5 ms·
DeepSeek-V3 Technical Report
- kristjansson 2y agoHasn't been updated for the -0324 release unfortunately, and diff-pdf shows only a few small additions (and consequent layout shift) for the updated arxiv version on Feb 18.
- skummetmaelk 2y agoThe fact that you can unironically put the "only" modifier on a training time of 2.8 million GPU hours is nuts.
- chvid 2y agoIf they have a cluster with 2,000 H800 GPUs (which is what they have stated in public) training would take 2,800,000 / (2,000 * 24 * 30) ~ 2 months. A cluster of 2,000 GPUs is what a second tier AI lab has access to. And it shows that you can play in the state of the art LLM-game with some capital and a lot of brains.
- ArtTimeInvestor 2y agoIsn't the price of an H800 like $30k? I don't know what your household budget is, but $60M might not be what most people associate with "some capital".
- andai 2y agoYesterday GPT asked me if I'd like to train a small LLM and I laughed out loud. That being said I'm amazed how far 1B models have come. I remember when TinyLlama came out a few years ago, it was not great. ($40K training cost iirc.) That was a 1B model, but these days even 0.5B models are remarkably coherent.
- chvid 2y agoIt is a lot less than what Google, OpenAI etc have. And the GPUs would be a shared resource so what you should calculate is what it would have cost to rent them - probably something like 2 m.
- charcircuit 2y agoAn H100 has 14592 CUDA cores. 2000 * 14592 already gives you more than 2 million cores.
- andai 2y agoCan someone put this into perspective? I'm finding heterogenous data on other models, i.e. number of tokens, number of GPUs used, cost, etc. It's hard to compare it all.
- system2 2y ago[flagged]
- MPSFounder 2y ago[flagged]
- rupertqin 2y ago[dead]
- litbear2022 2y agoYeah! Just steal new Boeing 6th-gen stealth fighter from slides.
- rupertqin 2y ago[dead]
- 0x008 2y agoThis model is open source and Beats all proprietary models in benchmarks. How is this stagnant?
- aoeusnth1 2y agoNo it doesn’t.
- starship731 2y agoDeepseek v3-0324 (new checkpoint) beats ALL but 1 proprietary AND non-thinking LLMs by a significant margin. Check livebench.ai & Artificial Analysis benchmark for details. The only non-thinking LLM the new V3 doesn't decisively thrash is GPT 4.5 which is more than 100 times more expensive than V3 and yet is only a few (essentially negligible) percentage points better than it.
- 2y ago
- gdiamos 2y agoNice to see a return to open source in models and training systems.
- jxjnskkzxxhx 2y agoCapitalism is beautiful.
- FirmwareBurner 2y agoWhen China schools the US on capitalism.
- jxjnskkzxxhx 2y agoTIL China invented the free source movement. Cool story bro.
- Centigonal 2y agoThe GPU-hours stat here allows us to back out some interesting figures around electricity usage and carbon emissions if we make a few assumptions. 2,788,000 GPU-hours * 350W TDP of H800 = 975,800,000 GPU Watt-hours 975,800,000 GPU Wh * (1.2 to account for non-GPU hardware) * (1.3 PUE [1]) = 1,522,248,000 Total Wh, or 1,522,248 kWh to train DeepSeek-V3 (1,522,248 kWh) * (0.582kg CO2eq/kWh in China [2]) = 885,948 kg CO2 equivalents to train DeepSeek-V3 A typical US passenger vehicle emits about 4.6 metric tons of CO2 per year. [3] 885,948 kg CO2 per DeepSeek / 4,600 kg CO2 per car = 192.6 cars per DeepSeek So, the final training run for DeepSeek-V3 emitted as much greenhouse gasses as would be emitted from running about 193 more cars on the road for a year. I also did some more math and found that this training run used about as much electricity as 141 US households would use over the course of a year. [4] [1] https://enviliance.com/regions/east-asia/cn/report_10060 https://enviliance.com/regions/east-asia/cn/report_10060 [2] https://ourworldindata.org/grapher/carbon-intensity-electricity https://ourworldindata.org/grapher/carbon-intensity-electric... [3] https://www.epa.gov/greenvehicles/greenhouse-gas-emissions-typical-passenger-vehicle https://www.epa.gov/greenvehicles/greenhouse-gas-emissions-t... [4] divided total kWh by the value here: https://www.eia.gov/tools/faqs/faq.php?id=97&t=3 https://www.eia.gov/tools/faqs/faq.php?id=97&t=3
- hugs 2y agothe nice thing about ai's energy usage is that no one complains about bitcoin's energy usage anymore. (i'm kidding, people still complain.)
- Centigonal 2y agoActually -- and this is insane -- the amount of electricity required to train DeepSeek-V3 would power the Bitcoin network for all of 5 minutes. DeepSeek would have to fully train a brand new V3 every week to approach the kinds of power consumption numbers that individual bitcoin mining facilities are doing. The energy use from BTC is ludicrous. (I'm assuming 155 TWh/yr for Bitcoin, using the low-end estimate from here: https://www.polytechnique-insights.com/en/columns/energy/bitcoin-electricity-consumption-comparable-to-that-of-poland/ https://www.polytechnique-insights.com/en/columns/energy/bit... )
- tmabraham 2y agohttps://x.com/iScienceLuvr/status/1905144432791609480 https://x.com/iScienceLuvr/status/1905144432791609480
- benob 2y agoI like that they give advice to hardware manufacturers: - offload communication to a dedicated co-proc - implement decent precision for accumulating fp8 operations - finer-grained quantization ...
- danielhanchen 2y agoRe DeepSeek-V3 0324 - I made some 2.7bit dynamic quants (230GB in size) for those interested in running them locally via llama.cpp! Tutorial on getting and running them: https://docs.unsloth.ai/basics/tutorial-how-to-run-deepseek-v3-0324-locally https://docs.unsloth.ai/basics/tutorial-how-to-run-deepseek-...
- behohippy 2y agoThese articles are gold, thank you. I used your gemma one from a few weeks back to get gemma 3 performing properly. I know you guys are all GPU but do you do any testing on CPU/GPU mixes? I'd like to see the pp and t/s on pure 12 channel epyc and the same with using a 24 gig gpu to accelerate the pp.
- danielhanchen 2y agoOh fantastic! Oh for MoEs like DeepSeek, technically GPUs aren't that necessary! I actually tested on 1x H100 I think it was 30 layers offloaded, and the other 30 are on CPU - it wasn't that bad at all!