This week is Computex 2024 in Taipei, and NVIDIA is kicking it off with the first keynote. In the keynote, we expect to hear a lot about AI and new platforms using the company’s GPUs. Although I am not traveling (although we have an away team for booth coverage), I am up with the baby and will be covering this live, so please excuse typos.
NVIDIA Computex 2024 Keynote
The keynote is starting out with a “The More You Buy, The More You Save” discussion. Most of Jensen’s keynotes have some flavor of this, and it has been going on for years.
NVIDIA AI Notebooks
This is going to be the Computex of the AI PC. NVIDIA highlights three ASUS and one MSI notebooks and says that it has not just AI PCs, but also AI PCs that run CUDA, the language of high-end generative AI.
This is going to be a really interesting challenge for NVIDIA. When Intel, AMD, Qualcomm and others say they have dedicated AI accelerators,
NVIDIA Grace Blackwell GB200
On the Blackwell side, Jensen said that the MTBF of a 10,000 GPU cluster is measured in hours. MTBF of a 100,000 GPU cluster is measured in minutes. That is why the GB200 has a RAS engine.
Here is the production Grace Blackwell GB200.
Here is a cool chart. Here is NVIDIA’s chart on how much energy it would take to train GPT-4 using each generation of NVIDIA’s GPUs.
Here are the heatsinks on the NVIDIA HGX B200 15kW board according to Jensen.
Here is something interesting. It seems like the GB200 NVL72 is now 100kW not 120kW.
Here is the 5th Gen NVLink Switch that makes the GB200 NVL72 possible.
NVIDIA showed off the passive NVLink spine. If the connectors look a bit off, they have to have a bit of play in them to support blind mating.
NVIDIA is also discussing Spectrum-X for the 400GbE era. NVIDIA is showing the 51.2T Spectrum-X800 Ultra for the ConnectX-8 era. It then has a roadmap to the ConnectX-9 NIC and 102.4T switch era for 1.6Tbps networking.
It seems to be planning for 800Gbps networking (which to a single host requires PCIe Gen6) in 2025 and then 1.6Tbps networking in 2026. So this seems to point to NVIDIA also planning for a PCIe Gen7 NIC in 2026. When we discuss PCIe speed steps accelerating, this is a solid signal.
Next up will be the NVIDIA Rubin GPU and that will have a NVIDIA Vera CPU. This is intersecting with the ConnectX-9 which is 2026. It seems like we get Blackwell in 2024, Blackwell Ultra in 2025, then Rubin in 2026 with Rubin Ultra perhaps in 2027.
NVIDIA NVL2
It is no secret that the NVIDIA DGX GB200 NVL72 is going to be out of reach for many, especially given its price tag of several million and 120kW per rack power budget. NVIDIA has sold the Grace Blackwell building block with only two GPUs instead of 72 and calls it the NVL2.
Final Words
NVIDIA has a lot going on. It is becoming apparent that while we discuss NVIDIA’s hardware as the company’s profit driver today, NVIDIA is spending much more time on its software, even in Taiwan, where its hardware primarily comes from. The multi-trillion-dollar question is how NVIDIA will monetize the software side.
good