Shows once again, not all engineers are capable of selling products as well. Too bad, I was assuming them to IPO at some point. Due to the Huggingface cooperation I expected them to sell more chips.
Xmos was around during my time at Bristol 06-10 although wasn’t until later I saw anything in the market. A few graduates went on to work there and apparently at least one is at graphcore, too
graphcore feels like a real outlier, hard to see this as a harbinger of some market trend towards consolidation. it will no doubt come because the market simply doesn't exist to support the number of companies developing hardware or the wide range of different hw and sw products they're developing, but I wouldn't read too much into this sale.
What's the market share like for proprietary single-use accelerators vs. just buying GPUs right now? Seems like the hyperscalers said no to startups and they're going to continue to drop off.
There's a kind of contradiction going on with that question. Although it's possible for it to happen the other way around - ie ARM tech in the cores in IPUs. That said, Graphcore could just as easily licence ARM etc IP, without being part of the same group.
@robinhodson9890 They could but I was just wondering if softbank could create some synergy between the 2 subsidiary tech companies 🤔 with an advantage only provided by having ownership of both like maybe no licencing fee or something 🤷🏾♂️
AI will be the next big thing when it is ready. Right now we are not dealing with the AI that people think of. Maybe in a decade or two it will be truly ready.
Credit to Theranos, at least it was a solution to an extent problem. The only problem AI solves is "can I lay people off and make the customer experience even more frustrating."
AI is nothing like a bubble, let alone a scam. It's so funny seeing people root for this to fail when it's already huge. Reccomender systems have been huge all over the internet since 2014, ChatGPT was the fastest adopted program ever to reach 100 million users, etc. AI isn't going anywhere, even if some small start up exits or even fails.
Ah you're right, I forgot in the moment. But it still rings true that TSMC doesn't have a large chip on N3, and it won't be Graphcore as the lead customer.
@@minhhungnguyen7867 I'm a (very) recent computer science graduate, I eventually went outside of AI chipmakers for a job (finance) but it'd be nice to try and apply again I guess!
Hi. It's Graphcore. Contrary to the 'Graphcore no more' title - we're here and in extremely good health. We're working on our next generation AI compute systems and we'll be delighted to share details with you when the time comes. Couple of factual corrections: IPU is Intelligence Processing Unit (rather than Inference Processing Unit) Nigel and Simon didn't have a connection to Bristol University. The Good Computer was a longer term project. The 1U M2000 and Bow2000 units went into IPU Pods. The most recent version of our Poplar software is v.3.4. We were already on v1.2 by July 2020.
Would love to see a follow up for Groq apparently they are increasing their forecasts and are trying to deploy 1.7 million chips by this time next year
@@TechTechPotato that's absolutely nuts. "I want the colour of the hardware I can't see in the room I can't enter without hearing protection to be pretty." Some people. Next you'll tell me the whole push for RGB was never driven by gamers but by data centres ;D
Seems preferred stock was totally liquidated as part of the sale, leaving no equity for the common stock. That is, softbank paid off graphcores debts and got what was left for nothing.
I’m very interested to see a follow up to what IBM are doing because they have been deploying AIUs in clusters and WatsonX and they are even using them to assist their supercomputers with inference
How did they last so long? When such an acquistion happens, the creditors get paid first and then employee and other stockholders get paid for their now fully matured stock. Value of each share may be small.
Nvidia have been so dominant since their hardware is so flexible in model architecture innovation with cuda. Going all in on hardware optimized for CNNs is a big gamble that did not pay off. What is Tesla dojo up to? Musk is bragging how many nvidia gpus he is buying, I guess dojo is dead in the water as well...
Most of these 'small' AI startups are upside down with expenditures far outweighing revenue, and investors wont want to keep them afloat forever. I bet in the next 4 years this list is more than halved, due to consolidations and bankruptcies, with companies like Intel, Apple, AMD, and Nvidia circling like vultures to pickup IP and talent. Graphcore going to Softbank may seem like a saving grace, but Softbank is so mismanaged that Masayoshi Son will almost certainly take a loss on Graphcore before either trying to sell it again or just dissolving it.
Graphcore's stabilised. Being acquired isn't assimilation; it's a type of investment where you have one guaranteed customer, you remain basically independent, and where that customer gets a kind of bulk purchase price deal, plus some say in your direction, and all your profits. But you can still pull out, although hardly anyone does. Their trouble is it took longer than anyone expected in the R&D phase, which caused initial investment to dry up, so they had to halt their planned expansion and take some cost-cutting measures. They are through their extended early building phase now: They have a product range which sells well, and a library of APIs and tools to use it. Their hardware make datacentres more efficient, and it plugs straight into existing systems and how those systems are used. They're one of the few companies pioneering WSI (Wafer-Scale Integration), which enables all sorts of improvements not otherwise possible. They're still innovating and pushing forward: The future looks bright. Disclaimer: This is only my own opinion of them and their situation.
I wonder if a FPGA compute card or a mix of FPGA + some kind of ASIC (with vector, matrix or graph functionalities) could be a valid long time investment approach? If there were to come a replacement to transformers then we will see the same thing again happening to the current transformers optimised chips, like what we saw with the "obsolete" CNN accelerators. If a company would make a mix of FPGA and ASIC as chiplets, then migrating to the next thing after transformers would perhaps be much easier(?) by adding the new ASIC type and keeping the FPGA ones, and perhaps also keep some transformers chiplets if that would make sense. An iterative and modular (reusable) design approach in other words.
The NVIDIA Blackwell is already quite optimized for AI operations with MXFP6 support. Etched is going toward ASICs for transformers, but it remains to be seen if they can pull it off in this space where nobody has anything to show for, despite billions in funding.
@@TechTechPotato more like asking for no more shitcoin content honestly. "AI" is just another attempt by the tech industry to find the golden unicorn and reverse the trend of steady decline that its currently exhibiting. Good luck with that one :)
@@modernsolutions6631bingo - I would have given them way more credit if they called it something clever like "int2", but instead they really went for "matmul-free"
@@aapje It will, but the crash will be bad. This isn't like Cryptocurrency where that was kept in a bubble unto itself, so when things hit the fan... the world kept spinning.
no it won’t. this isn’t the same as a housing crisis where it effects mainstream people. how will AI do the same? related stocks will crash but how is that going to crush the world economy?