06.08.2025-14:32:14 :: RIP Sleazy, we will miss you. Every man and every woman is a star now.


node:	06.08.2025-14:32:14
template:	4
parent:	HARDWARE :::
owner:	drakh
viewed by:
created:	06.08.2025 - 14:32:14

cwbe coordinatez:
101
63540
63541
1834976
9258189

ABSOLUT
KYBERIA

permissions
you:	r,
system:	public
net:	yes

som zvedavy co to bude zac

https://bolt.graphics/how-it-works/

a ako to zamiesa kartami ked tam bude rozumne implementovana CUDA

https://bolt-graphics.atlassian.net/wiki/spaces/EAP/pages/324468810/Zeus#Cuda

000001010006354000063541018349760925818909258985

George Hotz je presvedceny, ze to stoji a pada na SW, nie HW:

> I think the only way to start an AI chip company is to start with the software. The computing in ML is not general purpose computing. 95% of models in use today (including LLMs and image generation) have all their compute and memory accesses statically computable.

https://geohot.github.io/blog/jekyll/update/2023/05/24/the-tiny-corp-raised-5M.html

000001010006354000063541018349760925818909258310

Som zvedavy aj ja, ale to FP32 je nejake gramblave zatial.

000001010006354000063541018349760925818909258218

to je nejaka nova firma?

Żubr żuł żuchwą żurawinę

00000101000635400006354101834976092581890925821809258255

netusim co su zac.. kkz ale tie sloty na pamate ...

kebyze nvidia vendori take spravia kkz.. (ja viem nespravia, inak by nemali kazdy rok co nove predavat...)

0000010100063540000635410183497609258189092582180925825509258323

to bude pomale cez tie sloty. asi si to najde svoj ucel, ale cisty vypoctovy vykon neuzitocne

0000010100063540000635410183497609258189092582180925825509258311

Sloty su sice pekne, ale ako to bude fungovat na grafickej karte, ktora potrebuje super rychle pamete? Pre AI by to vsak mohlo byt fajn.

000001010006354000063541018349760925818909258218092582550925831109258327

no, ja dufam ze z toho dokazu vytrieskat aspon tolko ako apple zo svojej unified memory architektury ...

Trenovanie na tom by bol asi pain, ale inferencia z modelov ktore teraz dokazu natiahnut akurat tak "velke" nvidia clustre alebo jablka s vela pamate..

00000101000635400006354101834976092581890925821809258255092583110925832709258368

Kedze su to sloty na beznu so-dimm pamat tak o tom dost pochybujem.

0000010100063540000635410183497609258189092582180925825509258311092583270925836809258401

hej, pozeram akurat ze vdaka tomu SoC ma ta unified memory tak 10x vacsiu priepustnost..

Ale v porovnani s diskretnymi grafikami trebars od NVIDIE, ktore maju VRAM mimo cipu, ale soldered na doske tak aky je problem, to co je soldered spravit vo forme modulov (s tym istym clockom). co mi unika?

000001010006354000063541018349760925818909258218092582550925831109258327092583680925840109258421

GDDR7? jednak o par cisel vyssie nez DDR5 plus aj slotove DDR5 mali nizsi clock nez soldered LPDDR. zda sa mi, ze spociatku sa nieco vypisovalo aj o tom, ze v slotoch problem s tymi clockmi, ale nakoniec aj laptopy so slotmi pribudli, len to ma nizsiu frekvenciu (znizuje vykon iGPU)

000001010006354000063541018349760925818909258218092582550925831109258327092583680925840109258412

Keby sme nemali much profit oriented trh, tak samozrejme asi nie je problem aby GDDR existovala aj v standardizovanej slot verzi, ale takto Bolt asi nemal na vyber ak chcel ist touto cestou a jedina moznost bola DDR. Lebo na to neexistuje vytvoreny trh a nekupis to jednoducho niekde na Alze.