Remember when GPUs were the dominant tech to mine crypto, then ASICs and other spepcialized purpose-built hardware came out and made GPUs irrelevant for mining, wont that also happpen in regards to ai training/inference. like aws is attemppting to do with their trainium/inferentia chips. Sure they suck now, but won't they get better and thus lose the dependency on GPUs?
It's already happened hasn't it. Google's TPUs are apparently much more efficient at inference workloads. However the hardware isn't available to the public like Nvidia chips are.
I'm unsure about the energy efficiency of Tenstorrent cards, but they are at the moment much more cost effective than Nvidia. But software support is far behind, I suspect if it wasn't the cost would be more equal.
Modern NVIDIA cards are essentially TPUs - especially those spec'd out for AI data centres. TPUs aren't anything particularly special - they are just large vector processors, with hardware specialised for the data + instruction fetch patterns that are seen when running AI models. The graphics cards that NVIDIA sells for data centres are just large vector processors with hardware specialised for the data + instruction fetch patterns that are seen when running AI models.
Google has two different types of TPUs now. One for inference and one for training.
Do we actually know they're more efficient per watt or for the power used and heat produced ? Or are they just easier to scale (and manufacturer) en masse ?
There's also the CUDA moat which is hard to get around
Cuda is a moat for traditional data science, where you have a torch model and you want to easily run it on a GPU. CUDA is the happy path for doing this.
Nvidia's moat for LLMs isn't Cuda, it's installed capacity, arrangements with frontier labs, TSMC bandwith, and just the fact that it currently has no real competition.
Various articles recently claimed most of the NVIDIA deals are "handshakes with Jensen".
This Huang scandal has so many dimensions that are clearly more impactful on the world outside than to NVIDIA itself -- the blast radius seems to extend to the US cabinet -- that it begs questions if Jensen Huang wasn't id'd as the essential piece to remove to affect a comprehensive restructing in the segment.
So if deals fall apart because Jensen and his handshake have left the room, that may unclog many memory and processing supply lines that are currently spoken for by deals with Jensen. That may have been a goal from the start.
So the GPUs were a cherry on top, the real target was removing Huang? And Jensen huang'd himself by making himself the brittle link in a chain (his handshake deals being a big one).
CUDA is useful if you want a human expert writing new experimental kernels, perhaps while testing different algorithms.
But once you do that, translating CUDA into one of many available vectorized backends is trivial.
We have a bunch of custom kernels for geometric operations (like merging polygons) and they were rather difficult to write a while ago. AI rewrote them in HIP during my lunch break.
Basically capacity issues. NVidia + Apple have bought up a significant portion of TSMC's manufacturing capacity. The primary supplier for Google's TPUs is Broadcom, with TSMC as a second source, and they're trying to ink contracts with other suppliers like MediaTek, but there're only so many semiconductor fabs in the world and TPUs need a fairly modern process.
Within Google itself the TPUs are definitely preferred for AI workloads, you just can't always get them.
> The primary supplier for Google's TPUs is Broadcom, with TSMC as a second source
You've got it all mixed up.
Google designs the TPU in partnership with Broadcom. Many companies will write the Verilog RTL and then send this to another company like Broadcom to do the physical design and turn the Verilog along with lots of other IP blocks like serdes into GDS mask data.
Broadcom does not have any fabs. They send that GDS mask data to TSMC to manufacture the chip just like Nvidia, Apple, and tons of other companies.
I'm a physical design engineer if you have any questions.
Broadcom isn't a chip foundry so calling it Google's supplier doesn't make sense. TSMC has been the primary manufacturer of TPUs but Google's hoping to use Samsun and Intel also going forward (and these are the only leading edge foundries in the world). Broadcom is co-developer of the chip.
I don't know details at this level but I know that TPUs like GPUs need to be made on the most advanced available nodes and TSMC, Samsung and Intel are exactly the only enterprises pursuing those advanced nodes. The reason is that this is
I mean, Google tell me it's TSMC's 3 and 2 mm process - exactly what you'd expect. But you can Google as well as I can.
Presenting broadcom as an alternative to tsmc is like wet roads cause rain. The average knowledge of computer hardware development cycle in hn comment sections is frustratingly low.
Crazy that a company can use a contract to buy up all the supply to build an inferior product that’s less energy efficient and then sell it at a higher price than the purpose built counterpart. Market failure.
They're not really the same markets. NVidia is selling chips that AI runs on, Google is selling services that use AI. Google's play is to make better services than the competition can because they control the full infrastructure stack, thus preventing their markets from being commoditized. NVidia's play is to commoditize the AI providers so that the number of services expands rapidly, with NVidia chips powering all of them. Arguably it's Google's strategy that is the market failure, as it's designed to keep competition out, but...
Google's used this strategy a few times before, and it frequently doesn't work out. They kept MapReduce and GFS proprietary so they could build better big-data services than everyone else, but Doug Cutting from Yahoo clean-roomed it, and then eventually you had interviewees going "MapReduce? That's like Hadoop, right?" They had a long string of internal ML frameworks well before the rest of the industry (I remember in the '00s and early '10s Google was by far the leader in large-scale machine-learning) but had to open-source TensorFlow in the face of external competition, and then lost in the marketplace to PyTorch because TensorFlow was filled with internal WTFs and footguns needed to support specific Google projects. They had a big lead in datacenter technology over everyone else in the industry in the early 2010s, but kept it internal and proprietary until they were forced to sell it as a service by the huge growth of AWS and Azure, then had to play catch-up as a distant 3rd because they lacked the customer relationships and developer mindshare.
Both are fabless customers of TSMC, which is the choke point and center of the high-efficiency chip market.
Unlike gp, I do not believe its a market failures as much as the cutting-edge process node being a natural monopoly. Once TSMC processes got really good, TSMC started capturing most of revenue dollars and reinvesting them, which made TSMC products even better - a positive feedback loop that makes it impossible for Samsung or Texas Instruments to catch up. Intel rode on a similar wave for decades
Packaging is the choke point in AI. It's still TSMC but chip production is doing fine. It's the Chip-on-Wafer-on-Substrate build chain that is severly constrained.
Why is no one going to Intel then for this kind of stuff ? Or even a legacy producer like IBM or Ti ? I would imagine they could buy some slightly older ASML litho machines and make something on a slightly older process (or package it using the older process for the other stuff and just keep the main processor latest gen) ?
> Why is no one going to Intel then for this kind of stuff ? Or even a legacy producer like IBM or Ti ? I would imagine they could buy some slightly older ASML litho machines and make something on a slightly older process
Literally TI even uses TSMC for the digital only stuff they do. The investment for these cutting edge nodes (or even, not so cutting edge. Like say, 10nm) is absurd vs the payoff.
From what I’ve read it’s that TSMC is just so much better at it… much higher yields and much more reliable than any of the other fabs. I wanna say something like 2-3 years ago, before Blackwell, NVidia tried to use Samsung for example but had nothing but issues. Intel hasn’t proven its fabs capable of anything really useful in nearly a decade.
I think even if you have the capital to acquire something from ASML… by the time it’s setup and tuned correctly TSMC will likely have enough capacity to eat your lunch. If it takes 2-3 years to get a fab online TSMC will likely have a new node and their old ones will be available.
TSMC has really proven over the past 5-10 years that the fab is its own beast and requires its own immense level of expertise and investment. When you have the worlds largest chip companies shoveling all their manufacturing dollars through one single org, that seemingly does the right thing, it’s going to be pretty much impossible to beat them.
BigTable and MapReduce are differennt use-cases. The former is a database; the latter is a framework for running batch jobs. The (internal) successor to MapReduce is Flume (2010), which is quite a bit more capable, but didn't come out until Hadoop was already well-established. The external equivalent is Apache Spark (2009), but the gap between internal and external technologies had already shrunk to be nearly indistinguishable by then.
> Crazy that a company can use a contract to buy up all the supply to build an inferior product that’s less energy efficient and then sell it at a higher price than the purpose built counterpart. Market failure.
They're not buying up "all the supply to build an inferior product". They're buying it up build a different, high-demand product.
TPUs are a systolic array. It's not comparable to the hardware architecture that backstops CUDA, nor do they serve the same markets.
The writing was on the wall for TSMC ever since the 5nm node; Apple proved that anyone with enough liquid cash could block competitors from fab access.
Modern NVIDIA GPUs contain tensor cores that are optimised for MMA operations as well though right? The standard CUDA execution path isn’t comparable, but the tensor core path very much is?
I'm no expert but I think we're dealing with two different ends of the spectrum here. Nvidia GPUs are likely cutting edge within the individual board / processor itself. Their main limiting factor right now seems in effectively networking them together (which is why they bought Mellanox or whatever awhile back to further develop Infiniband).
TPUs are significantly less flexible and I would guess at an individual level not as efficient or fast as Nvidia. But they're designed specifically to be bundled together in large arrays with similarly large data pipes to storage and interconnects to cache, ram and other compute. I would guess Google didn't bother trying to sell them until recently because you needed to be deploying them at large scale for it to be even worth also developing the software for.
What we really need is innovation in super fast (bandwidth and latency) and reliable interconnects. I suspect the days of slotted in GPU and RAM will be over soon and we'll see a move everywhere to pin based sockets (like we've always had for CPUs). Stuff like SXM and LPCAMM.
We've started to vertically stack ram out of necessity but have yet to really move on from the 2D plane for the motherboard itself. Slots allow more devices to connect to one flat board. But if the board itself had more surface area (by being maybe an L shape or box) you could slot in a lot more chips physically
It's logical that company like TSMC sells capacity to the highest bidder.
The main thing is that chip manufacturing is the most highly capitalized industry in the world. Each manufacturing step has gargantuan costs and so no company wants to manage (and pay for) all the steps - so TSMC buys chip manufacturing equipment from ASML and sets up a system where corporations buy capacity. It's essentially a very, very, very expensive print shop. The print shop doesn't make its own printers nor it print a newspaper.
And naturally, with very, very objects like leading-edge chips, "inferior" or "superior" is a ridiculous simplification. Nvidia chips are incredibly powerful and useful. Nvidia has been more committed than any other company to general purpose GPU programming.
But Nvidia also is doing it's best to sell at high, monopoly prices. That's why they buy a lot of capacity. The thing about the situation is these companies did spend a lot of money reaching this monopoly situation so from their viewpoint, this monopoly pricing might be fair. But I personally wouldn't care and hope some regulation stops it. But that kind of regulation would still have to get these companies significant profit on these investments, so they'd keep making them.
If I come to a coffee supplier and say I can pay 50$ for a sack of coffee but I don't know how many I can buy yet as I am building up my coffee shops around the country and then Starbucks comes in and says we are willing to pay 55$ and pay ALL your capacity guaranteed, today, just sign here.
Where do you think most coffee sellers would sign?
Interesting. What manufacturing process is Broadcom producing them at ? I'm guessing they're not newly as small (lithography wise) as the best of the best yet but probably good enough
Broadcom is just a chip designer or “mercenary” as I read the other day. TSMC is still the fab for most of the TPU, with Mediatek, and Intel used for auxiliary/supporting components.
google doesn't use them, they buy them to make them available for google cloud customers that require CUDA or other nvidia-only workloads. All training and inference at Google is done on TPUs.
Are they modifiable after creation? If not that seems dangerously rigid until we reach a much more stable place with LLMs. It could be great for well worn architectures and models but building one now for LLMs seems like it'd be foolish when it'll be outdated in a few months.
Why would it be dangerous? It'd effectively be a stamped model at some point in time. The bleeding edge moves on every few months, but even now it seems to me that there's only a tenuous value proposition that's progressing for me. I'd rather have an instantaneous GPT 6, Opus 5.3, or Gemini Hydrogen that I own, rather than always leasing models I have no practical application for.
Dangerously rigid, or delightfully stable? The latest consumer models from claude and openAI are already good enough for so much, if they could run locally with minimal power usage.
The better question is, would someone be able to turn a profit offering a deepseek chip for a few hundred dollars? I wouldn't even care about the energy draw (within reason) having that running entirely locally at better than a few hundred TPS without going bankrupt to obtain VRAM would be a game changer.
You have basically managed to sum up why NVDA keeps dominating, no on can "just buy a few TPUs" and that has been the case since the first TPUs were announced years ago (yes yes GPU shortages and so on, but no one questions general availability of GPUs), and the current Tenstorrent model support page looks like an invitation to beta test HW that might (or might not) be able to run current and future models at some point.
But the next gen of Tenstorrent HW looks great on paper - problem is, I could have written that before the current gen was released.
All of the allegedly competing HW looks like that and has looked like that for years, either a lack of availability, capability, or both.
All of the hyperscale data center companies have their own AI chips and their own CPUs as well. They don't want to be dependent on Nvidia (or AMD or Intel)
They still buy Nvidia and x86 chips but this is what the WSJ article from last week was saying. The hyperscalers are moving away and Nvidia is committed to TSMC wafer orders so they are funding smaller companies to buy Nvidia chips in exchange for equity.
There are also the opposing architectural directions to extend toward:
one in the "Cerebras" direction of ever-larger scale general purpose HPC silicon running inference workloads. Cerebras may have skipped ahead to the last act of that scaling story, notwithstanding any further wafer-diameter enlargements past 300mm. Whether the ending is happy or not is still unclear. Lot of iterative room to grow into it further and they have a formidable patent moat waiting for conventional die-sized AI semi vendors when they inevitably run up against limitations of single-die designs.
the other is in the "Taalas" direction of freezing a specific model into something like high speed ROM / FPGA and getting orders of magnitude inference speed improvements - for that specific model. Like Cerebras, Taalas may also have jumped directly to the architectural telos of that philosophy, with only iterative improvements in cost per work unit left to grind, though like CBRS quite a lot of grinding left to do especially in getting cost per custom tapeout down to near-disposable (hopefully, recyclable) levels.
To be fair GPUs have already changed a lot in response to the AI demands. A lot of silicon these days are tensor cores which have very few applications outside of deep learning.
It still won't make normal GPUs irrelevant like they did with crypto, there will always be demand for training and SOTA models while there is no demand for GPU mining
> there will always be demand for training and SOTA model
Some, but how much? Once you have a mass produced ASIC with a model that's "good enough" much (but obviously not all) of the demand for the cutting edge could evaporate quite rapidly I think.
Something that could benefit from throughput/Wh but wouldn't benefit from flexibility and being cheaper than using bespoke chips feels like such a narrow set of uses though
Inference, yes, see https://taalas.com/ and https://chatjimmy.ai/ As expected, efficiency boost is extreme. But you've got to pick the model you want to enshrine in silicon months or maybe years before you actually get that silicon (and maybe there are additional difficulties if the model is large. Maybe in a few years everybody will just want to use the same four or five models, with no tectonic shifts on the horizon. Then it'll be a no-brainer to make ASICs for those models.
Training? I'm not so sure. You need something flexible so that it can be trained. Burning the weights into the board is the opposite of that.
I read into Taalas when they first were putting out info and they also are wiring up massive CPU cache to the thing. It's not just the burned in models.
As much as I love the idea of Taalas, their chatjimmy demo is a hard-wired (and probably very quantized) 8b LLM producing garbage extremely fast (17k tokens/s). Extrapolating from the picture of their PCIe board in their demo, I would expect a frontier model to require an IC not smaller than my desk.
At these dimensions, you are better off producing general-purpose inference ICs that can run any model, instead of baking in a single fixed LLM.
So, Taalas (now AMD) will probably target small models with these (maybe something like qwen3.8-27b, which seems too large already), if at all.
Apples and oranges, really. SHA-256 is a fixed function with no memory. So it lends itself to being hardwired into pure logic. A CPU or GPU was absurdly wasteful for it. That's why ASICs saw like 100x gains over general purpose hardware. Most of the repeatable tasks in AI workloads, like matrix and vector math at reduced bit widths, have already been encoded into hardware in modern GPUs.
Yes. Half dozen companies are designing special chips to bypass/replace the GPU.
Bitcoin was lucky for Nvida, then AI was lucky. Third times the charm?
On other hand, luck favors the prepared. Beyond luck, they did make their chips accessible, easier to program than other GPU's. So bitcoin miners and then AI could utilize them.
There are still a lot of people that want to run regular liner algebra.
Nvidia cards were the worst hashrate/purchase price of any cards. AMD were the lucky ones there if you consider anyone lucky in this situation. Consumers definitely weren't lucky because the GPU market hasn't recovered since the Etherium boom.
I think things changed, in the early days of BTC AMD (ATI at the time) was king. Later GPU made a comeback with Nvidia rigs doing well during Ethereum era.
Even in the later ethereum era everyone I know just had mining rigs full of RX480's since they were the cheapest 8GB cards (Before that banks of AMD 280X gpus until they didn't have enough VRAM). Then most people I know that were seriously into crypto quickly moved on to ASICs and sold off their GPU rigs.
I imagine the 1070/1070ti 1080/1080ti were decent at mining, but a lot more expensive to obtain for these gpu mining rigs with 6-10 gpus in them.
NVIDIA GPUs stopped being GPUs a while ago, that didn’t happen for crypto, but now it seems the will drop the entire customer graphics market long term they just haven’t committed themselves to it. They can’t go back to GPUs anyway there are not enough gamers to reach the revenue numbers they need now and soon gaming/graphics revenue will be irrelevant for them. Most US companies won’t maintain a boutique of products that don’t have synergy or are actively competing for internal resources long term they put all their resources where the margins and growth are best and for NVIDIA that’s not gaming anymore.
> NVIDIA GPUs stopped being GPUs a while ago, that didn’t happen for crypto
But Bitcoin mining moved GPUs to ASICs many many many years ago and Ethereum switched from proof-on-work using GPUs to proof-of-stake: GPUs are pointless now for Ethereum. And Ethereum was the big one for GPU mining. All the others cryptocurrencies are too tiny to register.
Ethereum switching from proof-of-work to proof-of-stake has been the death blow to GPU mining at scale. I'm not saying some have not been profitable mining other less important coins: I'm saying the market for GPU mining shrank to next to nothing.
Nvidia has multiple B2B customers that require the consumer graphics featureset, like Nintendo. There's sufficient motivation in keeping the GPGPU/raster capabilities on the same card even if PC gaming dies as a market.
I don't know. But my assumption is that they are. The crypto ASIC market was way smaller and less ambitious than this. It would have been just a relatively small cash grab for them to get into that market, whereas this market could represent an actual strategic direction.
If you look at the sessions from 2026's "Hot Chips" there's like a dozen competitors and possibilities and new architectures and the like, yes. Some very compelling. And OpenAI and Google and others are also making their own chips.
The thing is, it's only NVIDIA really that has a proven product in high volumes pumping through TSMC with the existing software stack that the community has already proven. That's a huge difference. Everything else is riskier if you're building out in volume. For now.
Relevent article from september 26. To me it’s obvious gpu are a « first response » solution. But I think nvidia can sell a lot before AI R&D slow down (because of the demand or the tech, infra, energy… or whatever reason it will be)
Mining ASICs are bound to a single cryptographic hashing algorithm, meaning they do one thing and they do it very fast. Something similar may make sense for running models locally, but I believe the more multifaceted Nvidia chips make more sense for training.
Taalas takes a similar approach but at a much higher level of complexity - instead of an array of nnn tiny specialized cryptographic cores running the same algo (or a mix of several with interlocking dependencies), it freezes a single entire LLM on to one die, running inference both far faster and cooler than any general purpose compute but with the same one trick pony obsolescence drawback. Question remains whether high volume fab tech can be wrangled into a form that allows this approach to drop unit costs near "disposable/recyclable" levels of cheapness per chip. This also may not matter when models all start to converge on a 'good enough' level of coverage for specific domains that the chips can remain productive for 3-4 years before needing a recycle.
>This also may not matter when models all start to converge on a 'good enough' level of coverage for specific domains that the chips can remain productive for 3-4 years before needing a recycle.
Yeah I was thinking this will probably be the case. We may eventually settle into a slower release schedule as well, where a new model only comes every year or two.
Check out Taalas, recently acquired by AMD. They printed a model directly to silicon and it generates 16,000 tokens per second. You can even try a demo here:
Whats amazing is that this is around 5x faster than the top end Cerebras wafer-scale silicon, at much less power/cooling cost per token. Of course Cerebras is a general purpose design, but perhaps Taalas could license CBRS wafer-scale patents to create a single monster chip that held either a single massive trillion parameter-scale model, or could an MoE design with a half dozen smaller specialized models on the same wafer with a shared SRAM pool.
I wonder if cooling such a beast, with very non-uniform hot spot distribution, would be too complicated to bother with though.
So will people buy gaming GPUs fron Nvidia again after how if abused them this time ? Or will the AI obsession be their downfall & example for others what not to do and how not to alienate the customers who get you where you are now?
This is why I think their pivot/venture to general robotics was smart. However, I doesn't seem like they will become profitable on that front because they were outcompeted pretty hard in most measures I can conjure.
Not to mention, many ASICs were outclassed by a few well-tuned GPUs in terms of power-efficiency. There were only a few coins that bucked that trend, usually just altcoin/shitcoin experiments that weren't worth mining in the first place.
I'm kinda bewildered, it's bizarre seeing people cheer on quickly-obsolete accelerators instead of demanding their OEM to adopt GPGPU. CUDA will be relevant for much longer than any ASIC can ever be, even (or especially) if the transformer gets replaced overnight with a better architecture.
Cerebras don't need to customise their chips. To my knowledge they are crazy fast but struggle to fit larger models and it's not clear they're economically favourable.
Ethereum's work function was designed to be memory bandwidth limited, which is something gous are already quite well optimised for. An asic would be faster but not by enough to be worth the investment, at least while gpu margins were less crazy.
The main thing is the actual number crunching and memory interfaces for LLMs are not crazy hard to implement, and the potential scale is huge, so NVIDIA's moat is not likely to be big enough to make their current margins sustainable in the longer run.
Yes, ASICs are application-specific e.g. "fixed". But, I think it's possible to target algorithm implementation that is inherently flexible, only limited by HW constraints (bit widths, RAM depth, etc) and thus cost. Yet, still fully pipelined and much higher performance than alternatives.
My background is ASIC design and more recently 3rd-party IP development for integration into ASICs. In our last design I made heavy use of resource sharing (RAM and holding regs for intermediary values, counters, LUTs, complex multipliers, adders, etc, even state machines).
That got the wheels turning. I believe such an approach can be generalized to support a wide range of computation and thus algorithm development. Even state machines can be configurable if designed that way. All controllable by programmable registers.
Just don't look for AI to take on such a task. Or go ahead and try, it already fails miserably when it comes to RTL design (verilog, vhdl). LOL
Yeah, that's my thesis since the beginning of this madness. I see it as a temporary blip (albeit it's already much much longer than I would expect !) and once the bubble pops, Nvidia will be back to their normal cruising speed. Which is a good company, making great products, but their valuation will be 1/10 of what it is now.
And if the boom continues, there is no way either that their clients won't develop a product of their own. It just doesn't make sense to buy chips by the 10s of billions $ on which Nvidia makes >50% margin.
Very interesting, thanks. I just started using claude code to drive filmcraft-cli for producing product videos, using elevenlabs for voiceover. The only part Ive been doing myself is the actual screen recording on mobile, but maybe cc can do that part too somehow. One of my recent edits https://www.instagram.com/p/DfJrVhwMSLP/
reply