phire

25 ms·

361.
by phire 2y ago
Yes, Deepseek introduced this optimisation of a common base "expert" that's always loaded. Llama 4 uses it too.
362.
by phire 2y ago
> It could have and should have. From a technical perspective, yes. You are basically talking about the Nokia N800 but with a cellphone modem and a bit of effort spent shrinking the bezels down. But from a product design perspective,
363.
by phire 2y ago
Actually, I'm wrong. AMD's deep dive [1] says register use is low during BVH traversal, so they are still using shared memory for the stack. According to the ISA doc [2] they have a special mode for the stack that overwrites older
364.
by phire 2y ago
BTW, I'd love to see a large model designed from scratch for efficient local inference on low-memory devices. While current MoE implementations are tuned for load-balancing over large pools of GPUs, there is nothing stopping you tuning
365.
by phire 2y ago
The main reason it's near impossible to bit-bang USB is that all devices to are required to use one of a few fixed clock rates (1.5 MHz, 12 MHz and 480 MHz), unlike SPI and I²C which allow variable/dynamic clock rates. If you simp
366.
by phire 2y ago
Yes... I could see that approach being desirable on a CPU, it would make all context switches cheaper. Are we re-inventing register windows? Of course, we are on a GPU here. Skimming through the RDNA 4 ISA document [1], looks like they expe
367.
by phire 2y ago
The feature was clearly designed for the ray-tracing use case, where the waves are independent and really shouldn't run into deadlocks as long as one wave can continue. That's why they have the deadlock avoidance mode, it guarante
368.
by phire 2y ago
It's not even per token. The routing happens once per layer, with the same token bouncing between layers. It's more of a performance optimization than anything else, improving memory liquidity. Except it's not an optimization
369.
by phire 2y ago
One of the earliest examples of "Artificial Intelligence" was a program that played tic-tac-toe. Much of the early research into AI was just playing more and more complex strategy games until they solved chess and then go. So LLMs
370.
by phire 2y ago
IMO, if there is any talk of "loyalty" in that first round of voluntary layoffs, you should always leave. Even if you were previously loyal, the company is changing and you won't like the results.
371.
by phire 2y ago
One more litetal translation of "ekranoplan" is "screen glider", so sea glider isn't that far away from it.
372.
by phire 2y ago
And when you think about c as the speed of causality, it becomes a lot more obvious why FTL travel/communicate is (almost certainty) impossible without breaking causality.... You are going faster than causality, of course it breaks. Th
373.
by phire 2y ago
"uninstall" points at a very specific type of exploit.
374.
by phire 2y ago
Not zero; In all bankruptcies, the shareholders get whatever is left over after liquidating all company assets and paying back all creditors. In most high-profile bankruptcies, there aren't enough assets to even finish paying creditors
375.
by phire 2y ago
This case is confusing because there were actually three sides. Wikimedia (and others) were arguing that the image was in public domain because animals can't hold copyright. PETA were arguing that monkeys should be able to hold copyrig
376.
by phire 2y ago
You know how YouTubers are always talking about how "the algorithm" didn't like this video, or loves that video. Or that "the algorithm" is a huge black block which nobody knows how it works. Youtube's "th
377.
by phire 2y ago
I'm not sure buying Traveller's Tales is a good idea. TT had a really good engine that allowed them to pump out Lego games at an impressive rate of 1-3 per year. They maintained that cadence from 2005 through to 2019, then they st
378.
by phire 2y ago
Arm64 has fast 128-bit loads. Not just with NEON, but with regular integer instructions, you can quickly load 128 bits into a pair of 64-bit registers. So it kind of makes sense to support fast shift by four. Though, it's more likely t
379.
by phire 2y ago
Traditional emulators also recompile games to the target device's native machine code. It just happens dynamically at runtime, hence the name "dynamic recompiler" or "Just In Time recompiler". This is just a static
380.
by phire 2y ago
I'm not entirely surprised, adding 40 bytes of SRAM to the 386 for a debug-only feature would have been hard to justify.
381.
by phire 2y ago
Mostly. ARM actually has instructions that allow a full 64-bit shift then add. But I checked the Cortex A78 optimisation manual. They take 1 cycle if the shift is 4 or less and 2 cycles in other cases.
382.
by phire 2y ago
There are some real gems in there. I especially liked the idea of CR2 and CR3 as scratchpad registers when memory access is really slow (386SX and cacheless 386DXs). And the trick of using ESP as a loop counter without disabling interrupts
383.
by phire 2y ago
It's an interesting performance comparison, the R4200 is well under half the size and actually has a higher IPC than the R4400 on integer code... when it hits the L1 cache (and both have 16KB L1i caches). The early R4400s and R4200 whe
384.
by phire 2y ago
Yeah, you are right... I've misunderstood radix-8 multiplication, missed that this post was only talking about a small part of the Pentium's multiplier. and jumped to conclusions... And annoyingly, hackernews doesn't allow
385.
by phire 2y ago
Obsolete. You can observe the evolution of multipliers in the MIPS line, which I've been studying, and happen to know. The R3000A had the same Radix-8 (aka 3-bit per cycle) integer multiplier in 1989. The R4000 had two Radix-8 multipli
386.
by phire 2y ago
Yeah... guaranteed after the teething issues. My understanding is that there were problems with the whole Open PDK, and that most designs would have run into similar issues even if Caravel wasn't preventing IO configuration. The didn&#
387.
by phire 2y ago
It's not a backdoor. Caravel is fully open source. You can audit it. There is no ROM (except for a project ID), just 1.5KB of RAM, a CPU, a few peripherals and (most importantly) the pad ring. Caravel is a (questionable) attempt to low
388.
by phire 2y ago
> The weirdest part is per-polygon mipmapping, which is not normal I suspect that per-polygon mipmapping is actually calculated on the CPU. That would mean the actual hardware doesn't really implement mipmapping, it just implement
389.
by phire 2y ago
Nvidia weren't chasing Sega. The NV1 was really just an evolution of Sun's GX graphics card, which did quad-based accelerated 3D in 1989 (though no texture mapping, GX was limited to flat-shaded quads, or emulating gouraud shading
390.
by phire 2y ago
It's an accounting simplification. The accountants want to say the item's book value is now $0 (because they can use the loss for a tax break). Destroying the item neatly proves this claim to any auditors, because if the wasn&#x27

More ›