第一部分 为什么电压必须往上走
01 开场：从机架讲到拓扑
谢谢。这个安排正好——Jared 讲的是机架层面，也讲了托盘层面；我讲的是拓扑和器件层面。所以顺序是合适的：我们从高层一路讲到低层。
我今天要讲的是 800 V 数据中心的 DC-DC 设计考量、一些拓扑比较，以及一些权衡。
[applause] Yeah, thank you. So, this is perfect. Jared talked about it at the rack level, and he talked about it at the tray level. I'll talk about it at the topology and component level. So, it's perfect — we're going high-level to low-level here. So, I'm going to talk about 800-V data center DC-to-DC design considerations, some topology comparisons, and some tradeoffs.
小结 这场演讲的定位很清楚：前面几场把 800 V 讲到机架和托盘，这一场往下钻到拓扑与器件——也就是「电到底怎么从 800 V 变成 1 V」。
02 一句论文里的话，用了二十年
我们先过一遍数据中心不断变化的功率需求。计算机性能是指数级增长的，功率需求也随之上升，而且这种趋势预计会继续下去。
这句话我二十年前的硕士论文里就写过——所以今天还在这儿。二十年前那句话写得完美，直接复制粘贴就行，每五年我都能重复一遍。
二十年前，我们根本不知道该怎么做到 200 W。你怎么给一颗 200 W 的处理器供电？在当时简直闻所未闻。
那就快进二十年。现在是什么情况？这是去年 OCP 上的内容。现在我们在谈兆瓦级的机架。二十年时间，我们从 200 W 级的处理器功率，走到了 2 kW 级。
我们是怎么走到这里的？或者说，我们已经走到这里了——这就是现实。
So, let's first go over evolving data center power demands. The computer has seen exponential growth in performance, and the power requirements have increased. These demands are expected to continue. I wrote that in my master's thesis 20 years ago. So, here we are. Twenty years ago, it was perfect — I can just copy and paste that. Every five years I can repeat that statement. So, 20 years ago, we didn't know how on Earth we were going to get to 200 W. How are you going to power a 200-W processor? It was just unheard of at the time. So, let's fast forward 20 years. Where are we now? This is from OCP last year. Now, we're talking about megawatt-type racks. So, in 20 years, we go from talking around 200-W processing power to around 2 kW processing power. So, how are we going to get there? Or how have we gotten there? This is reality now.
小结 同一句话在论文里放了二十年仍然成立，这本身就是问题所在：功率增长不是线性叠加，而是每五年换一个数量级。
03 从 50 V 母排到 800 V 母排
所以配电方式也在演进：从 12 V 背板起步，走到 50 V 背板。
左边是你今天看到的常规电源机架——交流转直流的电源托盘、PSU，就像上一场 Andy 讲的那样；接到 50 V 母排，再送进 IT 托盘。
底部你能看到 50 V 进来，经过热插拔、经过 IBC。IBC 把 50 V 降到 12 V 或 6 V 左右，然后是最后一公里：VR 和负载点把电压降到亚伏级，送给计算芯片。
那么，怎么走到兆瓦级机架？答案是提高电压。电压提高 16 倍，电流就能降到 1/16——这就是我们能去掉、能缩小那些粗大母排的办法。
你可以看到，800 V 这类架构有很多种实现方式，我只讲 IT 机架内部这一段。将来真正上到 800 V 母排的时候，我们再谈下一轮迭代。
那时候 IT 托盘长什么样？托盘内部的功率变换长什么样？你会发现非常相似：800 V 进来，有热插拔，有高压 IBC，然后可能还有一级低压 IBC——取决于你是走 800 转 50，还是考虑直接 800 转 12、800 转 6。然后再降到 12 V 和 6 V，最后降到亚伏级，送给处理器。
So, we've evolved the power distribution method. Started with 12-V backplane, moved to 50-V backplane. On the left is your conventional power rack as you see today. You've got your AC-to-DC power shelves, your PSU, like Andy talked about in the last presentation. Goes to your 50-V bus bar, feeds your IT tray. And then, on the bottom you can see you've got 50 V that comes in. I've got my hot swap, my IBC. My IBC will take us from 50 V to somewhere 12 or 6 generally. Then we have our last mile: our VRs and our point of loads are going to take it to our sub-volt compute chips. So, how are we going to get to these megawatt-type racks? We're going to go up in voltage. So, if you go up 16 times in voltage, you can go down 16 times in current. That's how we can remove and reduce the size of these massive bus bars. So, you can see there's a lot of different implementations of this 800-V type architecture. I'm just going to talk about inside the IT rack. So, we'll talk about a future iteration when we get to that 800-V bus bar. And what does the IT tray look like? What does the power conversion look like inside that IT tray? So, you can see it's very similar. 800 V comes in. I've got a hot swap. I've got a high-voltage IBC. Then I've got a low-voltage IBC, maybe. It depends on whether you do 800 to 50, or there's considerations to go directly 800 to 12, 800 to 6. And then I'm down to my 12 and 6 V again. And then that'll take us down to the sub-volt, down to the processor.
小结 整张图只有一个动作：把电压抬高、把电流压下去，母排才能从「铜的重量」里解放出来；托盘内部的级数其实没变，只是最前面多了一级高压变换。
第二部分 从 800 V 到 1 V 的四条走法
04 加一级，还是并两级
那么，最好的路径是什么？
从德州仪器的角度看——我们做半导体、做这类应用的芯片，所以我们想知道哪条路最好，力气该往哪儿使。
看这张托盘内部的框图：功率分配板（PDB），上面有热插拔、有遥测，然后是隔离屏障、加强绝缘——也就是安全。怎么把高压送进低压的 50 V 主板？
方案一很简单：加一级。加一个高压中间总线变换器，把 800 V 接进来，后面就能沿用相对常规的 50 V 主板设计。低压 IBC 做 4:1 降到 12 V，或 8:1 降到 6 V 左右，然后用常规的多相稳压器。
方案二：当我们开始追密度的时候——显然没人希望板子上多放功率器件。功率不赚钱，GPU 才赚钱。板上做功率变换的部分越少，能放上去挣钱的 GPU 就越多。所以想法是：为什么不把两级并成一级？方案二就是 800 V 转 12 V 的变换器。
方案三则是 800 V 转 6 V。这是我们考虑的三种方案。
再往后是方案四——直接做到集成式 IVR 之类的东西，那个更前瞻，我把它叫作方案四。
这些就是从 800 V 走到 1 V 的几条路。
So, what's the best way to get there? Looking at Texas Instruments — we make semiconductors, we make chips for these applications. So, we want to know what's the best way to get there and where should we focus our efforts to make the best products. So, looking at this block diagram of inside of a tray, like we talked about, we've got the power distribution board. Inside of that, you've got a hot swap, you've got some telemetry, and then you've got your isolation barrier. You've got your reinforced insulation, your safety. How do I get that high voltage into the low-voltage 50-V motherboard? So, option one is essentially: let's just add a stage. Let's add a high-voltage intermediate bus converter, takes this 800, bring that in, and then you can stick with a somewhat conventional 50-V motherboard design. I can have a low-voltage IBC with a 4:1 step down to 12 V or an 8:1 to get us around six, and then I can use my conventional multi-phase voltage regulators. And then let's talk about option two. When we start to try to push density, obviously, no one wants more power on the board. Power costs money, GPUs make money. So, the less power conversion you have on the board, the more revenue-generating GPUs you can have. So, the idea is maybe now, why not take two stages and make them a single stage? So, option two, I'll say, is an 800-V to 12-V converter. And then option three will be 800 to six. So, those are the three options that we'll consider. And this is going to build us into the future. This is option four, when we talk about do we do something like take 800 and then derive directly to an integrated IVR or something like that. That's more forward-looking, but I'll call that option four. So, those are the ways that you can get from 800 to 1 V.
小结 四种走法本质是同一道算术题：级数少一点、板上功率器件少一点，就能多放一颗 GPU；代价是每一级的难度都往上跳一级。
05 横向送电与垂直送电
我们来看看。看框图很容易，但不能只看单级效率——还要看板上配电的损耗，那里面损耗很多。
大致有两种做法。
一种是横向送电（LPD），非常直接：功率横着走过主板。800 V 进来，穿过 PDB，变成 50 V，然后横着走到板子正面的稳压器，再由它们给负载供电。
另一种是垂直送电（VPD）。你最终会问：如果横着走损耗太大、阻抗太高，为什么不把最后一公里——最后那一级稳压器——放到板子背面去？直接贴在芯片正下方驱动，电阻会显著降低。
这也给了你一些新的选择：它改变了 IBC 该放在哪里。因为我们谈的是极高的功率，电流也极高，尤其是在我们努力把电压往下压的时候。
上面那张图是一个四 GPU 托盘的示意：800 V 甩在侧面，让高压远离敏感的处理器，然后接到低压 IBC，再横向送到 12 V 或 6 V 的负载点，横着给 GPU 供电。
但一旦上了 VPD，就没有这最后一公里了——板子正面不再有那些低压稳压器，它们跑到背面去了。于是我可以把 IBC 挪得更近。
核心思路是：在你抵达负载之前，尽可能把电流降下来。
再说 800 转 12 V。德州仪器在低压处理器这块做得很多。我们的看法是：12 V 目前可以视为横向送电方案；而到了 6 V，就可以开始考虑垂直送电。
所以终局图景是：最理想的配电是直接从 800 V 到 6 V，然后紧贴 GPU 放置，由它驱动垂直送电，整体效率最高。
但这也是最难的挑战。首先我得做到极高的电压、极高的效率、极高的密度；你很可能还要和芯片共用同一块冷板；而且高压离 GPU、CPU 更近了。这本身就是难题。
这些就是可选项。下面我给大家看我们实际测出来的结果。
So, let's look at — it's easy to look at a block diagram, but let's talk about not only the stage efficiency. You've got to look at the power distribution on the board. There's a lot of losses in there. So, when you look at it, there's two general ways that you can approach this. You have lateral power delivery, and that's very straightforward. The power delivers laterally across the motherboard. Comes in at 800, goes in through the PDB, comes out, gets to your 50, and then that'll travel laterally to your voltage regulators, which are on the top side of the board, and then they power your load. The other option is vertical power delivery. Ultimately you say, well, if I have too much loss, too much impedance traversing that power laterally, why can't we put the last mile, that last voltage regulator, on the back side of the board? Significantly reduce the resistance by just driving directly underneath it. And that also allows you some options. It changes where you're going to locate your IBCs, because when we're talking about these very, very high powers, they have very high currents, especially as we try to reduce the voltage. So, you can see here at the top: this is a cartoon drawing of a four-GPU type tray. You have the 800 kind of out to the side, keep your high voltage away from your sensitive processor, and then it'll go to your low-voltage IBC. That'll go across to your 12-V point-of-loads or 6-V point-of-loads. It'll laterally power the GPU. But when you go to VPD, you don't have that last mile. You don't have those low-voltage voltage regulators on the top side of the board anymore. They're underneath. So, then I can move my IBCs closer. The idea is that you want to be able to reduce the amount of current required until you get directly to your load. And then, when we look at like 800 to 12: generally, TI works a lot in low-voltage voltage processors. We say that 12 volts right now we kind of see that as a lateral power delivery type solution. And then we kind of see at 6 volts you can start to look at vertical power delivery. So the envisionment of the end goal: the perfect power distribution is to go straight from 800 to 6. And then it's located very close to your GPU. It can derive the vertical power delivery. You get the highest overall efficiency. But it's the hardest challenge. Now I've got to get very high voltage, very high efficiency, very high density. You're probably going to be sharing the same cooling plate. And then also my high voltage is coming closer to the GPU, CPU. So that's a challenge in itself. But these are the options, and we'll show you what we found when we've looked at these.
小结 垂直送电省掉的是「最后一公里」的铜，代价是把开关电源挪到芯片背面——高度、频率、散热三件事同时变难，这也是为什么它效率更低却可能更划算。
（转自公众号：一超一世界）
