If we're going to waste electricity, I'd rather we waste it on AI than on mining Bitcoin. At least AI is providing some utility to its consumers and there's reason to believe it will provide even more in the future.
Bitcoin provides massive utility to customers. The banking cartels + global government massive taxes are destroying livelihoods daily.
The compute is "wasted" to pick winners and secure the system, but it's not actually wasted, it can be made more efficient by actually using the compute to do something, but even if that was not the case, it pays back its value 100x.
I think there's a reasonable argument that cohere might not capture the value and it might all go to the big players. But I can't understand how anyone still has the view llms themselves are not valuable.
Man I hate Nvidia and their monopolistic behavior. Making 800% margins on H100s is just absolute greed imo. They make good GPUs but they don't deserve to be an almost $3T company...
A ton of companies are dumping R&D into designing their own chips with seeming success. Nvidia will be in tough company quite soon, but for now demand greatly outstrips supply.
I hate the industry for never learning how to get along and make a real CUDA competitor. Nvidia can make 1600% margins for all I care, this is exactly the sort of future we deserve when nobody can agree on a common GPU API.
You want competition, you want vertical integration? Here it is. Compete.
The TPU is basically an ASIC as far as I know; it competes against CUDA in a very small subset of it's featureset. CUDA is essentially a composition layer on top of multiple GPU features that optimizes them for general-purpose compute. In essence, nothing is stopping Apple or Google from making an Open Source CUDA replacement and undermining the demand for specialized GPGPU compute. The problem is that CUDA is massive, and nobody wants to re-implement it (especially not for free).
So now Nvidia is in the privileged position of having both highly-flexible GPGPU compute hardware, as well as a highly-advanced software layer to use it with. TPUs and NPUs are neat, but fundamentally they are neither of these things; they have an extremely limited processing pipeline exposed by a high-level library, and that's usually it. CUDA is comparatively flexible, to the point that it doesn't even rely on AI to sell it's product.
To me, hating on Nvidia feels like being mad that a well-bred horse with great odds beat out the jockey you were betting on. Why should we hate them, for their "monopoly" on features that Apple and Khronos gave up developing? Because they're blocking-out their competitors by... not having working MacOS drivers per Apple's request? This is the causal and obvious outcome of letting businesses commoditize specialized compute. This is what the industry wanted, and it's rich watching the customers protest like they were fooled into thinking everything was fine.
> The TPU is basically an ASIC as far as I know; it competes against CUDA in a very small subset of it's featureset. CUDA is essentially a composition layer on top of multiple GPU features that optimizes them for general-purpose compute.
my understanding is that compilers can compile some straighforward JAX, TF, Pytorch programs to both Cuda and TPU, so they in direct competition in current hot topics (LLM, deep learning).
Right; but you can't cross-compile everything. This is really common in AI libraries, especially multi-target projects like ONNX: https://onnx.ai/
The math probably adds up in Google's favor with the TPUs, even if they end up being less efficient and slower per-unit than Nvidia hardware. They don't need to pay for the margins, and they can run them 24/7 for their intended purpose. The previous-generation TPUs can't be reused or resold for other purposes though, and if/when AI blows over as a trend you probably can't easily start mining crypto or doing HPC calculations like an Nvidia cluster would.
Our company can buy NVIDIA gear. Google TPU, well, is google's property I have no control over. This site is overrepresented by clod folks. In reality ~90% of the workloads are NOT cloud based. Unless one can buy Google/Amazon/Meta T/N/AI/PUs that unique optimizations are irrelevant for the most of the workloads.
If they are selling something that everyone wants and there aren't enough GPUs to around, how else do you suggest allocating those GPUs to the public except by increasing the price?
So say they mark them down to some reasonable level — they are still supply constrained and now the line just gets longer. It doesn’t really fix anything.
Pricing allows them to prioritize their customers, which seems reasonable.
if the line gets long enough, it allows other companies to compete with Nvidia and creates a healthier, more competitive market, in the same way AMD has made intel a more honest, fairer, less price-gouging company. Competition is the entire thesis of american capitalism
> Competition is the entire thesis of american capitalism
Then why aren't people competing with Nvidia? Why is OpenCL on life support and unsupported on major operating systems? Why are we doing this song-and-dance routine refusing to adopt certain GPU APIs but embracing closed ones instead?
I'd like to believe that a tipping point will be reached, but if not now then when? People have talked about upending Nvidia's GPGPU compute empire for years, but besides application-specific replacements and proofs-of-concept, we don't have a real CUDA-killer. Apple does not ship one, Google does not ship one, Microsoft does not ship one and AMD doesn't either.
So... when? If we continue along the current path, I suspect Nvidia will continue to find markets where CUDA is demanded and OEMS will continue to chase them down with half-measure solutions. Unless OpenCL is revived or someone commits to a proprietary CUDA-like platform, I suspect we'll be spinning our wheels and digging ourselves deeper.
> Then why aren't people competing with Nvidia? Why is OpenCL on life support and unsupported on major operating systems? Why are we doing this song-and-dance routine refusing to adopt certain GPU APIs but embracing closed ones instead?
Because Nvidia invested 20 years into its API platform, and this advantage is slowly getting realized.
Well therein lies my confusion; why did their competitors get 10 years into making a competing API, and then give up?
If the reason that OpenCL died is because Apple decided that they'd rather draw blood than work with the community, then yeah, this is a well-deserved failure on their part. Even Nvidia was willing to contribute to OpenCL; the only thing stopping us from living in a CUDA-agnostic world is the pointless and childish aggression between device manufacturers.
It feels less like we're slowly realizing things, and more like the persistent failure of Nvidia's competitors is forcing them back to the negotiations table. It's pathetic that American businesses are this willing to throw each other under the bus before they consider working together for the common good.
Tech hardware companies are the only people where I'm not annoyed at their prices.
Want a fence built for you? $3000 for 2 days work with 96 pieces of wood.
Want a basic android phone with 4 gb of memory, 128 gb of storage, a camera, speakers, touch screen, wifi, cell tower connection, GPS, battery, charger and cable? $60 on sale at Walmart for a Moto G Power Stylus edition. (My current phone and that's what I paid a few months ago)
If they sold them at 30% margins or whatever you think is reasonable, then scalpers would just pocket the 770%. Not sure how that is better for the ultimate purchasers or Nvidia.
It's because Microsoft/OpenAI and Meta are throwing multiple $10B's at NVIDIA. This is unsustainable and NVIDIA's stock will contract over 3-6 years unless Microsoft and Meta can translate hype into essential, profitable products which isn't a sure thing™.
Because monopoly means "wields monopoly power and engages in anticompetitive behavior that harms consumers." Having competition or not, having large market share or not doesn't matter one bit.
One of ways of wielding monopoly power is no longer being a price taker and being able to set prices at in practice to whatever you want.
Your definition of monopoly contains a reference to "monopoly", so I am still not sure what your definition of monopoly is. It sounds like it is something like "big company that charges a price you think is high". Why do you think you know what the "right" price is. Isn't it a good thing the company delivering the most value is capturing the most profits? They will likely continue to deliver breakthroughs with the profits they are generating. Why would you want to kneecap a company generating so much value?
No further explanation of what Cohere might provide to it's customer of value, just the ability to generate text.