MB
MasterBeef Group Ordinary Shares
Mentions (24Hr)
0.00% Today
Reddit Posts
Mentions
I believe when i use RISC-V on my phones and laptops. But the truth, it's a long journey and that's primarily because there isn't that big of a commercial incentive (cost, performance uplift) to overhaul the infrastructure (hardware and software) Maybe you can argue that with the best coding agent/LLM the software part can see a less painful, faster transition. But then someone still needs to validate that things works as expected. Apple and AWS (Graviton) maybe can pull that off. But they are not going to RISC-V for no good reasons. A reliable ISA (x86\_64, aarch64 and RISC-V) without a abnormal leap in process node technology improvements (e.g. lower Vmin @ high clocks 6GHz all cores or 100MB of SRAM L2 with 3nano-second latency) creates no incentives. Qualcomm can have RISC-V, a top notch SOC design but the process node part is where the story falls apart. https://preview.redd.it/dh0jwp45h6rh1.png?width=2262&format=png&auto=webp&s=66e57d93c5e5443bcb28b2d73b4f88ac9f6508b0
Sure buddy https://media.tenor.com/MB1xLfxOIUsAAAAe/stocks-stock-market.png
Where is that and what company. Where I am it's 24 euros for 300MB and 10euro for almost unlimited.
You've been wrong on many levels so far: 1. You didn't know how activations moved in bandwidth so you assumed the entire 500MB needed to be moved. This invalidated your first argument. 2. Next, you kept saying these chips wouldn't be good for training but I never said these chips are for training. 3. Now you're making some other stuff up. The goal post keeps getting moved.
>The rest of your post is wrong because your technical assumption is wrong. This is a strange claim to make considering how obviously verifiable the response is. And why are you only introducing concepts like weighting and vectors in your reply, not in your central thesis? Shouldn't you be explaining the architecture of your setup so you don't intentionally mislead people? The bottom line: yes, network bandwidth becomes an issue. Nvidia didn't invent NVLink and the entire industry didn't decide to develop a competing alternative because network bandwidth doesn't matter. Large training clusters are typically training on data sets vastly larger than what they can store on-net. Network communication is a requirement, and therefore networks communicate at network speed, not cache speed. Why on earth would you build a networked system and only train it based around the limit of a 500 MB cache? I will admit this is entirely possible - it's just a gross use of CapEx, and therefore highly cash inefficient. So your idea works, it's just every company out there can save money using systems with better architectures, and the Groq option is only cash-optimized when the training needs are exactly its capacity. Go 10% larger or smaller and its performance loses out to DRAM-based systems.
>If you combine Groq chips in an MCM (multi-chip module), you expand the SRAM memory, but now you're talking networking 2 chips together, which reduces system performance to DRAM performance at best. At that point DRAM wins. This is clearly wrong. You do not send all 500MB of weights across the network. You only send the activation vectors which are in kilobytes or megabytes. The network bandwidth never becomes the bottleneck. The rest of your post is wrong because your technical assumption is wrong.
So this all comes down to application and tradeoffs. If the limit of a Groq chip is just 500 MB that means most smartphones today have more DRAM. Then there's the application: even the desktop AI systems today include 128 GB of DRAM in an attempt to run the most-capable models. This means Groq is only suitable for small models if all you rely on is SRAM. If you combine Groq chips in an MCM (multi-chip module), you expand the SRAM memory, but now you're talking networking 2 chips together, which reduces system performance to DRAM performance at best. At that point DRAM wins. To even consider larger systems, involving multiple blades of Groq hardware, now you can't use MCM, you've got to resort to LAN networks using some NVLink or equivalent, maybe silicon photonics. What you end up with is: DRAM can scale as large and fast as the limits of network bandwidth today, meaning it is the memory of choice. The fact is SRAM is simply too limited a technology to provide any meaningful aid to AI training. If the AI model is small enough, SRAM can provide a boost, but now we're talking the pico-models that are even too small for smartphones. Very limited applications for any models like that these days.
"Having worked here for a few years..." is the crux. You can sell some of the peolle all of the time and all of the people some of the time. No comment on whether its BS or not but MB is better at beating that axiom than most.
Jensen and MB are playing the same game. We are sports fans cheering on dumbasses trying to climb a greased pole.
Dude is the ultimate salesman, ice to Eskimos. I think many of these big tech founders/execs have a lot of internal doubt and self loathing but I am 100% certain MB 100% believes the MB hype.
If this keeps going, a 32MB USB Stick will be $100 next year.
Export controls would hurt more than help, US produces the most oil but still uses significantly more than they produce (also the oil that gets exported is stuff the local refineries don't want). Granted most of the imports are from Canada who we're best buds with- Oh wait. As far as run until they physically can't, the current drawdown order will bring it to about 250 MB which is close. Realistically if a sane admin wanted anything in reserve in case of... hurricane, natural disaster, black swan, they wouldn't draw further. I don't think they're sane but the market sure seems to have thought so and still does.
I do love that they absolutely never show a picture of the actual person. I bet Christian Bale gets stopped in the streets and they think he's MB
Cisco´s success is kinda a mystery to me. Don´t get me wrong: I got no issue with their products. But the sell a router for 50MB/s thoughput. With a performance license it can do 100. Great. Scalabe with some buffer. Licensing as a business model. Fine. But then you can buy the boost license and get a gig out of that thing. And this is what I can´t wrap my head around. They dish out hardware capable of gigabyte traffic for 50 MB/s circuits. How is it no one manages to build 100 MB hardware routers and just massively undercut them on price?
Buying more nbis Bc fuck MB
Inflation would be even lower, if they replace the 32GB with 32MB RAM modules because of hedonistic improvements.
Lol NBIS. If they didn’t make that damn movie this MB retard would be a nobody
Are you also comparing cost of sever setup TODAY to cost of AI setup TODAY? You need to do a whole lot more research and math before you can make that statement. Back then you had to pay a significant portion of your income per MB of RAM, for example. Before AI you got gigabyte sticks for practically nothing.
I mainly trade momentum. So I will jump in with a .05 buffer from the current price as it is going up. Once I am in and the price starts to hesitate, then I will immediately close out my position. I don't have a set goal for profits, I just watch the candles/price action and make a decision. In this case with MB, once it hit around 9.50, the price started to hesitate and pullback, that was my sign to get out. Yes it did run a lot higher, but I've learned to take profits off the table and not get greedy. So whether it's $20 profit or $200 profit, I am happy with either one. $70 profit for your $1000 investment may seem small, but do that 2-3 times a day and it starts to compound significantly. Hope this helps.
Lol whats funny is that when people see someone is shorting they think that person wants to see the stock down 50, 60, 80%. WRONG. Shorters can hold that for 10 to 20% gain and get right out. They dont need to hold forever, few weeks at most. You think shorters like MB lose money, your regarded!
not a penny stock but MB is going nuts up over 400%
Everyone talking about HTZ but MB is up 426% rn lol. Halted 4 times now
These HTZ NPCs need to learn to buy a real meme stock like MB
MB is a scam lol, 222% in 1 day
MB is professional gambler and an gambling addict. Simple as that...
in the 90's I traded 8 MB of ram for a Chevette.
Im long LULU, ADBE, DKNG. Shorting NVDA pls subscribe to my $ newsletter for more great advice -MB
Everybody on his payroll is literally scared of him. It wouldn’t surprise me if some republicans give him their underage daughters at his desmmmmm MB bc-
Made the switch to a MB Air from PC and I finally get the Eco System hype. Had the iPhone since 2007 launch and I feel like I’m finally putting the last piece of the system together.
Who is "we" and who treats MB like a prophet? Literally no one. Get off your shit-baiting horse and do something productive.
I don't see any indication these won't require HBM. It's only got 64**MB** of on chip memory.