tedd4u

Yes, the "neural accelerators" in the M5 GPU cores really do 4x pre-fill performance over M4. You can find benchmarks online since M5 has been out for months now. I assume the M6 has whatever the next-gen version of those is. These are not the "neural processors" that have been there since M1 (those still exist) but are additionl matrix math accelerators inside the actual GPU cores. Like "tensor cores" on Nvidia GPUs.

Look at the "Q4_0 Prompt" column. That's the tokens-per-second processing the initial prompt/system prompt. This is where the "neural accelerators inside each GPU core" is seen most prominently.

Agreed, seems like the M5 has already made steps in that direction, with 4x prompt processing / prefill performance vs. M4. [1] And token generation also got a 10% boost. The data below is only for Pro & Max but I think the base M5 got the same relative boosts vs. M4 base.

    Chip         BW (GB/s)   GPU Cores   Q4_0 Prompt   Q4_0 Gen
    M4 Pro (20c)    273         20          439.78        50.74
    M4 Max (40c)    546         40          885.68        83.06
    M5 Pro (20c)    307         20      ~1500 to 1700    ~56
    M5 Max (40c)    614         40      ~3000 to 3500    ~92

[1] https://www.hardware-corner.net/m5-pro-m5-max-local-llm-4x-f...

Regarding the $25 price per GB RAM. These chips use LPDDR5X RAM. Looking at the Framework site, they are selling LPDDR5X LPCAMM2 modules for the following prices:

    16GB   $239 - $15/GB 
    32GB   $800 - $25/GB
    64GB $1,600 - $25/GB

The RAM in the M5 is LPDDR5X, not HBM. You're right of course that it's unified and used as VRAM in that sense.

Look, they are offsetting RAM-driven price increases: "In the U.S., pricing is now set at $9, down from $19 for the previous version of the cloth."

And AI

What would you propose as an alternative? Some concerns I have are: 1) my house not burning down due to electrical fire and 2) not exploding due to natural gas leak.

9-Aug-2026: SK Hynix to invest $38 billion building new memory chip plants as demand soars [1]

"In the AI era, technological competitiveness alone is not enough and the ability to supply the required volume at the exact moment customers need it is the ultimate competitive advantage," SK Hynix said in a press release. "We reached this investment decision after a thorough review of market demand."

[1] https://www.msn.com/en-us/money/general/sk-hynix-to-invest-3...

FYI it takes less than 30 seconds do clone, build, and run.

I wonder if this is at least partially motivated by Anthropic's need to know what training materials are themselves Claude-generated.

1 •--||--• 2

Figure 1 block diagram

Seems like, here in the 21st century, one could reasonably have a model of the plant that tests could be executed against. Anything like that exist?

Oof, that's bad. Iran got burned real bad by Stuxnet, you know they would love to score big with PLC sabotage.

The joke about KFC's "11 secret herbs and spices" is that they are

• Salt

• Pepper

• MSG

Posted many times. Here's the most recent time from 2024 (also has most discussion).

https://news.ycombinator.com/item?id=42325011

That one I know: it's 750ml

As long as you also know a gallon is 4.0… quarts

"It was produced by ISO Technical Committee 34 (Food products), Sub-Committee 8 (Tea)."

Here's another option, comparison to jpeg2png at the bottom of the README.

https://github.com/ilyakurdyukov/jpeg-quantsmooth

Ukraine fighting Iran has a benefit for Ukraine as well in terms of US perception ... the enemy of my enemy is my friend. Makes it harder for MAGA types like Tucker Carlson to sell the idea that actually Russia is the US's friend.

To build this (on macOS 15) I also had to do this (in Package.swift):

    @@ -4,8 +4,7 @@ import PackageDescription
     let package = Package(
         name: "TurboFieldfare",
         platforms: [
    -        .macOS(.v26),
    -        .iOS(.v26),
    +        .macOS(.v15)
         ],
         products: [
             .library(name: "TurboFieldfare", targets: ["TurboFieldfare"]),

I went back to 15 after accidentally upgrading to 26 because on MBA 13" screen the new UI design uses a lot of extra padding everywhere wasting screen space which is already at a premium (especially vertical). Hoping 27 fixes a lot of these issues.

And, “TSMC and ASML, currently, the only supplier of tools that could make cutting-edge 2nm chips, have added a ‘kill switch’ that will disable TSMC’s EUV machines remotely. According to Bloomberg.”

https://www.tomshardware.com/tech-industry/tsmcs-euv-machine...

And one cool trick is that if one develops step function more efficient training, one can still “open source” the model without revealing the training techniques.

True, but it's not all roses. If you like to sit at the front of economy to exit as fast as possible when arriving, you will be sitting next to a heavily-used lav which is not enjoyable. You have to sit 4-6 rows back from the lav to avoid the annoyance.

Consider the 7600x3d also if you're in the market for something like this. It has a slightly higher base and boost clock, and half the TDP. It's 6 cores instead of of 8 but if you're building a gaming machine that may not make a big difference. If you're putting it in a small form factor the lower TDP might be helpful.

The corps ruled the US until Trump. Now it’s moving to an oligarchical / kleptocratic mode like Russia. Sure the corps are still involved, but not on top. Will they remain after the current administration? I think it will be hard to put the genie back in the bottle.

Should we compare release date to release date? Or training date to training date?

An article exploring research and citations of "first sleep" and "second sleep"

https://www.patheos.com/blogs/geneveith/2013/08/first-sleep-...