English
Andrei Stan
378 posts
@Orduh @Keir_Starmer Human has evolved to be intelligent and have dialogue and diplomacy!
English
I condemn in the strongest terms the overnight Iranian strike on a Qatari gas facility.
We are working towards a swift resolution to the situation in the Middle East, in the best interests of the British people – because there is no question that ending the war is the quickest way to reduce the cost of living.
English
English
@ayzddzya @Keir_Starmer Iran should have attacked Israel instead. Is the logic not clear enough?
English
@__tinygrad__ I have thought about this for a long time, what is it actually forwarding, the TLPs? How does BAR mapping work in this case
English
If we can tunnel PCIe over USB3, why not also tunnel it over Ethernet? extra/remote/serve.py on remote, REMOTE=<ip> on local.
English
@sama All extremely complex software is written character by character.
English
I have so much gratitude to people who wrote extremely complex software character-by-character. It already feels difficult to remember how much effort it really took.
Thank you for getting us to this point.
English
@bubbleboi Cya at the next earnings call, tho i am a Intel perma bull, hope Jensen stops jestermaxxing around and gets to the real stuff
English
I sincerely think Intel will overtake Nvidia’s market cap in the next 5 years.
English
I’m hedged against everything.
Oil prices.
Interest rates,
My gf cheating on me.
It’s mathematically impossible for me to lose.
English
@NLwartracker This will be absolutely crushed by mostly 3d printed conventional designs that can be deployed by the 10s of thousands. The problem isn't the drone is the whole supply chain., which has to be onshored, not dismissed to Shenzhen
English
@NLwartracker This is at most useless, future non nuclear wars will be all about speed, cost and adaptability. What you have here is over complicated drone with carbon fiber and aluminium frame, complicated fittings, great for the eyes of finance and c-suite people.
English
"Spotted at BEDEX 2026 (Brussels): Thales' kinetic interceptor drone. Max speed: 360 km/h. Targets Shahed-class UAVs via sapient-compliant radar. Fuselage: titanium-reinforced carbon fiber. Brushless motors (low KV 400-800) with optimized props for top speed. Pure ram-kill using titanium nose. #CUAS #Thales #BEDEX2026
English
@elonmusk @atomic_semi This will result in $INTC $ASML and maybe as wild as it sounds even $TSLA fall. The latter due to the market not beliving in Elon executing fast enough in a this extremely competitive market.
English
Prediction, @elonmusk will unveil Terafab using @atomic_semi hardware, it will be a low yield process but that won't matter as models already already lossy and AI is "bit flip" resistent. Probably Elon will even acquire them.
SuspendedCap@ContrarianCurse
This might be the stupidest thing Elon has ever tried. Kiss that 20b goodbye
English
@SemiAnalysis_ @MediaTek Why does only the 6t have this problem?
English
TSMC spent billions on the nanosheet transition for N2 and SRAM bitcell area barely moved. But it’s even worse than that – as @MediaTek showed at ISSCC, the N3E High Current bitcell actually regressed vs N5. N2 doesn’t fare much better, barely improving.
Two full nodes of going nowhere. While logic density continues marching forward, all the caches, register files and on-chip memory that limit real chips? Flatlined.
Even Intel, with their struggles, managed to shrink by 0.77x with 18A!
English
Photonics stocks in layman's terms
$TSEM = $TSMC
$LITE = $NVDA
$COHR = $AMD
$BESI = $AMAT
$XFAB = $INTC
$AAOI = $MU
$POET = Groq
$ALMU = Cerebras
$AXTI = $SNDK
$CIEN = $CRDO
$SMTC = $ALAB
$SIVE = $LITE Lite
English
These Pakistani men are the victims of horrible Islamophobia.
Western Invasion@WesternInvasion
🚨🇬🇧Pakistani Miah brothers r*ped two 7-year-old boys in a mosque. They drugged and prostituted 30 underage school girls. They cut off the toes of two girls who tried to escape. The court found a total of 62 crimes. This happened in Leeds UK. Illegal immigration continues to destroy UK.
English
@FelixCLC_ Not only double the bw but also add a constant to it
English
@imnotharsh @jukan05 > ARM-based limits of 8 GPUs to full 72-GPU NVL72-like coherence.
The GH200/GH300 NVL72 have full "coherence". Grace is linked to NVLink network with C2C and can access any GPU in the nvlink domain
English
With $NVDA GTC 2026 coming up, there are some expectations of what they announce with $INTC Intel.
@jukan05 mentioned sell-side analysts believe they will announce the x86 CPUs here.
I have some thoughts: Nvidia’s strategic equity stake of $INTC Intel stories its push for x86 and positions it for Agentic AI workloads and scaling AI infrastructure.
x86 architecture enters Nvidia’s ecosystem where Intel will design and manufacture custom x86 CPUs specifically tailored for Nvidia’s AI platforms, enabling tight integration with NVLink for hyperscale/enterprise AI systems.
The core driver for x86 adoption is the shift toward agentic AI. x86's vast software ecosystem and superior multi-threading/compatibility make it far more practical than $ARM (Grace) for the emerging workloads, avoiding ecosystem fragmentation.
Intel likely fabricates Nvidia’s-custom x86 server CPUs (likely Xeon-derived or specialized variants) that Nvidia integrates directly into its rack-scale AI supercomputers (e.g., expanding NVLink domains beyond current ARM-based limits of 8 GPUs to full 72-GPU NVL72-like coherence).
x86 unlocks broader hyperscaler adoption (many still prefer x86 for legacy/enterprise software) and addresses CPU bottlenecks seen in Blackwell deployments, while future Rubin/Feynman (which is about to get announced next week, probably) generations are expected to increase CPU core counts and support x86 options.
On the client side, the partnership yields Intel x86 SoCs with integrated NVIDIA RTX GPU chiplets (branded as "Intel x86 RTX SoCs"), targeting notebooks and consumer PCs for better AI-accelerated personal computing, but the data-center custom CPUs remain the bigger strategic win.
—————————————-
Ultimately, Agentic AI requires CPUs:
- Orchestration, planning, tool calling, API handling, memory management, and sub-agent coordination are CPU-intensive (not GPU-parallelizable).
- GPUs handle model inference; CPUs manage everything else in complex, branching workflows.
English
@jukan05 Where they even doing asic? I thought they wanted to use ai6 for training as well
English
@racetrack275 @jukan05 No way nvidia goes to Ethernet scale-up. The moat is too big, they will showcase nvlink nics for scale-up scale-out bridge
English
@jukan05 Yes it’ll be part of Feynman reveal.
Ethernet in scale-up options too.
Nvidia will talk a lot about TAM expansion into wider enterprise/sovreign since hyperscale capex growth will naturally be more limited from here.
For those you need x86 and wider interoperability.
English
@jukan05 I think is happening faster then expected, probably already in the fab. Intel allocated fab capacity to it thats why they came out surprised that they did not anticipate demand in q4
English
@SemiAnalysis_ How many GPUs are running the inference on NVL72 in this case? Is it NVL36x2 or NVL72?
English
At 167 tok/s/user interactivity on Deepseek 670B MoE at 8k context length, it would cost $0.96 per million output tokens on GB200 NVL72 FP4 verus $2.3 per million output tokens on B200 even with DeepSeek system optimizations like disaggregrated PD & wide EP enabled.
English
@__tinygrad__ Solar panels are made in China and are heavily tarrifed. Elon said it in his latest apperance on Dwarkesh
English
The price and difficulty of getting electricity in America is infuriating. Why aren't we building this? I nominate Arizona to be turned into a huge solar panel.
English