@zeeshanp_ The frontier labs only have openings for a tiny minority of Nvidia employees, including in design and directly AI related roles. It is a massive company.
Ben Intel'deyken ilk kez mühendis olmayan bir CEO göreve geldi.
Intel’in ilk dört CEO’su tamamen teknik/mühendis kökenliydi:
Robert Noyce (fizik PhD)
Gordon Moore (kimya/fizik PhD)
Andy Grove (kimya mühendisliği PhD)
Craig Barrett (malzeme bilimi PhD)
Paul Otellini şirketin tarihinde ilk mühendis olmayan CEO’ydu. Ekonomi mezunu + Berkeley MBA’si vardı, kariyeri finans, satış ve pazarlama tarafındaydı. 1974’te Intel’e katılmış, uzun süre içeriden gelen bir isimdi ama teknik altyapısı yoktu.
Intel'in 2000'li yıllarda piyasa değeri 500 milyar doların üzerindeydi. Intel, Otellini döneminde "Biz 50 dolardan ucuza işlemci satmayacağız" dediği için iPhone çip fırsatını kaçırdı. Ardından da AI/GPU yarışını NVIDIA'a karşı kaybetti.
Demem o ki, bu Jensen Huang tipi adamların yerine Colette Kress gibi insanları koyarsanız şirket tepe taklak olur. Bu insanlar teknoloji de sevmiyor, mühendislik de sevmiyor. Bunlar ancak domates satsın, coca cola satsın. 3'e aldım 5'e sattım. Gerisine kafası basmaz.
It is hard to do justice to just how bizarre this story is.
Last year, the FT cried foul because Nvidia's disclosures attributed 22% of sales to Singapore. The FT and others questioned this figure as suspiciously implausible, given the size of Singapore's market. The explanation was simple. Multinational buyers used Singapore affiliates to centralize invoicing, while physical shipments went elsewhere. Nvidia explicitly disclosed this context in its FY2025 annual report.
In Q3, Nvidia switched to reporting revenue by customer headquarters, rather than billing address. This resulted in Singapore ceasing to be a material line item. Nvidia explained that, because billing addresses “often reflected a customer’s centralized invoicing location, even though our products were almost always shipped elsewhere,” reporting by customer headquarters “provides a better representation of the geographic profile of our revenue.” Nvidia also recast prior periods on the same basis.
Now, the Financial Times runs the false headline: “Nvidia stopped disclosing Singapore sales as revenues hit $104bn.” The sales still are disclosed but based on where customers are headquartered rather than location of accounts-payable departments. The same FT that faulted Nvidia for doing the latter rather than former, is now faulting Nvidia for doing the former rather than the latter.
What is even weirder is that Nvidia continued to disclose the prior Singapore-booked metric in the financial statements filed by its local Singapore subsidiary. That is how the FT got the figure that it claims is being concealed. So, in effect, the public markets are getting both, while the FT is making it sound like they are getting neither.
Finally, and perhaps most disingenuously, the article tells readers in paragraph four that Nvidia “previously declined to comment on the reasons for changing its reporting methodology,” only to say to add several paragraphs later “beyond its financial filings", while never quoting the straightforward explanation sitting inside those filings.
What a joke.
@not_ellington I am hearing from multiple companies that engineers are one-shotting simple analog and RF IC designs with Claude. It's only a matter time before manual labor in those areas is gone too.
@amitisinvesting The comeback plan? Dump the Anthropic shares at IPO and invest it all in short-dated call options of highly correlated tech companies.
@happydot1108 That sounds like some weird executive strategy stuff not applicable to most meetings. Try doing this for a design review and you will quickly learn it's counterproductive.
Increasingly bullish on compilers, hardware, and framework itself for maximizing inference
Need to optimize a “complex” kernel? opus5.5 + compiler and arch-aware harness + 5-8 hours + $1-2k budget
Need to improve the inference Pareto? setup above x10-15 + accuracy constraints + performance targets + roofline analysis + 3-6 days + $5-10k budget
We now can write custom inference engine per model from scratch in a couple days that outperform vLLM/SGLang in certain narrow use cases
I'm pretty sure I figured out the issue with the Harry Potter scene's bad color. TLDR: the HDR to SDR tonemap is incorrect, so the gamma is wrong. Here is the scene with the correct gamma. (No grading changes were made at all. Just a technical transform).
Ok, now for the long description: When you make a show in HDR, you need to properly tonemap the final grade to SDR in order to watch on an SDR device, or to share clips on social media (because most social media sites don't handle HDR well, or at all).
Tonemapping is when you take a high dynamic range video source, and remap the color and luma values to fit within a standard dynamic range container. You're essentially "translating" a large format into a lesser format.
HDR in cinema/streaming typically has a gamut of Rec.2020 and a gamma (transfer function) of ST.2084. This needs needs to be converted down to Rec.709 Gamma 2.4.
The Rec.2020 gamut got converted to Rec.709 just fine, which is why the hues look correct in the original clip. But somebody forgot to do the other half of the conversion, which would be to take the ST.2084 transfer function and convert it down to Gamma 2.4. This explains why the contrast looks completely wrong in the original clip. Here's what it looks like when you tonemap the gamma.
NOTE: There are multiple ways to tonemap an image. Some ways handle high brightness and high saturation values better than others, and give you tools to dynamically adjust those values (this is what Dolby Vision is, in case you didn't know). The tonemap method I used for this demonstration is just a simple global one though, so the saturated values of the lamps would probably look better using a more robust/dynamic method.
I have a counterexample on TPU v7x. In a matmul kernel, issuing matmuls in bursts triggers power-management throttling, which in turn stalls weight delivery. Spacing out the matmuls yields a 25.9% speedup over an issue-when-ready greedy scheduler and 26.4% over vanilla Pallas.
@Ortho_Dixie 530 AD is an entirely wrong analogy to the US today. The Gracchi brothers are a better historical parallel, particularly the descent of the Roman Republic into political violence that eventually led to its downfall.