新智元报道
这两天，韩国SK海力士甩出了一份能把同行看哭的财报。
一个季度，营业利润约620亿美元，同比暴涨632%，营业利润率80%。相当于每卖一块钱的内存，就有八毛是纯赚的。
让人倒吸凉气的，是它顺手放出的预告——
内存供不应求的局面，得一路延续到2030年；2027和2028两年的产能，一大半已经被长约提前抢光了。
未来这波「内存吞金兽」，就是自动驾驶和机器人。
再往前倒一周，特斯拉三季度财报会上，马斯克也认了这个现实：AI5芯片要到明年4月左右，才能真正装进擎天柱机器人。
为了把内存、封装、逻辑芯片这几道卡脖子的难关彻底打通，特斯拉还要自己建Terafab芯片厂。
光是到2028年初，就得先砸进去几百亿美元；往长远算，总投入甚至可能过千亿。
如今，模型参数规模越来越大，芯片和内存根本跟不上了。
那问题来了：没有特斯拉那么厚的家底，也不打算自己盖晶圆厂，这条路还能走得通吗？
今天，中国一家成立不到三年的芯片公司，交出了它的答案。
XHEART新芯航途发布新一代大模型芯片X7 Ultra，直接冲着十亿参数级智驾大模型去，算力700 TOPS。
在模型表现相当的前提下，X7 Ultra对标行业旗舰芯片，内存带宽消耗降了80%
与此同时，兄弟公司Momenta同台分享R7 Ultra世界模型的最新进展。
而XHEART破局的解法，浓缩起来就一句话：让大模型来定义芯片。
过去的智驾，是一堆小模型拼起来的：感知、预测、规划各管一段，像流水线一样往下传。
信息每过一道手就被压扁一次，天花板也低，小模型能做多好，全看工程师想得多周全。XHEART CEO李宗霖在台上说得直白，「那是人脑驱动的系统，工程师写规则，机器照着执行」。
但Transformer一出来，端到端大模型直接把这套老桌子掀了，智驾真正变成了吃数据的怪兽。
算法换了内核，跑算法的芯片，立刻就撞上了三堵新墙。
- 纸面算力，能不能在真实系统里发挥出来？
算力标多少TOPS，就像是跑车的表底数字：标着能飙300码，早高峰堵在路上照样只能开20。
对大模型来说，堵车的地方是数据搬运：参数和中间结果在内存和计算单元之间来回倒腾，计算单元算得飞快，然后干等数据。
更致命的是精度。关键计算为了提速砍掉精度，车上的决策就会「走样」。用户不在乎纸面数据，只关心车在路上反应快不快、判断及不及时。
- 今天的芯片，跟不跟得上明天的模型？
芯片从定义到上车动辄好几年，模型却半年一个大版本。设计时没预判模型往哪儿长、没留够资源，车卖出去没多久，OTA升级就会撞墙。
- 能不能用合理的能耗，换来需要的能力？
提算力最直接的办法是堆晶体管，可车上的散热永远有一条硬边界，功耗暴涨的方案很难大规模上车。今年内存价格又翻了好几倍，数据在内存里多跑一趟，花的都是真金白银。
现实需要，这三道题得放在一起答。
特斯拉十年前跟Mobileye分手、自研芯片，看透的就是这一层。只是它的芯片从来只给自己用。
行业常规是先做芯片，再让模型去适配。
XHEART的思路是反着来的——先搞清楚模型在车上到底是怎么跑的，再反过来设计芯片。
通用芯片像商场成衣，谁都能穿，谁穿都不太合身。XHEART是量体裁衣，尺寸从大模型身上量，量尺握在兄弟公司Momenta手里：在上百万台量产车上跑着的真模型、真数据。
芯片的架构怎么搭、资源怎么分、怎么干活、往哪迭代，全围着大模型转。
落到X7 Ultra这颗芯片上，XHEART有几刀切得非常准：
第一刀，直接上超大核NPU。
传统做法是一堆小核分头干活，核与核之间传数据，速度在每秒百GB这个级别。
X7 Ultra把它们聚成一个超大核，核内数据交互直接拉到每秒TB级，相当于把乡间小路换成了八车道。再配上大容量、高并发的多级缓存，数据尽量在芯片里面转，少往外面的内存跑，计算等待和能耗都跟着降下来。
第二刀，专用算子加速引擎。
早年的NPU为卷积网络设计，到了Transformer时代，一些计算只能甩给CPU，来回切换还多读写内存。业界的办法是在NPU里加向量处理器。
这一次，X7 Ultra给它们配了专门的硬件加速，原生支持FP16高精度计算，快和准两头都要。
第三刀，解锁NPU硬件并行。
计算单元同时开工，算东西和加载权重同步进行。好比生产线，原料提前送到工位，机器才能一直满负荷转。
第四刀，重构AI工具链。
常规推理每算完一步，就把中间结果写回外部内存，下一步再读回来。
X7 Ultra的工具链把整张网络融合成一个「大算子」，中间结果不再进出外部内存，外部内存只管输入、输出和权重，整网内存零拷贝。
那这些改动换来了什么？
X7 Ultra面向十亿参数级的智驾大模型设计，算力狂飙700 TOPS。单看这个数，不算吓人。
狠的是它的考试规则：跑的是量产智驾大模型，芯片输出的可部署精度要达到「四个九」，也就是99.99%。
在这套考纲下，成绩单非常亮眼：
- 用合作伙伴的Transformer量产智驾模型实测，计算利用率98.7%。标称算力几乎一滴不剩全用上了。
- 在模型表现相当的条件下，对标行业旗舰芯片，内存带宽消耗下降80%。
- 同样条件下，与行业旗舰产品相比，推理能效提升100%。
98.7%答的是第一道题，能效翻倍答的是第三道。至于SK海力士那个「缺到2030」，别人还在排队交内存税，XHEART先给自己打了个两折。
在配置上直接拉满，X7 Ultra用上的Armv9车规级CPU，XHEART一举拿下中国首发量产资格。
激光雷达已经卷到上千线，X7 Ultra专门配了加速器，每秒处理2000万个点。摄像头这边，X7 Ultra专门加了本地计算的多核处理器，实测摄像头处理稳定性比竞品提升100%。
另外，全新芯片功能安全满足ASIL-B，内置MCU拿到最高等级的ASIL-D，信息安全满足国密二级。
以上所有的设计，出发点都一样：先理解大模型，再设计芯片。
第一代X7已经证明，同样的芯片资源，能部署的模型参数量约是通用芯片的10倍。
技术讲得再漂亮，车企最后只问一句：敢不敢把车交给你？
车企选芯片，其实是在赌未来。一颗芯片从定点到上车要好几年，车上市后还要再卖好几年，可模型半年就迭代一轮。
光有前瞻思路还不够，在车圈，信任这东西只能靠真刀真枪攒出来。
先看带兵的人。CEO李宗霖在芯片行业深耕了二十多年，在联发科做到无线通信事业部负责人，后来当过OPPO旗下哲库的COO。
2023年重新扎进创业浪潮时，他和曹旭东聊出两个判断：大模型会改变智驾，也会改变芯片的设计方式；汽车是大模型惠及普通人的起点，芯片得把性能、能效、成本一起做好。
核心团队一半来自头部芯片公司，一半来自一线智驾算法企业。
再看具体打法，主打的就是一个「快准狠」。
Momenta把量产模型提前放进芯片的硅前验证，芯片还没造出来，模型已经在仿真里「试住」过了；制造交给台积电，整车表现跟车企一起磨。
所以X7回片之后，速度快得离谱：不到24小时点亮全芯片，17天完成整车闭环，21天就接受合作伙伴的实车检验。客户坐进车里自己开一圈，比看一百页PPT都管用。
接下来，还要看省不省心。XHEART不当Tier1，但参考设计按主机厂量产标准验证，车企拿去就能开发。其中双X7方案用统一「双大脑」架构，一套硬件兼容L3和L4，关键环节全冗余，已经拿到多个车型定点。
最后拿战绩说话。X7系列量产8个月，累计出货20万颗，搭载10款量产车型，拿下超过40款车型订单。
首发车是上汽大众旗舰SUV ID.ERA 9X，之后是荣威家越07、MG 07，还有凯迪拉克、奇瑞的部分车型。从30万级的豪华旗舰到十几万的国民家用车，全线通吃。
值得一提的是，资本也在用双脚投票。就在前几天，这家独角兽刚敲定新一轮融资，累计将近50亿。
投资人名单里，有欧洲车企雷诺，还有国产内存龙头长鑫科技。一家内存厂主动入股一家大模型芯片公司，这步棋，值得细品。
这一次大会上，「兄弟公司」Momenta带来了下一代R7 Ultra世界模型。
参数量比上一代涨了5到10倍，安全性能提升5到10倍以上，「推演万物」的能力开始涌现。
X7 Ultra托着R7 Ultra，双方一起搭了个「世界模型基座平台」：Momenta把模型往哪长提前告诉芯片，XHEART按需求调架构；车端算力的实际条件，又反过来喂给模型研发。真实道路上跑出来的经验，再推着两边一起往前走。
Momenta CEO曹旭东曾提出「智驾摩尔定律」——智驾的安全性和性能，每两年提升十倍。而大模型芯片，正在加速「智驾摩尔定律」，进一步提升至每年10倍。
这也让他做出了新的判断，2028年前后，L4便会迎来行业引爆点，20230有望进入规模化落地。
这条定律想成立，模型得一直长，芯片也得一直跟得上。
这边，XHEART按「量产一代、开发一代、预研一代」的节奏往前推，下一代X9已经在研发，瞄准的就是更高要求的L4自动驾驶和具身智能。
特斯拉证明了软硬一体是最硬的护城河，但它关起门自己用，还得自己盖厂扛账单。Momenta和XHEART把模型、数据、芯片打通，再把这套软硬一体能力开放给全球车企。
马斯克为这件事准备了十年，还在砸钱。
XHEART这边，用李宗霖的话说，从第一个提出「大模型定义芯片」，到「第一个」让大模型芯片大规模量产，不到三年。
三道题的答卷已经交了。接下来打分的，是每天开着这些车上下班的普通人。
编辑：桃子
不喜欢
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
New Ji-won.
In the last two days, South Korea's SK Hercules has thrown out a financial report that allows his colleagues to cry.
In one quarter, operating profits amounted to about $62 billion, an increase of 632 per cent over the same period and a profit margin of 80 per cent. The equivalent of 80 cents for every dollar of memory.
And those who breathe the cooler, it is the warning which it sends forth.
The gap between the supply of and demand for memory continues until 2030; most of the capacity in the years 2027 and 2028 has been looted in advance.
In the future, this wave of "internal gold swallowers" is autopilot and robotics.
A week later, at Tesla's three-quarter press conference, Mask recognized the reality that the A5 chip will not be actually installed in the Optimus robot until around April next year.
Tesla had to build its own Terrafab chip plant in order to get the memory, seal and logical chips through the neck.
By early 2028 alone, tens of billions of dollars will have to go into it; in the long run, the total investment may even be over a trillion.
Today, model parameters are growing in size, and chips and memory are simply not up to date.
The question arises: can this road work without Tesla's wealth and his plans to build a crystal circle?
Today, a Chinese chip company that has been in existence for less than three years has given its answer.
XHEART's new core route publish, a new generation of LLM chips X7 Ultra, is heading straight for LLM, a billion parameter class, to calculate 700 TOPS.
With comparable model performance, X7 Ultra has lost 80% memory bandwidth consumption to flagship chips in the flag industry
Meanwhile, Brother Momenta shares the latest developments of the R7 Ultra World Model.
And the solution to the XHEART break-up is concentrated in one sentence: let LLM define the chip.
The old smart drive was a bunch of small models: sense, prediction, planning of the tube, going down like a current line.
The information is crushed once in a hand, the ceiling is low, the small models are good, and the engineers think well. XHEART CEO Lee Zhongzheng said, "It's a brain-driven system, engineers write rules, machines follow."
But when Transformer came out, LLM just blew up this old table, and Jiji became a data-eating monster.
The algorithm changed the kernel, ran into three new walls.
- It's paper power. Can it work in the real system?
To calculate how much TOPS is, it's like the bottom figure of a sports car: it's a 300-yard mark, and early peaks are stuck on the road and can only drive 20.
For LLM, traffic is blocked by data handling: parameters and intermediate results are inverted between the memory and computing units, which are counted fast and then dry.
More lethal is precision. The key calculation is that, in order to cut off precision at a speed, the decisions made in the car will be "disturbed". Users do not care about paper data, but only about the speed, judgement and timeliness of the vehicle on the road.
- Today's chip, can you follow tomorrow's model?
The chip moves from definition to the car for several years, and the model is a large version of six months. Unpredicted models were designed for long distances and without sufficient resources, and soon after the car was sold, the OTA upgrades hit the wall.
- Can we use rational energy to replace the capabilities we need?
The most direct method of calculating power is through stacking transistor tubes, but there is always a hard boundary in the heat of the vehicle, and it is difficult to get on a large-scale train in a programme with a high capacity. This year's memory price has increased several times, and the data goes on a trip to the memory, all with real money and silver.
Realism requires that these three questions be answered together.
Tesla broke up with Mobilee ten years ago, studied his own chip, and that's the floor. It's just that it's never used for itself.
Industry usually makes chips before models fit.
XHEART's idea is the opposite -- first figure out how the model runs in the car, then design the chip.
The generic chip is like a mall dress, and it doesn't fit anyone. XHEART is tailored to the size of LLM and held in the hands of Brother Momenta: real models, real data running in millions of production cars.
How the chip is structured, how the resources are divided, how the work is done, where it goes, all around LLM.
On this chip, X7 Ultra, XHEART had a few cuts that were very accurate:
First cut, straight to the super-nuclear NPU.
The traditional practice is to work in small nuclear fractions, passing data between nuclear and nuclear at 100 GB per second.
X7 Ultra brings them together into a super-nucleus, where the interactivity of the core data is carried directly to the TB level per second, which is equivalent to the replacement of the country road by eight lanes. The data are then accompanied by large-capacity, high-quality, multi-scale cachees, with as little data as possible in the chip, less running to the outside memory, and counting the waiting and energy consumption down.
Second cut, special calculator acceleration engine.
Early NPUs were designed for a volume network, and by the time Transformer era, some calculations could only be dumped on CPUs, with multiple reading and writing memory. The industry approach is the NPU Liga vector processor.
On this occasion, X7 Ultra provided them with specialized hardware acceleration, with natives supporting high-precision FP16 calculations, both fast and accurate.
Third knife, unlock NPU hardware in parallel.
The calculation unit starts at the same time, counting and load weights are synchronized. It is like a production line, where the raw materials are delivered to the posts in advance so that the machine can remain fully loaded.
Number four, reset the AI tool chain.
Once the normal reasoning is completed, the intermediate result is returned to the external memory and next step is read back.
The X7 Ultra tool chain integrates the entire network into a "big counter" with the intermediate result that no longer enters into or exits the external memory, except for input, output and weight, and completes the network with zero copies.
So what's the change?
X7 Ultra designed for a billion-scale smart driving LLM, a force-run 700 TOPS. That's not scary.
It's the rules of its exams: it's a mass drive of LLM, and the chip output has a deployable accuracy of "four nines," or 99.99%.
Under this examination, the report card is very bright:
- 98.7 per cent utilization factor was calculated using partners ' knowledge of the Transformer-Quantities Driving Model. It's almost empty.
- A reduction of 80 per cent in the consumption of memory bandwidth for flagship chips in the flag industry under conditions of comparable model performance.
- Under the same conditions, the reasoning is 100 per cent more efficient than the industry ' s flagship products.
98.7 per cent answered the first question and the third doubled the energy efficiency. As for SK Hercules, who was "defunct at 2030" and who was still queuing to pay the memory tax, XHEART gave himself a discount.
The X7 Ultra X-Ultra vehicle code Armv9 CPU and XHEART each took China ' s initial production qualification.
Laser radars have reached thousands of lines and X7 Ultra is specially equipped with accelerators to process 20 million points per second. On this side of the camera, X7 Ultra was specially equipped with a locally calculated polynuclear processor, which increased the stability of the actual camera processing by 100 per cent over the competition.
In addition, the new chip is fully functionally compliant with ASIL-B, the built-in MMU is at the highest level of ASIL-D, and the information security is at the level II.
All the above design, starting from the same point: understand LLM, then design the chip.
The first generation of X7 has proven that the same chip resource, the model parameter count, can be deployed approximately 10 times the generic chip.
The technology is so beautiful, the only thing the car company wants to ask is, "Will you give it to me?"
Cars pick chips, actually, in the future. A chip takes several years from the point to the car, after which it will be sold for several more years, but the model will be repeated in six months.
It is not enough to be forward-looking, and in the circles of cars, trust in this thing can only come from real guns.
The CEO Lee Jong-ho has been working in the chip industry for more than 20 years, has been the head of the wireless communications department in UNDGC, and has since become the Coo of Chukaku under OPPO.
In 2023, when he and Cao Xudong resumed the wave of entrepreneurship, he and he made two judgements: LLM will change the drive and the design of the chip; the car is the starting point for LLM to benefit ordinary people, and the chip has to be done with performance, energy efficiency and cost.
Half of the core team came from head chip companies and half from front-line smart drive algorithm companies.
And if you look at it, it's a "quick and hard" hit.
Momenta pre-positioned the mass model to the silicon, which was not produced, and the model was "tried" in the simulation; it was delivered to the station, and the whole car was polished with the car.
So after the X7 return, the speed is out: less than 24 hours to light up the whole chip, 17 days to close the whole car and 21 days to undergo physical inspection by partners. The client sits in the car and drives himself in a circle.
And then we'll have to look at it. XHEART is not a good Tier1, but the reference design is validated according to the host plant volume standards and the car will be developed. The two-X7 programme uses a unified "brain" structure, a set of hardware compatible with L3 and L4, with redundancies in key links and multiple model locations.
I'll end up talking about the score. The X7 series produced a total of 200,000 goods over a period of eight months, carrying 10 volume models and taking over 40 car-type orders.
The first car was SUV ID.ERA 9X, followed by Vongwei ' s 07, MG 07, and some of the Cadillac and Chire models. From the 300,000 class flagship ship to the 100,000 national cars, it's available online.
It is worth mentioning that capital is also voting with two feet. Just a few days ago, the unicorn just finalized a new round of financing, totalling almost $5 billion.
The list of investors consists of European automobile companies Reno, and the country's domestic memory is a long-standing technology. A LLM chip company, in which a memory plant actively shares, is worth a fine piece.
At this conference, Brother's Momenta brought the next generation of R7 Ultra World Models.
parameter count is 5 to 10 times higher than the previous generation, safety is 5 to 10 times higher, and the ability to "do everything" is beginning to emerge.
X7 Ultra holds R7 Ultra, and the two parties together set up a "World Model Base Platform": Momenta informed the chip of how long the model went, XHEART was structured according to demand; and the actual conditions of the end-of-pipe computing, which in turn fed the model development. The experience of running out of the real road, pushing both sides together.
Momenta Ceau Xudong has proposed a tenfold increase in the security and performance of smart driving every two years. The LLM chip, however, is accelerating to 10 times more per year.
This also led him to make a new judgement that, around 2028, L4 would meet the industrial detonating point and 20230 was expected to enter the scaleed landing site.
It's a law that needs to be in place, model long and chip up to date.
Here, the next generation, XHEART, is now being developed, aiming at a more demanding L4 self-piloting and intelligence.
Tesla proved that hard and soft were the strongest moat, but it closed its doors and had to pay for it. Momenta and XHEART have been able to access models, data, chips and then open this soft and hard-core capability to global companies.
Mask prepared for this for 10 years and is still smashing money.
On this side of XHEART, in the words of Lee Jong-hoon, from the first to the first to "LLM" definition chip to "first" to "first" large-scale production of LLM chip, less than three years later.
Three responses were received. The next score is the average person who leaves work in these cars every day.
Edit: Peaches
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.