is worth pursuing in the traditional sense yet.
Nvidia has NVLink, why would they try to license Infinity Fabric or anything from Apple? Maybe I-LSI from TSMC but it would still utilize NVLink on top of that.
JamesSneed
We are on the same page. At this point I would say its 90% likely Nvidia consumer graphics card chips are on 5nm. It shouldn't be 6nm because that probably doesn't get them the density and power savings they will need. Its probably not 4nm due to higher costs. So that leaves you with 5nm.
tunejunky
NV Link isn't for linking chiplets.
but i daresay you're right about them doing that once they have an interposer like Infinity Fabric or TSMC/Apple's M1 Ultra interface
Astyanax
yes it is.
https://nvidianews.nvidia.com/news/nvidia-opens-nvlink-for-custom-silicon-integration
tunejunky
wow
good for them (really no sarcasm)
this means that Nvidia will be in the game for the future. it's a big deal as both AMD and Intel have mcm capabilities.
but as the node it's on is questionable i would suspect at first it will be joining full blown chips rather than chiplets as that's an easier job than 4-8 chiplets so i suspect it will be like the M1 Ultra, just like emib/fovereros is quite large but can be (eventually) engineered to a smaller node.
but it's not for Hopper (yet), external NV Link is
cucaulay malkin
same as 7nm ampere and 8nm ampere
only this time the difference will be small,not humongous.4nm non-plus vs 5nm are very similar.
tunejunky
:D:D:D
not quite but you're onto it
8n Ampere was a clusterfuck from Samsung - not in the finished product, but due to lower yields than Nvidia expected and Samsung promised. the TSMC product has far better yields from a more refined process.
but other than picking that nit you're right
