I don't think there's necessarily anything about this that points to problems at Intel. For now, I would tend to think that sticking with TSMC for large ASIC development is just easier (Intel is already using their tool chain). Obviously, given Raja's departure, there are significant problems at Intel Graphics other than node selection.I'm surprised Celestial is still on TSMC, doesn't bode well for Intel's process nodes.
I have no idea of what to think of Intel process nodes anymore, I just consider them unpredictable.I'm surprised Celestial is still on TSMC, doesn't bode well for Intel's process nodes.
Both a long way off apparently:
So that sounds like over a year until Battlemage and around 3 until Celestial.
Last time I saw a leaked GPU timeline from Intel there was also an Alchemist+ slated for Q3/Q4 2023. I hope the "+" means more than using screws instead of adhesive in GPU assembly. This may be enough to keep the ball running for them, in terms of being able to offer a competitive value oriented product until Q2/Q3 2024 when Battlemage is supposed to arrive.2H2024? That really doesn't bode well.
It's not like their competitors will be sitting on their laurels.
If they aren't using the product to fill their own fabs, then what's the point of an uncompetitive GPU? I just don't see why this makes sense for Intel.I don't think there's necessarily anything about this that points to problems at Intel. For now, I would tend to think that sticking with TSMC for large ASIC development is just easier (Intel is already using their tool chain). Obviously, given Raja's departure, there are significant problems at Intel Graphics other than node selection.
Given the long delay of Alchemist, it also suggests that they needed to reboot the design of Battlemage, since they are normally already working on the new generation before the previous gen reaches production.2H2024? That really doesn't bode well.
It's not just fab filler, they presumably want to enter the same professional markets that Nvidia is dominating.If they aren't using the product to fill their own fabs, then what's the point of an uncompetitive GPU? I just don't see why this makes sense for Intel.
Well, there was a bit of speculation when all the driver issues became apparent:I have no idea of what to think of Intel process nodes anymore, I just consider them unpredictable.
Last time I saw a leaked GPU timeline from Intel there was also an Alchemist+ slated for Q3/Q4 2023. I hope the "+" means more than using screws instead of adhesive in GPU assembly. This may be enough to keep the ball running for them, in terms of being able to offer a competitive value oriented product until Q2/Q3 2024 when Battlemage is supposed to arrive.
GPUs don't make sense for Intel in the short term. They never did. Despite the rosy projections needed to sell the idea to shareholders, it was always going to be a slog, IMHO. That said, yeah, in the long term these are Intel fabbed GPUs, or 1/2 of the business case is gone. Non-Intel fabbed gaming GPUs can still produce good profits, and that's not nothing.If they aren't using the product to fill their own fabs, then what's the point of an uncompetitive GPU? I just don't see why this makes sense for Intel.
It's worse than that. They;re supporting their main process competitor who's tightly integrated with their main product competitor.If they aren't using the product to fill their own fabs, then what's the point of an uncompetitive GPU? I just don't see why this makes sense for Intel.
Eh? Intel is not the one ponying up for development, Apple is.It's worse than that. They;re supporting their main process competitor who's tightly integrated with their main product competitor.
Eh? Intel is not the one ponying up for development, Apple is.
TSMC will still be making a healthy original profit on every wafer they sell to Intel, and that funds R&D.
His yet-to-be-named firm would either partner or acquire Yotta, which has data centers in the cities of Mumbai and Noida, said Koduri, whose aim is to challenge Nvidia Corp's (NVDA.O) grip on the digital movie and video game markets.
Jensen would dirty his trousers from laughing so hard at the idea of Raja making a dent in nvidia's grip.Jensen cowers in fear as Koduri sets his sights on eating Nvidia's lunch, AGAIN! https://www.reuters.com/technology/...ata-center-deals-ai-startup-india-2023-04-05/
That depends. If Nvidia has important employees in their Indian offices (not sure what goes on there), Raja could poach those employees.Jensen would dirty his trousers from laughing so hard at the idea of Raja making a dent in nvidia's grip.
That depends. If Nvidia has important employees in their Indian offices (not sure what goes on there), Raja could poach those employees.
As one of my cousins told me many years ago upon hearing a ridiculous utterance: "Well yeah, and monkeys could fly out of my butt."
Anything could happen, but I don't believe the lynchpin to NVidia's success resides in a hard-working, under-appreciated engineer in one of their Indian offices, whose talent can be spotted by Raja and harnessed to bring about the downfall of JHH and his company.
Those monkeys might even reproduce the works of Shakespeare after they emerge.
Jensen would dirty his trousers from laughing so hard at the idea of Raja making a dent in nvidia's grip.
It seems like a competitor would have an easy time getting more AI productivity out of a chip that has only AI cores. How much of the NVidia GPU dies Transistor budget going to non-AI? I'd bet it's probably over 75% non AI. A competitor could build an AI chip half the size and devote all the transistors to AI and have much more AI capability.
It seems that Nvidia's strategy is to tout the generic abilities of their cards, so you can do all the pre- and post-processing on the cards, rather than require other hardware to do that.
Jim Keller seems to think that there is room for different kinds of solutions for different use cases, and I wouldn't bet against him.
@guidryp
Most AI processing is going to be in running the models, not training them. Then you need to get the data in a format suitable for the model. For example, you might need to do image processing to detect relevant elements of an image, or for chatgpt, you need to transform a text into tokens, etc. Then afterwards, you may need to convert the result, or even just render it, if the output is an image.
For training you need to process huge amounts of data, so even though you won't have the postprocessing step, you'll still have to do preprocessing to get the data in the right format. Depending on the kind of preprocessing, a GPU may be far faster and more efficient at it than a CPU.
See:
End-to-End AI for NVIDIA-Based PCs: An Introduction to Optimization | NVIDIA Technical Blog
This post is the first in a series about optimizing end-to-end AI. The great thing about the GPU is that it offers tremendous parallelism; it allows you to perform many tasks at the same time.developer.nvidia.comFast AI Data Preprocessing with NVIDIA DALI | NVIDIA Technical Blog
Editor’s Note: This post has been updated. Here is the revised post. Training deep learning models with vast amounts of data is necessary to achieve accurate results. Data in the wild…developer.nvidia.com