phire

54 ms·

661.
by phire 3y ago
Not really accurate. The switch from NRZ to PAM4 actually massively increased the bit error rate. They switched away from the 8b/10b style line code and replaced it with forwards error correction. PCIe 6.0 uses 256 Byte frames, with 24
662.
by phire 3y ago
My understanding: Delay slots only make sense when you don't have a branch predictor. I get the impression that many people massively underestimate how powerful branch prediction is. Even a simple 2 bit saturating counter will correctl
663.
by phire 3y ago
You could actually run a small chess engine a few moves ahead to sort the full list of legal options into the order of obviousness. Then you can use a variable length encoding (Huffman coding?) to encode the moves, with more obvious moves t
664.
by phire 3y ago
Register Windows were a very targeted solution to the problem of spilling temporary registers across function calls, mostly the demand that it put on the memory subsystem. This extra hardware only helped that one problem, which is pretty wa
665.
by phire 3y ago
Yes, you could probably fine-tune (or even zero-shot) a LLM to handle the "knowing when to jump in" use case. The real problem is that it's simply too computationally expensive to continually feed audio and video into it one
666.
by phire 3y ago
The "magic" of the fake Gemini demo was the way it seemed like the LLM was continually receiving audio + video input and knew when to jump in with a response. It appeared to be able to wait until the user had finished the drawing,
667.
by phire 3y ago
I agree that the "forcing valid json output" is super cool. But it's unrelated to the problem of LLM hallucinations. A hallucination that's been validated as correct json is still a hallucination. And if your problem spa
668.
by phire 3y ago
It's been 3.5 years since GPT-3 was released, and just over a year since ChatGPT was released to the public. If it was possible to solve LLM hallucinations with simple Chain-of-Thought style agents, someone would have done that and rel
669.
by phire 3y ago
There is no 12v starter motor, it always just uses one of the large motor-generator units to start the engine, and in the Prius drive chain, they require 600v (boosted from 200v from the HV battery). So if the HV battery is flat, you can&#x
670.
by phire 3y ago
Yeah, I started wondering the same thing as I was writing that response, as I do actually own a Lexus with the Prius drive chain. For starters, it would be pretty hard to get into that state in the first place. My understanding is they actu
671.
by phire 3y ago
I'm ok with "edited for latency" or "only showing the golden path". But the most impressive part of the demo, was the way the LLM just seemed to know when to jump in with a response. It appeared to be able to wait u
672.
by phire 3y ago
No. For safety, hybrids and electric vehicles physically disconnect the high-voltage battery from everything when it's off with massive relays. So you can end up in the somewhat stupid situation where your high-voltage battery is fully
673.
by phire 3y ago
It was pretty blatantly obvious from just the video. The engine didn't just stop producing power, that would leave typically the propeller spinning, windmilling in the air. (unless the engine seized, but he was faking fuel starvation)
674.
by phire 3y ago
The detachment mechanism adds significant complexity, weight and costs. And the result is much less accurate, and the warhead is smaller. Instead of actively flying the warhead as close to the target as possible, the drone now needs to calc
675.
by phire 3y ago
This release added cosmo_dlopen, which works on many platforms (but not all). Should be enough to allow you to find the vulkan loader and dlopen that. I'm not sure what you would do on mac, statically link moltenvk? Or maybe it would
676.
by phire 3y ago
Yes, but a normal hybrid only has a single mile of range. Much, much lower capacity than your typical PHEV.
677.
by phire 3y ago
Don't forget, the CPU needs to know the current PC for relative branch instructions too. Relative branch instructions and calls are pretty common, so flushing the entite pipeline just to get the current PC would be way too expensive. S
678.
by phire 3y ago
All 32bit/64bit x86 machines. I believe you can still buy new 8086 class and 286 class 16bit x86 cores in random SoCs. And I know you can buy SoCs with 386 class x86 cores. It might be more accurate to describe them as 486 class cores
679.
by phire 3y ago
On the plus side, every day of the month would always be the same day. The 1st would always be a Sunday, the 25th would always be a Thursday, etc.
680.
by phire 3y ago
So... it turns out anyone can file a lawsuit for any reason. Doesn't mean it's true. This lawsuit was later dropped, so it seems to have been without merit. And it was part of a power struggle that happened on the Uber board after
681.
by phire 3y ago
She pled guilty to 7 counts. 1 Count of actual wire fraud and 6 counts of various "Conspiracy to commit wire fraud, commodes fraud, securities fraud and money laundering". They have a total maximum sentence of 110 years. Prosecuti
682.
by phire 3y ago
Sam didn't create a bank. If he did, this wouldn't have been fraud and he wouldn't be going to jail. All documentation, marketing, interviews, statements, tweets, general understanding, and government licensing had FTX implem
683.
by phire 3y ago
No. Both Uber and WeWork were "gambling" investor money, which is totally allowed and it's what those investors were expecting the companies to do with that money. There was never a chance of a "run on the bank" bec
684.
by phire 3y ago
Probably, but the A100 BIOS probably can detect it is running on something other than a A100 and bail. The lack of memory would be the most obvious difference. The A100 has 80GB, this has 8GB. And I really suspect Nvidia probably some way o
685.
by phire 3y ago
Driverless taxies could actually be an integral part of a good public transit system. Think about it as last-mile public transit. Shared driverless taxies (or call them dynamically-scheduled self-driving mini-buses) that take groups of peop
686.
by phire 3y ago
It's only unconditional jumps that might have NOPs following them. For conditional jumps (which are pretty common), the extra instructions in the packet will be executed whenever the branch isn't taken. And instruction scheduling
687.
by phire 3y ago
But if you don't have a uop cache (which is quite feasible with a RISC-V or AArch64 style ISA), then decode bandwidth is much more important than a few NOPs in icache. Presumably your high performance core has at least three of these 6
688.
by phire 3y ago
I think if you have 64bit packets, you might as well align jump targets to the 64bit boundary. I'd rather have an extra nop or two before jump targets than blindly throw 1-3 instructions worth of decoding bandwidth on jumps (which are
689.
by phire 3y ago
It doesn't matter how optimised the length decoding is. Not doing it is still faster. For an 8-wide or 10-wide design, the propagation delays are getting too long to do it in all in single cycle. So you need the extra pipeline stage. T
690.
by phire 3y ago
One advantage of just sticking with only 32bit instructions is that nobody needs to write packet-aware instruction scheduling. Even with decent instruction scheduling, you are still going to end up with a bunch of instruction slots filled w

More ›