111–120 of 377 posts
Oh come on Jeff, you forgot to buy GPUs for your AI cluster. Such a beginner mistake. All you needed to do is buy 4x xtx 7900 used on ebay and build a four node raspberry pi cluster using the external GPU setup you've come up with in one of your previous blog posts [0]. [0] https://www.jeffgeerling.com/blog/2024/use-external-gpu-on-r...
Earlier quoted context omitted.
The cost effective way to do it is in the cloud. Because there's a very good chance you'll learn everything you intended to learn and then get bored with it long before your cloud compute bill reaches the price of a desktop with even fairly modest specs for this purpose.
Yeah... Looks like can get about $1/hr for 10 small VMs, ($0.10 per VM). So for $3000, that's 3000 hours, or 125 days, (if just wastefully leave them on all the time, instead of turning them on when needed). Say you wanted to play around for a couple of hours, that's like.. $3. (That's assuming there's no bonus for joining / free tier, too.)
The desktop equivalent of your 10 T3 Micro instances is about $600 if you buy new. For example a Lenovo ThinkCentre M75q Gen 2 Tiny 11JN009QGE has 8x3.2GHz processor with hyperthreading. That's 16 virtual cores compared to the 20 vcpus of the T3 instances, but with much faster cores. And 16GB RAM allows you to match the 1GB per instance.
If you don't have anything and feel generous throw in another $200 for a good monitor and keyboard plus mouse. But you can get a used crap monitor for $20. I'd give you one for free just to be rid of it.
That's a total of $800, or 33 days of forgetting to shut down the 10 VMs. Maybe half that if you buy used.
Granted not everyone has $800 or even $400 to drop on hobby projects, renting VMs often does make sense
Earlier quoted context omitted.
The internet says 100W idle, so maybe more like $40-50 electricity, depending on where you live could be cheaper could be more expensive. Makes me wonder if I should unplug more stuff when on vacation.
100W continuous at 12¢/kWh (US average) is only ~$9 / month. Is your electricity 5x more expensive than the US average?
Still only $50/month, not $150, but I very much care about 100W loads doing no work.
I thought the conclusion should have been obvious: A cluster of Raspberry Pi units is an expensive nerd indulgence for fun, not an actual pathway to high performance compute. I don’t know if anyone building a Pi cluster actually goes into it thinking it’s going to be a cost effective endeavor, do they? Maybe this is just YouTube-style headline writing spilling over to the blog for the clicks. If your goal is to play…
Another fun fact, the network module of the pi is actually connected to the USB bus, so there's some overhead as well as a throughput limitation.
Fun fact, the Pi does not have a power button, relying on software to shut down cleanly. If you lose access to the machine, it's not possible to avoid corrupted states on the disk.
Despite all of this, if you want to self host some website, the raspberry pi is still an amazingly cost effective choice, from anywhere between 2 to 20000 monthly users, one pi will be overprovisioned. And you can even get an absolutely overkill redundant pi as a failover, but still a single pi can reach 365 days of uptime with no problem, and as long as you don't reboot or lose power or lose internet, you can achieve more than a couple of nines of reliability.
But if you are thinking of a third, much less a 10th raspberry pi, you are probably scaling the wrong way, way before you reach the point where a quantity matters ( a third machine), it becomes cost effective to upgrade the quality of your one or two machines.
On the embedded side it's the same story, these are great for prototyping, but you are not going to order 10k and sell them in production, maybe a small 100 test batch? But you will optimize and make your own PCB before a mass batch.
I thought the conclusion should have been obvious: A cluster of Raspberry Pi units is an expensive nerd indulgence for fun, not an actual pathway to high performance compute. I don’t know if anyone building a Pi cluster actually goes into it thinking it’s going to be a cost effective endeavor, do they? Maybe this is just YouTube-style headline writing spilling over to the blog for the clicks. If your goal is to play…
Earlier quoted context omitted.
I can get a Xeon E5-2690V4 with 28 threads and 64GB of RAM for about $150. If you need cores and memory to make a lot of VMs you can do it extremely cheaply
Source? That seems like something I would want to take advantage if at the moment...
It also means it performs like a 10 year old server CPU, so those 28 threads are not exactly worth a lot. The geekbench results, for whatever value those are worth, are very mediocre in the context of anything remotely modern: https://browser.geekbench.com/processors/intel-xeon-e5-2690-...
Like a modern 12-thread 9600x runs absolute circles around it https://browser.geekbench.com/processors/amd-ryzen-5-9600x
Earlier quoted context omitted.
He is not a YouTuber. And even if he was - what's the problem ? Jeff has many useful OSS software used by many companies around the world daily - including mine. What have you created ?
"He is not a YouTuber"... what? https://www.youtube.com/c/JeffGeerling "978K subscribers 527 videos" Jeff's had a pattern of embellishing controversies, misrepresenting what people say, and using his platform to create narratives that benefit his content's engagement. This is yet another example of farming outrage to get clicks. I don't understand why people drool over his content so much.
I then used many of his ansible playbooks on my day to day job, which paid my bills and made my career progress.
I don't check youtube so I didn't know that he was an "youtuber", I do know his other side and how mucH I have leveraged his content/code in my career
Earlier quoted context omitted.
I did some calculations on this. Procuring a Mac Studio with the latest Mx Ultra processor and maxing out the memory seems to be the most cost effective way to break into 100b+ parameter model space.
Depends on how heavy one wants to go with the quants (for Q6-Q4 the AMD Ryzen AI MAX chips seem better/cheaper way to get started). Also the Mac Studio is a bit hampered by its low compute-power, meaning you really can't use a 100b+ dense model, only MoE feasibly without getting multi minute prompt-processing times (assuming 500+ tokens etc.)