1,086 karma · joined December 21, 2010
All that said I actually don’t think that matters much. I think we are dragging attention economy concepts in to ai responses, and it doesn’t matter. Both options saved me hours per week, and the difference between 3 and 1 minute may not be worth the additional cost.
Also there are times when the model output is much better with anthropic, but it’s not all the time. I think it becomes a question should we be using the best model for all questions?
In practice they are also not very flexible when compared to gpus.
In general, this solution would be expensive and targeted at data lakes, or areas where you want to run computation but not necessarily expose the data.
With regard to DRM, one key thing to remember is that it has to be cheap, and widely deployable. Part of the reason dvds were easily broken is that the algorithm chosen was inexpensive both computationally, so you can install it on as many clients as possible.
I don’t think the prices have adjusted because of that. Additional during Covid the prices were very high and this is baked into the pricing.
I would guess their interconnect technology is what NVIDIA wants. You need something like 75 adapters for an 8b parameter model they had some really interesting tech to make the accelerator to accelerator communication work and scale. They were able to do that well before nvl 72 and they scale to hundreds of adapters since large models require more adapters still.
We will know in a few months.
Look up Intel pfr.
I don’t know that the first two have changed significantly.
It’s more likely a set of products that were shipping directly from factories disappears from the market. For example, the direct from factory Halloween costume.
It could end up being a step backwards in living standards and access to daily luxuries.
Overall, there is a continued challenge with CPU temperatures that requires much tighter tolerances both in the thermal solution. The torque specs need to be followed and verified that they were met correctly in manufacturing.
The protection here is to ensure the vms are isolated. Without doing this there is the potential you can leak data via speculative execution across guests.
The part that I’ve been scratching my head at is whether we see a retreat from aspects of this due to the high costs associated with it. For cpu based workloads this was a workable solution, since the price has been reducing. gpus have generally scaled pricing as a constant of available flops, and the current hardware approach equates to pouring in power to achieve better results.
For you that means focusing on a growing area of the company, and finding new areas to grow your team in. You also need to have a team of managers, who are growing their scope as well.
They reach the conclusion here they are more reliable.
The bigger issue I think is most of the cars are teslas, which didn’t behave like a normal automaker for better or worse. For example the work done during the pandemic to avoid supply chain crunches may result in a maintenance headache a few years from now.
Fab + design... its apples to oranges.
TSMC for example has 77k employees and looking to add 23k more.
Packaging and testing are labor intensive, and require folks to be added in different geographies.
I'd also like to know more about what its doing.
It is insane especially if you think emulation is performant enough to allow for a switch.
From what I've seen, but haven't heard discussed much, the naive implementation vs AVX512 is a huge gain, but AVX2 vs AVX512 was not very impressive for the application I was looking at. The complexity this code added, and the cases where we needed it to run on AMD (for other reasons), basically made taking advantage of the feature undesirable for a single digit gain.
Things like VNNI or AMX are better wins, but they are only needed in very specific cases. VNNI in particular looked to be a 30% improvement in a BERT workload.