chessgecko

I think people might prefer the lower idle power consumption from lpddr over the better bandwidth in hbm in battery powered stuff. That said right now the price is definitely preventing us from finding out.

pretty sure its lpddr not hbm.

The real issue is that housing is heavily underweighted in the cpi basket. How many people do you know that are only spending 12.9% of their after tax take home on housing, water and fuel? Only people with paid off mortgages.

Wouldn’t say never harder, but it’s been pretty flat.

https://fred.stlouisfed.org/series/FYFRGDA188S

My girlfriend loves it. She’s an immigrant from Kazakhstan and apparently the Kazakh community on Threads is very active. Reminds her of home.

if you bought a nvidia h100 at wholesale prices (around $25k) and ran it 24/7 at commercial electric rates (lets say $0.1 per kwh), then it would take you over 40 years to spend the purchase price of the gpu in electricity. Maybe bump it down to 20 for data center cooling.

I don't think the cost of the ai is close to converging to the price of power yet. Right now its mostly the price of hardware and data center space minus subsidies.

Actually since the qe to fix 2008 foreign investors/governments cut down on bond buying. Treasury has foreign bond holdings up about 2T since 2014 and current account deficits sum to around 7T.

Nowadays we're mostly selling companies and real estate. (though there was a nice jump in bond buying a few years when rates went up)

Or maybe a lot of it can't deliver returns in line with the current price. The world realizes that and stops sending us stuff for those investments. Then we're stuck with limited manufacturing, high inflation and relatively low ownership of our existing high value exports.

If the pie gets bigger too slowly then I think we still lose.

It is tracked! there are a few different numbers.

by the BEA: https://fred.stlouisfed.org/graph/?g=gX0f

by the treasury: https://ticdata.treasury.gov/resource-center/data-chart-cent...

It's a little hard to figure out how important it is, but there are "good" numbers.

1800 on the h100s is with 2/4 sparsity, it’s half of that without. Not sure if the tpu number is doing that too, but I don’t think 2/4 is used that heavily so I probably would compare without it.

I remember reading that it’s too hard to get good memory bandwidth/l2 utilization in the fancy algorithms, you need to read contiguous blocks and be able to use them repeatedly. But I also haven’t looked at the gpu blas implementations directly.

I guess it's hard to know how well this would compete with integrated gpus, especially at a reasonable pricepoint. If you wanted to spend $4000+ on it, it could be very competitive and might look something like nvidias grace-hopper superchip, but if you want the product to be under $1k I think it might be better just to buy separate cards for your graphics and ai stuff.

I think we don't really disagree, I just think that this shape isn't really a gpu its just a cpu because it isn't very good for graphics at that point.

For ai inference you definitely have other options, but for low end graphics? the lpddr that apple (and nvidia in grace) use would be super expensive to get a comparable bandwidth (think $3+/gb and to get 500GB/sec you need at least 128GB).

And that 500GB/sec is pretty low for a gpu, its like a 4070 but the memory alone would add $500+ to the cost of the inputs, not even counting the advanced packaging (getting those bandwidths out of lpddr needs organic substrate).

It's not that you can't, just when you start doing this it stops being like a graphics card and becomes like a cpu.

GDDR isnt like the ram that connects to cpu, it's much more difficult and expensive to add more. You can get up to 48GB with some expensive stacked gddr, but if you wanted to add more stacks you'd need to solve some serious signal timing related headaches that most users wouldn't benefit from.

I think the high memory local inference stuff is going to come from "AI enabled" cpus that share the memory in your computer. Apple is doing this now, but cheaper options are on the way. As a shape its just suboptimal for graphics, so it doesn't make sense for any of the gpu vendors to do it.

One day is doing some heavy heavy lifting here, we’re currently off by ~3-4 orders of magnitude…

The real reason is that it's basically impossible to produce a cheap new car that is a better deal than a Toyota with 80k miles on it.

For decode steps it depends on the number of inputs you run at a time. If your batch size is 1 then it runs in line with active params, then as you get to like batch size 8 it runs in line with all params, then as you increase to 128ish it runs like the active params again.

For the context encode it’s always close to as fast as a model with a similar number of active params.

For running on your own the issue is going to be fitting all the params on your gpu. If you’re loading off disk anyways this will be faster but if this forces you to put stuff on disk it will be much slower.

Not just gpus, the k20 was at 3.9 Tflops (fp32) and the new iPhone is at 4.3 (fp16). If you don’t need the precision it got passed by the phones

You’d get better perf training on a current gen phone than that gpu, but it probably functions

Are you one of the authors?

I feel like that blogpost was almost just ragebait for ai researchers. It goes between calling not including the +1 an error (which to me implies it would improve training losses, which it doesn't really https://news.ycombinator.com/item?id=36854613) and saying possibly it could help with some types of quantization (which could very well be true but is a much weaker statement) and the author provides basically no evidence for either.

My hypothesis for why this works that it mitigates the downsides of rope

to eli5:

rope is the modern strategy used to give information to the model about how far a query and a key are apart when doing attention. It's the best strategy we have now, but has a major downside, where it makes some connections between tokens that are far apart much stronger than you would like them to be. Xpos (https://arxiv.org/pdf/2212.10554) is another paper by microsoft tackling issues with rope and you can see figure 1 on page 4 to get a visual interpretation of the sinusoidal attention strength (you would like it to be smooth).

I think a big reason differential transformers is working so well, especially on long sequence stuff, because when both q1 and q2 don't match a token, the rope relative strength will still have the same value and the noise will cancel out. Leaving intended matches, but at the cost of somewhat dampening the original value rope brought.

Just a hypothesis though. It would be easy to test by running this experiment against a baseline where both use alibi attention (https://arxiv.org/pdf/2108.12409) which has a different set of tradeoffs this wouldn't mitigate, but still a really interesting result.

I think they mitigated the extra memory/compute from this by using half the number of overall heads and doubling V and O. Without actually checking the math I think it should be equivalent in flops, not counting the extra (cheap) multiply by const and subtract.

I wonder how much of the value here is from canceling out the positional noise rope produces. I would love to see a table comparing an alibi version of this to an alibi baseline in addition to the rope models here.

Crazy gains though congrats to the researchers

Their numbers are for a single input, I assume the throughput is much higher given the prices they are quoting and the cost of a single cs3.

On die communication isn’t free, a lot of things here are sequential and within matrix multiplies the cores have to transfer output and mem loads have to be distributed. It’s really fast but not like one cycle

*edit neverming below this is a character level model that probably has a small vocab so it wouldn’t make a massive difference

Is this taking into account the parameters in the embedding and the output ffn? Because normally when models are really small and the vocab is large they can account for an extremely large number of parameters and would explain why the optimal number of layers here is unusually small.

I suspect it isn’t being taken into account because doubling the embedding and cutting the number of layers in half only holds the parameters constant if you forget the embedding and output, but id need to see more details on the config (mainly the vocab size he used) to confirm.

Having played with this stuff its definitely spots in the expert buffers (the other comment in the thread has the link to explanation) and not the extremely small differences in floating point arithmetic. The effect from this is much much less than any change in quantization, i.e. almost impossible to see from the outputs.

jorbs manages to get 15 win streaks on a20 beating the heart. I wouldn’t call it a slot machine