46
0Episode: OpenAI Jalapeño: The Chip That Beat NVIDIA
Duration: approximately 9 minutes
Level: B1 (Intermediate)
---
[Mike]: Welcome back to Learn English with Podcasts! Sarah, here's a strange question to start - what do a Japanese pork cutlet, an Indian curry, a bowl of chickpeas, and a chili pepper have in common?
zh:欢迎回到 Learn English with Podcasts！Sarah，开场先来个奇怪的问题——日式炸猪排、印度咖喱、一碗鹰嘴豆和一根辣椒，有什么共同点？
[Sarah]: Hmm... somebody planned a very international lunch?
zh:嗯……有人计划了一顿非常国际化的午餐？
[Mike]: Funny you say that - those are real product names inside OpenAI's data centers. The tray holding CPUs is called Katsu. The chip tray is Vindaloo. The network switch is Chana.
zh:你说巧了——这些是 OpenAI 数据中心里真实的产品名。装 CPU 的托盘叫 Katsu，装芯片的托盘叫 Vindaloo，交换机叫 Chana。
[Sarah]: Wait, seriously? Engineers named computer parts after dinner?
zh:等等，认真的吗？工程师用晚餐的名字给电脑零件命名？
[Mike]: And the star of this week - OpenAI's first chip designed completely in-house - is named after the pepper. It's called Jalapeño.
zh:而本周的主角——OpenAI 第一颗完全自研的芯片——是用那个辣椒命名的。它叫 Jalapeño。
[Sarah]: Okay, a pepper chip. So what did it do? Because I'm guessing you didn't come here with food news.
zh:好，一颗辣椒芯片。那它干了什么？我猜你不会专程来报美食新闻。
[Mike]: It picked a fight with the king. An independent testing lab called SemiAnalysis put Jalapeño side by side with NVIDIA's strongest AI chips - GB200 and GB300. And the pepper beat them. On every chart.
zh:它向王者下了战书。一家叫 SemiAnalysis 的独立测试实验室，把 Jalapeño 和英伟达最强的 AI 芯片 GB200、GB300 放在一起对比。结果辣椒赢了。每张图上都赢。
[Sarah]: Every chart? Give me a number I can feel.
zh:每张图都赢？给我一个有感觉的数字。
[Mike]: My favorite one: on a model called GPT-OSS 120B, Jalapeño pushed out 1459 tokens per second. Tokens are the little pieces of words that AI generates one by one.
zh:我最喜欢的：在一个叫 GPT-OSS 120B 的模型上，Jalapeño 每秒输出 1459 个 token。token 就是 AI 一个个生成的小文字碎片。
[Mike]: NVIDIA's best managed about 535 on the same job. Not even half.
zh:同样的活儿，英伟达最强只跑出大约 535 个。连一半都不到。
[Sarah]: 1459 versus 535? That's not winning a race, that's lapping people.
zh:1459 对 535？这不是赢比赛，这是套圈了。
[Mike]: And analysts went wild. Dylan Patel, a famous semiconductor analyst, wrote that not just Blackwell got surpassed - even NVIDIA's newest Rubin was beaten. Even Sam Altman commented, and his whole statement was: we built a chip that is absurdly fast.
zh:分析师们也炸锅了。著名半导体分析师 Dylan Patel 写道：被超越的不只是 Blackwell，连英伟达最新的 Rubin 也被击败了。连奥特曼都出来表态，他的原话就一句：我们造了一颗快得离谱的芯片。
[Sarah]: Confident. Okay, but speed per second is one thing. When my food delivery says thirty minutes, what I care about is when the doorbell rings. Total time.
zh:够自信。不过每秒速度是一回事。外卖说三十分钟送到时，我在乎的是门铃什么时候响。总时间。
[Mike]: Same story there. One task: Jalapeño finished in 1.65 seconds. GB300 needed almost 6 seconds - about 3.6 times slower.
zh:总时间也是同样的故事。同一个任务：Jalapeño 用 1.65 秒完成，GB300 要将近 6 秒——慢了约 3.6 倍。
[Sarah]: Four extra seconds sounds harmless for a chat, honestly.
zh:说实话，多 4 秒在聊天里好像没什么伤害。
[Mike]: For chatting, sure. But think about AI agents - programs that run long jobs step by step, like booking flights or filling in forms. Dozens of steps, each one a little slow... the delays multiply. Slow chips make painfully slow agents.
zh:聊天确实没事。但想想 AI agent——那些一步步执行长任务的程序，比如订机票、填表格。几十步，每步慢一点……延迟会相乘。慢芯片造出慢到让人痛苦的 agent。
[Sarah]: Ah, that's why everyone suddenly cares about milliseconds. Okay, there's also a wild number going around online - something about 104 times?
zh:啊，所以大家才突然在意毫秒。对了，网上还流传一个疯狂数字——好像是 104 倍？
[Mike]: 104.3 times, and it needs context. Engineers pushed GB300 up to its own maximum decoding speed - around 169 tokens per second. At exactly that point, Jalapeño's throughput was 104.3 times higher.
zh:104.3 倍，这个数字需要上下文。工程师把 GB300 推到它自己的极限解码速度——大约每秒 169 个 token。恰恰在那个点上，Jalapeño 的吞吐量是它的 104.3 倍。
[Sarah]: So where the champion collapses out of breath, the newcomer is still cruising?
zh:所以冠军喘不上气的地方，新人还在匀速巡航？
[Mike]: Perfect image. To be fair, at each chip's best setting the gap is smaller - between 1.5 and 1.9 times across three models. Still stunning for a first attempt.
zh:画面感满分。公平地说，各自跑在最佳设置时差距小一些——三个模型上是 1.5 到 1.9 倍。但对第一次尝试来说依然惊人。
[Sarah]: What about electricity? Chips this fast usually drink power like crazy.
zh:那耗电呢？这么快的芯片通常是电老虎吧。
[Mike]: Another shock. Jalapeño runs at 700 watts. GB300 needs 1400. Half the electricity, double the work.
zh:又一个震撼点。Jalapeño 跑在 700 瓦，GB300 需要 1400 瓦。一半的电，双倍的活。
[Sarah]: Even my hair dryer feels embarrassed now.
zh:连我的吹风机现在都觉得丢脸。
[Mike]: Ha! Count cooling and everything in, and running one Jalapeño costs about 1.56 dollars per hour - almost identical to an older H100, and far below NVIDIA's newest Rubin at 3.61 dollars.
zh:哈！把散热全都算进去，跑一颗 Jalapeño 每小时成本约 1.56 美元——跟老将 H100 几乎一样，远低于英伟达最新 Rubin 的 3.61 美元。
[Sarah]: Cheaper per hour than my streaming subscriptions... okay wait. Is the pepper even showing its final form here?
zh:每小时比我的视频会员还便宜……等等。这颗辣椒这是最终形态了吗？
[Mike]: Not even close! They ran these tests WITHOUT speculative decoding - a clever guessing trick that speeds up AI output. Experts estimate that trick alone could cut the cost per token by more than two thirds.
zh:差远了！他们是在没开投机解码的情况下测的——一个加速 AI 输出的聪明猜词技巧。专家估算，光这一招就能把每 token 成本再砍掉三分之二以上。
[Sarah]: Hold on. Rewind. This is OpenAI's FIRST chip ever. Chip companies usually spend years learning this craft. How is this possible?
zh:等一下。倒带。这是 OpenAI 有史以来第一颗芯片。芯片公司通常要花多年修炼这门手艺。这怎么可能？
[Mike]: Two reasons. First, raw speed: design to factory took nine months. The industry norm is 18 to 36 months.
zh:两个原因。第一，纯粹的速度：从设计到工厂只花了九个月。行业常规是 18 到 36 个月。
[Sarah]: Nine months! Some online shopping deliveries take longer!
zh:九个月！有些网购快递都比这慢！
[Mike]: Reason two is the wild one. OpenAI brought an AI into the design room - their strongest model, called GPT-Astra. Probably the same model everyone keeps whispering about as GPT-6.
zh:第二个原因才是真离谱。OpenAI 把一个 AI 请进了设计室——他们最强的模型，叫 GPT-Astra。很可能就是大家一直悄悄传的 GPT-6。
[Mike]: It explored design options, compressed the whole build-measure-test loop, and fine-tuned tiny circuits. One math unit shrank by 8%, another by 10% - while getting faster and cooler.
zh:它帮忙探索设计方案，压缩整个"搭建-测量-验证"循环，还微调了微型电路。一个运算单元缩小了 8%，另一个缩小了 10%——同时速度更快、发热更低。
[Sarah]: So an AI shrank circuits for a chip whose entire job is running AI?
zh:所以是一个 AI 为一颗芯片缩小了电路，而这颗芯片的全部工作就是跑 AI？
[Mike]: Yes. AI designs the chip. The chip runs smarter AI. Smarter AI designs the next chip. Round and round. Engineers call it a flywheel - heavy at first, then almost impossible to stop.
zh:没错。AI 设计芯片，芯片跑更聪明的 AI，更聪明的 AI 再设计下一代芯片。一圈又一圈。工程师管这叫飞轮——起步很沉，转起来之后几乎停不下来。
[Mike]: One more detail shows how extreme it is. The main computing area measures about 840 square millimeters. The lithography machine's biggest possible canvas is 858. They used nearly all of it.
zh:还有个细节能看出它有多极端。主计算区域约 840 平方毫米。光刻机能画的最大画布是 858 平方毫米。几乎全用上了。
[Sarah]: Like packing your suitcase until the airport scale reads 22.9 kilograms.
zh:就像把行李箱打包到机场秤显示 22.9 公斤为止。
[Mike]: Exactly. And the next version is already in production, promising about 25% better performance per watt.
zh:就是这样。而且下一版已经在生产了，预计每瓦性能还会再高约 25%。
[Sarah]: Impressive engineering. But NVIDIA has more than chips, right? Years ago every engineer learned their software... CUDA?
zh:工程上确实漂亮。但英伟达手里不止芯片吧？多年前每个工程师都学过他们的软件……CUDA？
[Mike]: Right - CUDA has been the famous moat. A moat is the deep water ring around a castle. Nearly twenty years of tools and habits sat around NVIDIA's chips. Switching hardware meant relearning everything from zero.
zh:对——CUDA 就是那条著名的护城河。护城河就是城堡周围那圈深水。围绕英伟达芯片积累了将近二十年的工具和习惯。换硬件意味着一切从零重学。
[Sarah]: So did the pepper swim across?
zh:那这颗辣椒游过来了吗？
[Mike]: SemiAnalysis wrote a line nobody can stop quoting - the CUDA moat may be dead. Their reasoning: NVIDIA's hardware isn't weak. Rubin actually reached the factory one month before Jalapeño. What lost the race was how fast the software around the chip grew up.
zh:SemiAnalysis 写了一句大家疯狂引用的话——CUDA 护城河可能已经死了。他们的推理是：英伟达硬件并不弱。Rubin 其实比 Jalapeño 早一个月进工厂。输掉比赛的是芯片周边软件长大的速度。
[Sarah]: And starting from zero helped OpenAI there, right? No old code dragging behind.
zh:而从零开始反而帮了 OpenAI，对吧？没有旧代码拖后腿。
[Mike]: A blank page lets you draw anything. Now, ready for the ironic part?
zh:白纸才能随便画。现在，准备好听最讽刺的部分了吗？
[Sarah]: Always.
zh:随时奉陪。
[Mike]: Those clever OpenAI models that helped design the pepper? While they worked on it, they were running on NVIDIA GPUs.
zh:帮着设计这颗辣椒的那些聪明 OpenAI 模型？它们干活的时候，跑的就是英伟达 GPU。
[Sarah]: Wait. So NVIDIA's machines raised their own challenger? Like training the apprentice who comes back to take over your shop?
zh:等等。所以英伟达的机器养大了自己的挑战者？就像培训了一个学成回来接管你店面的学徒？
[Mike]: And they're not stopping at one pepper. Last October, OpenAI signed a deal with Broadcom for 10 gigawatts of custom chips - roughly ten nuclear reactors running flat out.
zh:而且他们不会止步于一颗辣椒。去年 10 月，OpenAI 跟博通签了 10 吉瓦定制芯片的大单——差不多相当于十座核电站满负荷发电。
[Sarah]: Ten reactors' worth of chips. And Jalapeño is only generation one?
zh:十座核电站级别的芯片。而 Jalapeño 只是第一代？
[Mike]: Generation two is deep in development, generation three taking shape, mass production climbing through 2027. Though there's one shadow: their data center chief, Chris Malone - who oversaw the huge Stargate project - just left the company.
zh:第二代在深度开发，第三代正在成形，量产会在 2027 年爬坡。不过有一道影子：他们的数据中心掌门人 Chris Malone——负责庞大的星际之门项目——刚刚离职。
[Sarah]: Right before a possible IPO, too. Risky timing. But honestly, Mike, the picture I can't shake is much quieter than all this drama.
zh:还赶在可能的 IPO 前夜。这个时机挺危险。不过说实话，Mike，我脑子里挥之不去的画面比这些大戏安静多了。
[Mike]: Which picture?
zh:哪个画面？
[Sarah]: Somewhere in a lab sits a chip named after a snack pepper, resting in trays named after dinner - and it quietly beat the most valuable chip company on Earth.
zh:某个实验室里坐着一颗以佐餐辣椒命名的芯片，躺在以晚餐命名的托盘里——而它安静地击败了地球上最值钱的芯片公司。
[Mike]: And today NVIDIA still wears the crown. But notice something - the challenger didn't queue at the front gate. It's already inside the server room, fans spinning.
zh:今天英伟达仍戴着王冠。但注意一件事——挑战者没有在大门口排队。它已经在机房里面了，风扇正转着呢。
[Sarah]: Spicy times ahead. That's all for today! If you enjoyed this episode, share it with a friend who follows tech news.
zh:接下来的日子够辣的。今天就到这里！如果你喜欢这期节目，分享给关注科技新闻的朋友吧。
[Mike]: See you next time on Learn English with Podcasts!
zh:下期 Learn English with Podcasts 再见！
