4 ms·
K3 is natively trained to mxfp4, if they cannot get a hold of Blackwell chip, it is meaningless. Hopper does not do native 4-bit floating math. Either they hav
by ycui7 2mo ago
K3 is natively trained to mxfp4, if they cannot get a hold of Blackwell chip, it is meaningless. Hopper does not do native 4-bit floating math.
Either they have Blackwell with native 4-bit floating math, or they use have Chinese domestic NPU that support mxfp4 natively.
The article’s statement does not make sense.
- orliesaurus 2mo agoWhat do you mean by "it's meaningless" can you expand?
- sho 2mo agoThey mean that K3's native floating point is MXFP4, a newer standard that's not supported in Hopper, an older nvidia chipset, so it's impossible the GPUs in question were Hopper (the article specifically mentions H200s), which is correct. Probably either the author or the people they interviewed got the specific GPU details mixed up or wrong. Maybe it was B200s.
- orliesaurus 2mo agoOh, okay, thank you for the explanation. I was under the impression that we, the United States, were banning exports of such GPUs.
- gpugreg 2mo agoNotably, MXFP4 was introduced at the (much less costly) supervised fine-tuning stage after pretraining, so the number of B200/B300 GPUs could be relatively small in comparison to the number of H200 GPUs used during pretraining (or maybe not, who knows). (Kimi K3 tech report section 4.1.1 https://arxiv.org/pdf/2607.24653 https://arxiv.org/pdf/2607.24653)
- throwa356262 2mo agoHuawei A910 is supposed to have native support for MXFP4. The problem is that huawei doesn't have the capacity to build more GPUs right now. Those yield numbers must be really terrible. Personally, I think 20K nvidias is a stop gap solution because they really don't have the capacity to serve their models to earn any money right now.
- Tempest1981 2mo ago> Either they have Blackwell From the article: A person familiar with Moonshot’s procurement strategy confirmed that the company does indeed have a channel for accessing Blackwell processors via Southeast Asia. They didn’t specify whether this was a rental channel...
- basiccalendar74 2mo agoWeights will be fetched 4-bit from memory, but compute happens in FP8 (W4A8) or BF16 (W4A16). This is still better than fetching 8-bit weights from memory. Both W4A8 and W4A16 schemes are supported by Hopper GPUs and commonly used to serve mxfp4 Kimi models on Hopper.