StraightEdge
SENIOR MEMBER
- Joined
- Jan 21, 2021
- Messages
- 2,262
- Reaction score
- -6
- Country
- Location
When the US dropped updated AI restrictions, we thought the US locked down every single loophole conceivable. To our surprise, Nvidia still found a way to ship high performance GPUs into China with their upcoming H20, L20, and L2 GPUs. Nvidia already has product samples for these GPUs and they will go into mass production within the next month, yet again showing their supply chain mastery.
Furthermore, one of the China specific GPUs is over 20% faster than the H100 in LLM inference, and is more similar to the new GPU that Nvidia is launching early next year than to the H100! Today we will share details about Nvidia’s new GPUs, the H20, L20, and L2. The detailed specs include FLOPS figures, NVLink bandwidth, power consumption, memory bandwidth, memory capacity, die size, and more. We will also discuss performance in more detail. The simple specs are here:
For an explanation of the US controls, see here. In short, Nvidia is perfectly straddling the line on peak performance and performance density with these new chips to get them through the new US regulations.
We are also sharing our revised Nvidia quarterly forecast for FY24 and FY25 based on these revelations. As a reminder, we have had H800 and A800 bans included in our model for many months now. With the new China specific GPU, we no longer have CoWoS capacity catching up to demand in late Q2 and are now revising volumes back up. Note that only the H20 uses CoWoS.
Lastly, we will also share specs and thoughts on the new Huawei Ascend 910B, a SMIC 7nm based GPU that lands between the A100 and H100 in performance. It is being purchased by multiple Chinese hyperscalers.
www.semianalysis.com
Furthermore, one of the China specific GPUs is over 20% faster than the H100 in LLM inference, and is more similar to the new GPU that Nvidia is launching early next year than to the H100! Today we will share details about Nvidia’s new GPUs, the H20, L20, and L2. The detailed specs include FLOPS figures, NVLink bandwidth, power consumption, memory bandwidth, memory capacity, die size, and more. We will also discuss performance in more detail. The simple specs are here:
For an explanation of the US controls, see here. In short, Nvidia is perfectly straddling the line on peak performance and performance density with these new chips to get them through the new US regulations.
We are also sharing our revised Nvidia quarterly forecast for FY24 and FY25 based on these revelations. As a reminder, we have had H800 and A800 bans included in our model for many months now. With the new China specific GPU, we no longer have CoWoS capacity catching up to demand in late Q2 and are now revising volumes back up. Note that only the H20 uses CoWoS.
Lastly, we will also share specs and thoughts on the new Huawei Ascend 910B, a SMIC 7nm based GPU that lands between the A100 and H100 in performance. It is being purchased by multiple Chinese hyperscalers.
Nvidia's New China AI Chips Circumvent US Restrictions - H20, L20, and L2 Specifications
2024 Revenue, EPS, H20, L20, and L2 Full Specifications
www.semianalysis.com
