Nvidia shows signs in [2023]

[trinibwoy]

Performance of 4800 what? This has to sting. China is a huge market and it doesn’t look like the US will give up the trade war anytime soon.

These trade restrictions apply to Nvidia. What’s stopping some company in India from buying GPUs and reselling them to China?

 

[pharma]

Performance of 4800 what? This has to sting. China is a huge market and it doesn’t look like the US will give up the trade war anytime soon.

These trade restrictions apply to Nvidia. What’s stopping some company in India from buying GPUs and reselling them to China?

I think they meant RTX 4080.

 

[pharma]

Apple investirait des milliards de dollars pour muscler son IA générative

Apple aurait l'intention d'investir des milliards de dollars pour prendre toute sa place dans la course à l'intelligence artificielle générative. Et cela passe par une commande massive de serveurs spécialisés pour les calculs nécessaires au modèle de langage maison.

www.01net.com

October 24, 2023

Apple reportedly intends to invest billions of dollars to take its rightful place in the race for generative artificial intelligence. And this requires a massive order of specialized servers for the computations necessary for the in-house language model.

According to Ming-Chi Kuo, who is usually well-versed in Apple's behind-the-scenes, the company intends to acquire 2,000 to 3,000 AI servers this year, and 18,000 to 20,000 units in 2024! This represents 1.3% and 5% of global deliveries, respectively. Apparently, Apple has done what all other companies have done and set its sights on Nvidia servers equipped with specialized HGX H100 GPUs.

At list price, a unit costs $250,000: Apple could therefore gobble up $620 million in 2023, and $4.75 billion in 2024! We can imagine that in view of the order, the manufacturer will get a small discount, but it's not even certain: demand is very strong and Apple is following other companies that were quicker to pull out the checkbook.

Even if the volumes are impressive, Apple's AI computing capabilities will still be a notch below those of Meta, which plans to buy 40,000 servers next year. Ming-Chi Kuo indicates that Apple could develop its own servers, but there is no visibility on this project at this time.

These servers will be used to train the company's large language model (LLM), called Ajax, which will support Apple's future generative AI services and features. Siri could be one of the first to take advantage of this, but it will be several months before we see the first effects.

 

[Lurkmass]

Performance of 4800 what? This has to sting. China is a huge market and it doesn’t look like the US will give up the trade war anytime soon.

These trade restrictions apply to Nvidia. What’s stopping some company in India from buying GPUs and reselling them to China?

The US government will probably want them to implement a form of online authentication system in order for other foreign to be able to utilize any sensitive technology. It's either that or the US government proceeds to block other countries who won't cooperate with their export control regime which would be even more catastrophic ...

 

[DavidGraham]

Performance of 4800 what?

Aggregate TOPS x bit length of 4800 or more. TOPS x bit length of 4800, described as “Total Processing Performance” (TPP). This means that chips such as Intel’s Gaudi 2, Gaudi 3, AMD’s MI250X, and MI300 are all blocked as well.

There is also a new performance density threshold too. Performance density is TPP divided by die area. This prevents shipments of chips with smaller die size that have less absolute compute power alone but are still dense/efficient from a computing standpoint. These figures stand at 5.92 for an absolute ban and 3.2 for license. They also come with a few tiered performance levels as well.

Wafer Wars: Deciphering Latest Restrictions On AI And Semiconductor Manufacturing

China's Countermove: How Beijing is Dodging New Semiconductor Restrictions

www.semianalysis.com

 

[DavidGraham]

https://twitter.com/x/status/1717338591037390862

 

[Kaotik]

https://twitter.com/x/status/1717338591037390862

So the news of them entering PC cpu market is just the old MediaTek collab?

 

[pharma]

So the news of them entering PC cpu market is just the old MediaTek collab?

Working with a company that has 36% global market share of the smart phone market is a smart move.

According to a report from DigiTimes Asia, sources outline that they expect MediaTek to “integrate an Nvidia GPU into its next-generation flagship mobile processor.” The report adds that these mobile processors can land in the market by as early as 2024. If this stands to be true, it could mean that smartphones can soon feature Ada Lovelace or the rumored Blackwell GPU architecture.

While the pc collaboration is "old" news, if true Nvidia won't have to wait until the end of 2024 to enter the ARM PC master race. In 2025 they will be able to enter with their own chip, adding the experience of working with MediaTek to the knowledge gained from running ARM chips (Nvidia Tegra 3/4) on Windows RT (2012).

Good times ahead. Will be an interesting to see if anything develops from these ventures.

 

[DavidGraham]

A new round of AI benchmarks for consumer GPUs. In OpenAI Whsiper, the 2060 is faster than 7900XTX. In Stable Diffusion, the 2080Ti is faster than 7900XTX.

RX 7900 XTX > RTX 4060??? AI 그림/자막 그래픽카드 벤치마크

퀘이사존 AI 소프트웨어 50종 그래픽카드 벤치마크 지난 2022년. 오픈AI의 챗GPT(ChatGPT…

quasarzone.com

 

[pharma]

A new round of AI benchmarks for consumer GPUs. In OpenAI Whsiper, the 2060 is faster than 7900XTX. In Stable Diffusion, the 2080Ti is faster than 7900XTX.

RX 7900 XTX > RTX 4060??? AI 그림/자막 그래픽카드 벤치마크

퀘이사존 AI 소프트웨어 50종 그래픽카드 벤치마크 지난 2022년. 오픈AI의 챗GPT(ChatGPT…

quasarzone.com

English translation errors (MS Edge) in their conclusion but think the Korean language conclusion is correct ...

Edit: Interesting that a WebUI version (uses CUDA) gives faster performance on an RTX gpu than running the C++ version.

 

[DegustatoR]

NVIDIA RTX 40 SUPER rumored specs emerge, RTX 4080 SUPER with full AD103 GPU and 10240 CUDA cores - VideoCardz.com

NVIDIA RTX 40 SUPER gets first CUDA core spec rumors Initial specs for RTX 40 SUPER from ‘several weeks’ ago. This week has been filled with a surge of RTX 40 SUPER speculations, culminating in detailed specifications shared by a prominent NVIDIA insider, Kopite7kimi. The disclosed information...

videocardz.com

Nvidia seem to be refreshing the lineup between 4060Ti and 4090 - as a result of 7800XT launch and recent 7900 retail price changes probably.

 

[pjbliverpool]

NVIDIA RTX 40 SUPER rumored specs emerge, RTX 4080 SUPER with full AD103 GPU and 10240 CUDA cores - VideoCardz.com

NVIDIA RTX 40 SUPER gets first CUDA core spec rumors Initial specs for RTX 40 SUPER from ‘several weeks’ ago. This week has been filled with a surge of RTX 40 SUPER speculations, culminating in detailed specifications shared by a prominent NVIDIA insider, Kopite7kimi. The disclosed information...

videocardz.com

View attachment 9915

Nvidia seem to be refreshing the lineup between 4060Ti and 4090 - as a result of 7800XT launch and recent 7900 retail price changes probably.

Oh man if that 4070 Ti Super spec is true I'd be sorely tempted to upgrade if could get a decent price for my existing 4070Ti.

 

[Flappy Pannus]

NVIDIA RTX 40 SUPER rumored specs emerge, RTX 4080 SUPER with full AD103 GPU and 10240 CUDA cores - VideoCardz.com

NVIDIA RTX 40 SUPER gets first CUDA core spec rumors Initial specs for RTX 40 SUPER from ‘several weeks’ ago. This week has been filled with a surge of RTX 40 SUPER speculations, culminating in detailed specifications shared by a prominent NVIDIA insider, Kopite7kimi. The disclosed information...

videocardz.com

View attachment 9915

Nvidia seem to be refreshing the lineup between 4060Ti and 4090 - as a result of 7800XT launch and recent 7900 retail price changes probably.

Somewhat disappointing if true, as no potential bus/vram upgrade for the 4070 - vram upgrades are really what the midrange needs, not necessarily performance.

 

[DegustatoR]

Somewhat disappointing if true, as no potential bus/vram upgrade for the 4070 - vram upgrades are really what the midrange needs, not necessarily performance.

4070 Ti Super using a mix of AD102 and AD103 chips imply that it will get >192 bit bus and thus 16GBs are likely.
4070 will probably fall down to $500 so its 12GB will be less obnoxious.
4070S though will remain with 192 bit and I doubt that it will get >12GBs.

 

[Kaotik]

Nvidia shares drop after report of canceled China orders

Shares of Nvidia Corp dropped by about 5% to a near five-month low on Tuesday following a Wall Street Journal report that the artificial intelligence (AI) giant may be forced to cancel up to $5 billion worth of advanced chip orders to China in compliance with new U.S. government restrictions.

www.reuters.com

NVIDIA shares took a 5% dip when it was revealed that the company is looking at cancelling 5 billions worth of orders to Chinese companies next year due sanctions.

Meanwhile Korea's top search portal Naver has switched from NV GPUs to Intel CPUs in their AI servers for their Naver Place map service due price hikes & chip shortages

Naver replaces Nvidia GPU with Intel CPU for its AI map app server - KED Global

South Korea’s web portal giant Naver Corp. has replaced the main chip supplier of its artificial intelligence server for its map service, Naver Place, fr

www.kedglobal.com

 

[Father_Murphy]

An interesting interview with the CEO of Cerebras, who clearly has no love lost for Nvidia.

https://www.hpcwire.com/2023/10/30/cerebras-ceo-andrew-feldman-sounds-off-on-nvidias-roadmap-and-chiplets/

He is not an Nvidia fan for sure. I chuckled a bit at his insinuation that Nvidia's (and others' in truth) approach to silicon design and AI/HPC workloads lacks inventiveness.

He also contends Nvidia's recent road map announcement was "predatory" and likened the announcement to Cisco in the 90s trying to buy time to fend off competitors. There's probably some truth to that, though only time will tell.

There's more, it's a good read. Sorry if it has been posted already, I haven't seen it.

 

[TopSpoiler]

Meanwhile Korea's top search portal Naver has switched from NV GPUs to Intel CPUs in their AI servers for their Naver Place map service due price hikes & chip shortages

I understand that Nvidia's GPUs have issues with price and lead time.. but why CPU server and not Gaudi?

 

[trinibwoy]

Nvidia shares drop after report of canceled China orders

Shares of Nvidia Corp dropped by about 5% to a near five-month low on Tuesday following a Wall Street Journal report that the artificial intelligence (AI) giant may be forced to cancel up to $5 billion worth of advanced chip orders to China in compliance with new U.S. government restrictions.

www.reuters.com

NVIDIA shares took a 5% dip when it was revealed that the company is looking at cancelling 5 billions worth of orders to Chinese companies next year due sanctions.

Meanwhile Korea's top search portal Naver has switched from NV GPUs to Intel CPUs in their AI servers for their Naver Place map service due price hikes & chip shortages

Naver replaces Nvidia GPU with Intel CPU for its AI map app server - KED Global

South Korea’s web portal giant Naver Corp. has replaced the main chip supplier of its artificial intelligence server for its map service, Naver Place, fr

www.kedglobal.com

The China ban is a problem but pissing off customers in other countries is probably the bigger risk. Nvidia’s chip shortages and high prices will encourage people to invest in other ecosystems that are cheaper and more readily available even if they aren’t technically better than Nvidia’s stuff.

 

[DavidGraham]

Meanwhile Korea's top search portal Naver has switched from NV GPUs to Intel CPUs in their AI servers for their Naver Place map service due price hikes & chip shortages

This is in the future and for Inference only, right now they are buying as many NVIDIA GPUs as they can for AI training.

https://twitter.com/x/status/1719496546952077788

Naver, Intel team up to replace Nvidia GPUs with CPUs for AI inference

Naver, a well-known South Korean AI company, has reportedly teamed up with Intel to develop solutions that can replace Nvidia's AI accelerators, which are essential components for improving AI services. The move comes as AI developers struggle to access Nvidia GPUs due to supply shortages and...

www.digitimes.com

 

[troyan]

Dont like the regulation of AI? Buy your own sea plattform with 10k H100 GPUs:

Del Complex

Our mission is to accelerate human potential through the symbiosis of AGI, neural prosthetics, robotics, clean energy, resilience solutions, and fundamental scientific research.

delcomplex.com