calaphos

Sadly LCOE is a very bad metric when looking at intermittent and non dispatchable generation sources in a power grid. Especially so as they approach a meaningful fraction of total generation. The market goal of power generation after all is not to produce as much kWh as possible but to satisfy demand at a specific time. System level LCOE which take dispatch into account are a better way of looking at power generation costs.

Here is a pretty good paper on the topic, albeit a bit older so some parameters might have changed slightly.

https://research.google/pubs/pub46289/

I assume it's int8 operations (FMA would count as 2). At least that's the case with FLOPs, TOPs for AI accelerators is basically the same measure, with the changed acronym reflecting the non float data format.

Even if technically possible the economics of nuclear power plants don't favour only using them for peak loads. With gigantic initial captial investment and very low fuel costs (almost?) all nuclear power plants are used for base loads.

It's hard to find current numbers, but capital investment for combined cycle natural gas plants seems to be about a tenth of modern reactor projects in western countries.

> I can't see any way to go much lower than that without freezing in the winter. 2 kWh/day seems crazy low to me.

Apart from insulation energy required for heating changes massively with local climate and scales directly with living space. So directly comparing those numbers without additional information is quite meaningless.

Here in Germany, which seems to have a roughly similar climate, ~120kwh/m2 and year is the average for a detached home. Modern buildings are usually a lot less though, 30-50 kWh/m2year seems standard. With specialised construction ('passive house') and heatpumps 10 kWh/m2year is well feasible.

Capacity has decreased - mainly due to a lot of french nuclear reactors offline for scheduled and (correlated) unscheduled maintenance. France went from Europe's largest exporter to largest importer of power this year. Add to that a dry summer woth historically low water levels in rivers, disrupting cooling as well as things like coal transport for power plants.

And natural gas, which usually covers the more elastic power demands is dramatically more expensive.

The number is for FP64 double precision terraflops. For vector operations, not matrix operations as well.

Double precision is important for a lot of simulation codes because of numerical issues. For Deep learning applications it's basically unused, single precision or half precision formats dominate training. The same numerical issues don't show up, so this is a lot more efficient.

Additionally a lot of the 'AI' flops are achieved through matrix cores, fitting Deep learning the workload well. This brings an A100 from only 17 to ~200 tflops.

> But can a 10-node-iphone-cluster really match an A100 for pure flops?!

The iphone number is probably fp32 if not lower precision. However the combined SoCs of 10 iphones have tripple the transistor count and a larger chip area among them. So the comparison is not totally off.

Similar classes of machines have been build and used for simulations in industries for decades now.

In fact if you look through the historic top500 lists a lot of the space outside the top10 is occupied by industry clusters. Classically oil and gas has been one of the big consumers. In recent years there are more and more cloud clusters in there as well.

So this Tesla cluster isn't really anything unusual.

> It makes we wonder if something like the Australia-Asia power link would make sense.

There is the AAPowerLink project which proposes exactly this. It's a ~20GW solar farm in Australia with battery storage and a 4500km HVDC undersea cable to export power to Singapore. Seems currently in a planning and Project seems to be in a planning and exploration stage with funding not fully secured though.

> Yeah, seperate stations (edit: platforms, they can share the building) for high-speed trains will probably never happen, sadly. And as long as high speed trains share the platforms with regional trains, disruptions and delays will always affect both.

This is the case for some stations, e.g. Berlin Hbf or Dresden Hbf. Anecdotally I've never experienced delays due to waiting in front of the station there, unlike e.g. Hamburg or Hannover where platforms are shared.

A major problem would he that the temperature delta for storage will be way smaller. There isn't really space for the same 500C below ambient in that direction :)

With more and more miniaturization such stochastical behaviour becomes more and more prevalent. Same with higher and higher speed interconnects. Forward error correction is standard on e.g. NAND flash, VRAM for GPUs and being adopted by faster Ethernet standards. It's mostly just the consumer dram market that doesn't have robust error correction.

They have been shipping multi die CPUs for quite a while, but the interconnect is closer to a PCIe connection (slower, longer range, less contacts).

Intels upcomming Saphire Rapid server CPUs are extremly similar, with wide connections between two close dies. Crossectional bandwith is in the same order of magnitude there.

In benchmarks amperes cores are roughly as fast as a hyperthread on current gen Epyc. Sometimes more on heavily execution bound workloads, sometimes less.

The same issue exists with Jax. XLA compilation can take up quite a bit of time, especially on larger NN models. And theres no persistent compile cache, so even if you don't change the jitted function you need to wait for compilation again as you restart the process.

Per transistor performance in M1 isn't any better than the x86 opposition, quite the opposite in fact. The standard M1 has 60% more transistors than a current gen Ryzen mobile chip, while having halve the large cores (and significantly worse multi ore performance).

We had similar discussions in Germany on taxes on any CO2 emissions. There were proposals of redistributing some of the earnings to people with very low emissions, a bit similar to negative income tax rates in concept. Sadly never really went anywhere.

One could even eliminate the heavy and potential dangerous battery with some sort of overhead wire or rail contact. And since the cars can only move in one direction within a tunnel, one could eliminate the need for steering by placing them on a set of rails. And maybe use some sort of automatic control system to make sure the train stops when needed. Wait..

I don't think the goal of this is to make games look better.

Especially the GTAV dataset is often used as a synthetic dataset for research. The appeal is that you can extract normal/material and other semantic information from the game engine for a driving scene.

Basically avoid expensive segmentation labels on real data by working the other way around. Things like this synthetic to real translation can then be used for other downstream tasks (semantic segmentation, distance/normal prediction).

High mandated standards on Fire/Noise Protection, Insulation, accessibility on new buildings make it really hard to build "cheap" rental apartments. With the mandated standards in place there is little to no cost difference between high cost and low cost residential buildings.

Those standards obviously have their benefits or are unavoidable (fire protection) and result in new housing being high quality. But they make it hard to address housing shortages in a non long term way.

The current ARMeilleure seems to emit x86 instructions directly and execute them, so never really using the intermediate representation of Jit compiler of the .Net runtime.

However the previous cpu emulator apparently emitted .Net byte code and used RyuJIT (the .Net jit) to emit the final native code.

> if it's running on an ARM processor does it just run the instructions directly?

So maybe? There are some passes in between but one would think it would at least result in very similar instructions.

Given the enourmous costs of decommissioning nuclear power plants already I fail to see how a reactor containment that lasts 10-20 years at best improves the situation.

For ground based targeting radar at the destination at least. Not sure how relevant this is with satellite based radar or infrared detection systems.

It is absolutely possible and exactly how connections between the different grids are done, e.g. between the power grids of North America or between the European grid and Russia.

HVDC of course has the disadvantage of extra conversion equipment. HV circuit breakers are also significantly more complex as an arc will form as long as current is flowing. With AC this happens at the zero automatically, nothing like this with DC.

Cymer seems to only do DUV light sources, the EUV stuff is developed by Trumpf and Zeis in Germany.

A better analogy for context switching would be stowing everything away you used for making cookies before starting the next task.

What you are describing would work as an analogy for hyperthreading / smt though. Not all ressources are used because one needs to wait for things to be ready. So you do something else to keep busy in the same space as the rest of your kitchen.

If you like some practical project I can highly recommend Day tracing in one weekend by Peter Shirley.

Apparently first threshold for cutting of hydro storage pumps below 49.85 Hz was undercut very briefly. Immediately afterwards the frequency reaches that threshold again, so I assume those measurements were working as planned.

Below 49Hz larger consumers would be cut off in steps of ~10% of the network load. AFAIK this includes large industrial consumers who are legally required to have the equipment in place to allow for that.

> I’d be curious how the unified memory architecture shifts the cost dynamic for GPU acceleration.

Correct me if I'm wrong, but is this actually different from regular integrated graphics that have been in intel and amd chips for decades? I remember there being some initiatives from amd proposing similar offloading under the name HSA almost a decade ago. I don't think there are actually any software really using it.

You can still benefit from AVX for short workloads, the latency of them is no more than other instructions.

Running things on some accelerator (gpu, etc.) usually involves writing a specific kernel in a language subset, manually copying data and generally long latencies. Unless there is a lot of data it won't be faster.

With AVX in the best case the compiler can just vectorize some loop, speeding it up 5x without any added latency or source code changes.

Not relay the same. This predicts an additional depth channel using CNNs which you can then use to create some sense of 3D by translating the image elements differently based on depth when rotating the camera a limited amount. However you won't be able to e.g. inspect the backside of the object or model things like cavities using this approach. It's more like a heightmap or relief.

With nerfs you get an actual volume representation with a mapping from every point in 3d space to whether it's inside a volume or not. It's an actual 3d model instead of a 2d image with a depth channel.