Recent content by HansSchulze

Upgradable costs even more. Nvidia spent years trying to convince Quanta MSI et al to use a slotted GPU, cost adder was around 100$. Even memory sticks are soldered down these days to save 5$. Robotics would help, but those are expensive and still need supervision and programming and fine...

Youtuber northwestrepair comes across nvidia gpu boards like 4080 with the gpu and memory removed, sold on the street for list price. That's a good way to supply china.

Sorry, the cluster needs all gpus to be within about 100m for speed. If you added distance, they would need 5x more GPUs to get the same performance.

And how long would it take to build a 1GW NGPP next to this data center?

One significant source of increased power consumption near launch is when the software guys figure out some optimizations or realize that they didn't understand how much better the hardware was if they just tweaked a few methods, and then they implement using the new tricks, and voila, the power...

Exactly. If Nvidia prioritizes biz by price paid, China might not get anything. I'm sure they want Mellanox NICs before those are deemed restricted. Nvidia should sue back for the hacking. BTW if the author wants to and feels qualified to report on tax management strategies, it might be more...

I'm looking for an android cpu str3ss t3st, prim3s would have been nic3

How did we get from October mass production to January shipping? Heat density is a thing. Even with liquid cooling it is difficult to continuously move thousands of watts from a to b. Temporary fix is to maybe install 2 of 3 boards? Not earth shattering. Problem is that if they change the...

Sorry, I don't get your question or point. How many were being sold at list price? My point was that no one outside of NV or AMD or Intel or Broadcom have an idea how expensive it is to make square inch silicon at TSMC, and write good software for it. Also, what's the difference between...

There are huge costs with increasing performance by 50% per year, and every other year, doubling of mask prices and pancaking of yield levels to less than 50%. Take a look what the prices are for a brand new N4 TSMC wafer from scratch, it's in the billions of dollars. Then 6 months go by where...

Note that this estimate assumes 100% duty cycle but ignores the power needed to cool the servers, two effects which tend to cancel each other. I read that original paragraph, and as a power engineer, I don't understand any of it. The rest of the website seems reasonable, but skips over a lot...

Peak power is needed during very intense operations that are compute bound. It is hard to achieve full peak power consumption in most apps unless they are hand tuned. Some workloads like inference are more memory bound, so GPU isn't running as hard, and memory power is a small fraction of the...

Yes, and all the temperatures listed are out of spec. That case air temperature will overheat the GPUs. Copper has a gradient, heat doesn't just run from one end to the other. The reason heat pipes are used is that steam travels at sonic speeds as opposed to inter-electron thermal speeds...

Um, a block of metal is like a cup filling with water when likened to heat. Yes, a bigger bucket lasts longer before it overheats. You still need surface area to exude the heat to the nearby air, unless you only need 2 hours of gameplay a day. You could just use a block cooler pump and a...

Nope. The article says that ChatGPT3 didn't tell you what slot to put your graphics card into. I put graphics cards in any and all PEX slots. Even PCI did that correctly. Just to point out that the author was grabbing at straws to technically criticize the AI.