TopSpoiler
Regular
Here's a list of NVIDIA's sessions at the upcoming GDC (except the last one):
NVIDIA discussion [2025]
- Thread starter SlmDnk
- Start date
DavidGraham
Veteran
Samsung Securities estimated total CoWoS capacities (L/S/R) for 2025: NVIDIA 64%, Broadcom 14%, Amazon 12%, AMD 8%.
On other news, OpenAI just signed a 5 years contract worth 12 billion $ with CoreWeave (an NVIDIA backed data center company), for renting GPUs (H200/B200).
On other news, OpenAI just signed a 5 years contract worth 12 billion $ with CoreWeave (an NVIDIA backed data center company), for renting GPUs (H200/B200).
DavidGraham
Veteran
TSMC has proposed a joint venture with NVIDIA, AMD, Qualcom and Broadcom to run Intel foundries.
Nope, if you go to actual source (https://www.reuters.com/technology/...v-nvidia-amd-broadcom-sources-say-2025-03-12/) it says right in the beginning that "TSMC would take a stake of no more than 50% in joint venture operating Intel foundries-sources"
US government has been clear that Intel Foundries can't be sold outside.
If it's held private, Intel has to own 50.1% minimum, if it's made public, no outside investor can hold over 35 %.
The 50.1% requirement was set into place by the CHIPS Infrastructure act here in the US, specifically because Intel was receiving funding. If the current administration makes good on their threats to kill the CHIPS act, would the 50.1% rule still be binding? Even if technically true, would it be enforced? Given the proclivities of the current US leadership, I'm not so sure...
arandomguy
Veteran
I think there's 2 aspects of this. There might not be a restriction on Intel having to maintain a set amount of ownership but there is almost certainly going to be restrictions on foreign ownership, if anything more so with the current administration.
From my perspective this was always going to be the issue with selling Intel and/or splitting it's foundry. The companies that would be most interested and capable in buying either would be non US entities which would almost certainly be blocked regardless of who is the government. While with US companies the ones most interested and capable are almost certainly going to run into anti-trust concerns.
From my perspective this was always going to be the issue with selling Intel and/or splitting it's foundry. The companies that would be most interested and capable in buying either would be non US entities which would almost certainly be blocked regardless of who is the government. While with US companies the ones most interested and capable are almost certainly going to run into anti-trust concerns.
Agreed re: foreign ownership, however keep in mind NVIDIA, AMD, Qualcomm and Broadcom are all "American" companies, only TSMC is the foreign investor here.
And so, with a proposed consortium of mostly American companies to satisfy a cursory foreign investment challenge, I refer back to my prior comment.
And so, with a proposed consortium of mostly American companies to satisfy a cursory foreign investment challenge, I refer back to my prior comment.
Well that was an interesting sales pitch. Jensen basically shat all over Hopper and waved away Blackwell Ultra. Must feel great for all the companies who spent billions on those chips.
Rubin and the silicon photonics stuff do look insane though. Hope TSMC has their CoWoS stuff sorted. The photonics chips are using CoWoS too.
Oh Blackwell is no longer one GPU, two dies they're walking that back. It's now 2 GPUs, 2 dies. Making Rubin Ultra 4 GPUs, 4 dies.
Rubin and the silicon photonics stuff do look insane though. Hope TSMC has their CoWoS stuff sorted. The photonics chips are using CoWoS too.
Oh Blackwell is no longer one GPU, two dies they're walking that back. It's now 2 GPUs, 2 dies. Making Rubin Ultra 4 GPUs, 4 dies.
DavidGraham
Veteran
NVIDIA announced B300, releasing in 2H 2025, 50% faster than B200, and with 50% more VRAM. Then, Rubin will launch in 2H 2026, with 3x the performance of B300 (measured in FP4). Rubin will be two reticle sized GPUs fused together, Rubin Ultra will follow in 2H 2027, with 4x the performance of Rubin, it will be 4 reticle sized GPUs fused together.
In 2028, there will be the Feynman GPU arch, NVIDIA gave no details about it though.
wccftech.com
wccftech.com
In 2028, there will be the Feynman GPU arch, NVIDIA gave no details about it though.
NVIDIA Rubin & Rubin Ultra With Next-Gen Vera CPUs Start Arriving Next Year: Up To 1 TB HBM4 Memory, 4-Reticle Sized GPUs, 100PF FP4 & 88 CPU Cores
NVIDIA laid out plans for its next-gen AI powerhouses, the Rubin & Rubin Ultra GPUs, along with Vera CPUs, taking the segment to new heights.
wccftech.com
NVIDIA Unveils Next-Gen Feynman GPU In Updated Roadmap, Arriving in 2028 With Next-Gen HBM Memory
NVIDIA has just named its next-generation GPU that arrives after Blackwell and Rubin, called Feynman, and it will use a new HBM memory.
wccftech.com
DegustatoR
Legend
- NVIDIA Blackwell Ultra “B300” Unleashing In 2H 2025 – 1.1 Exaflops of FP4 Compute, 288 GB HBM3e Memory, 50% Faster Than GB200
- NVIDIA Rubin & Rubin Ultra With Next-Gen Vera CPUs Start Arriving Next Year: Up To 1 TB HBM4 Memory, 4-Reticle Sized GPUs, 100PF FP4 & 88 CPU Cores
- NVIDIA Unveils Next-Gen Feynman GPU In Updated Roadmap, Arriving in 2028 With Next-Gen HBM Memory
nVidia announced new workstation systems with Grace and Blackwell Ultra: https://www.nvidia.com/en-us/products/workstations/dgx-station/
288GB HBM3e and 500GB from Grace.
288GB HBM3e and 500GB from Grace.
Nvidia not in talks with TSMC to take over Intel foundry: Jensen Huang
No immediate impact from Trump tariffs but more onshoring needed, CEO says
asia.nikkei.com
Last edited:
DavidGraham
Veteran
GB300 is revealed to have 4 B300 GPUs each is equipped with 288GB HBM3e memory, GB200 only had two B200 GPUs (each with 192GB HBM3e).
It was already right in the slide they showed at the keynote, though?
How is this any different to the GB200 NVL4?
www.tomshardware.com
Nvidia introduces a new merged CPU and GPU AI processor — GB200 Grace Blackwell NVL4 Superchip has four B200 GPUs, two Grace CPUs
Nvidia's GB200 NVL4 solution takes things to a new level by implementing four B200 GPUs and two Grace CPUs on a single motherboard.
It's now called a single unit instead of NVL4 configurationHow is this any different to the GB200 NVL4?
Nvidia introduces a new merged CPU and GPU AI processor — GB200 Grace Blackwell NVL4 Superchip has four B200 GPUs, two Grace CPUs
Nvidia's GB200 NVL4 solution takes things to a new level by implementing four B200 GPUs and two Grace CPUs on a single motherboard.
View attachment 13343
It's now called a single unit instead of NVL4 configuration
They're advertising GB300 NVL16 and NVL72 so I would expect them to stick with NVL4 for this too.
Yeah, I mixed it in my mind with Rubin Ultra which was 4 instead of 2 GPUs per "single GPU"
Yep, Jensen said they're sticking with the old naming for Blackwell and changing it for Rubin. Something about aligning the naming with NVLink topology.
Similar threads
