News Intel GPUs - Falcon Shores cancelled

[NTMBK]

 

[darkswordsman17]

I feel like I'd expect the curve to be trending upward more than tapering off for dGPU, as far as performance being the vertical axis. And perhaps that's more for the integrated/entry stuff, with the big block (pushing higher and also further on the timelines) indicating substantial improvements as they move up in tiers, but that will happen over time (so the initial stuff will be integrated/entry).

 

[Hitman928]

I feel like I'd expect the curve to be trending upward more than tapering off for dGPU, as far as performance being the vertical axis. And perhaps that's more for the integrated/entry stuff, with the big block (pushing higher and also further on the timelines) indicating substantial improvements as they move up in tiers, but that will happen over time (so the initial stuff will be integrated/entry).

That's how I read it as well.

 

[zinfamous]

I love charts without real axis. Convinces me that the person showing me the chart is just trying to blow smoke up my cornhole.

 

[Hitman928]

I love charts without real axis. Convinces me that the person showing me the chart is just trying to blow smoke up my cornhole.

Lol, I share your disapproval of arbitrary axes, but here I don't know if it's so much blowing smoke as it is not ready to reveal. We're still over a year a way from launch and this graph makes it seem that launch will only be integrated/entry level solutions to begin with as well.

The only thing this graph really tells us is that they plan on scaling their architecture up to cover from integrated to enthusiast to DC markets. I don't think that was known before now, was it?

 

[Despoiler]

I'm reading their graph that the whole stack is represented within the Xe moniker. The generational line is dotted and the product stack boundary is solid. If they wanted to stay with generational performance wouldn't the left most enclosing line have stayed dotted? More tongue and cheek I read the graph as pure fantasy. I could see December of 2020 if they were technically going to make it.

 

[Phynaz]

I'm reading their graph that the whole stack is represented within the Xe moniker. The generational line is dotted and the product stack boundary is solid. If they wanted to stay with generational performance wouldn't the left most enclosing line have stayed dotted? More tongue and cheek I read the graph as pure fantasy. I could see December of 2020 if they were technically going to make it.

Why do people think a chart on a marketing slide would be to any kind of actual scale? Marketing slides are for illustrative purposes.

 

[Despoiler]

Why do people think a chart on a marketing slide would be to any kind of actual scale? Marketing slides are for illustrative purposes.

I think everyone understands that scale isn't involved because we have no numbers on either axis. It's about the illustrative purposes. What are they trying to convey given there are at least two different interpretations.

 

[Mopetar]

Why do people think a chart on a marketing slide would be to any kind of actual scale? Marketing slides are for illustrative purposes.

While that's true, you'd probably want your marketing slide to suggest exponential growth as opposed to logarithmic.

Look at just about any AMD or NVidia marketing slide and you see the same exponential curve where the last 3-5 years barely rise above the axis and whatever is coming out in two or three years is almost off the top of the graph.

Sure, it's all marketing BS, but you might as well at least use good BS.

 

[NTMBK]

While that's true, you'd probably want your marketing slide to suggest exponential growth as opposed to logarithmic.

Look at just about any AMD or NVidia marketing slide and you see the same exponential curve where the last 3-5 years barely rise above the axis and whatever is coming out in two or three years is almost off the top of the graph.

Sure, it's all marketing BS, but you might as well at least use good BS.

Oh come on, what muppet uses a linear chart to depict something with exponential growth. You use a logarithmic scale. For example:

 

[Mopetar]

Marketing muppets. Here's the standard chart that I'm talking about that NVidia or AMD trot out with an exponential growth curve:

Go look at the chart you posted at the top of the page. It looks linear up to G9, but then because of the way they draw the lines, it looks like it's starting to taper off and if you extrapolated further, it would start to look logarithmic. There's no scale given other than "performance" so without additional information it's a bad chart. Even at best, it implies that performance over time isn't quite keeping the same pace as previously.

Also, the average person doesn't even understand logarithms anyways, so you'd never want to use that scale in marketing slides. The whole point is to make your product look impressive and an exponential curve on a linear scale gives the impression that the new stuff is massively better. Put the Intel chart next to that and the average person wonders why Intel sucks in comparison.

 

[jpiniero]

FWIW, the Intel Supercomputer announcement more or less confirms that at least the dGPUs that will be used will be fabbed on a non-Intel node, given that the timeframe would be too early for 7 nm.

 

[Arachnotronic]

FWIW, the Intel Supercomputer announcement more or less confirms that at least the dGPUs that will be used will be fabbed on a non-Intel node, given that the timeframe would be too early for 7 nm.

Why not 10nm?

 

[jpiniero]

Why not 10nm?

Because it only barely works.

 

[JasonLD]

Because it only barely works.

They wouldn't have gotten the contract if that is the case. It also uses future Xeon which is most likely 10nm.

 

[Arachnotronic]

Because it only barely works.

So how are they going to supply the CPUs then?

 

[maddie]

FWIW, the Intel Supercomputer announcement more or less confirms that at least the dGPUs that will be used will be fabbed on a non-Intel node, given that the timeframe would be too early for 7 nm.

These type of contracts are often a stealth way of subsidizing research. Many major US tech companies get these at times from DOE, DOD, etc. Also, does that timeframe assume completion of said contract or when it first becomes available for use.

 

[Dayman1225]

Because it only barely works.

I think it’ll be A-Ok by 2021, when they’re delivering the hardware..

 

[Ajay]

They wouldn't have gotten the contract if that is the case. It also uses future Xeon which is most likely 10nm.

Wait, if Intel can implement a die sufficient for a large many core Xeon CPU, then they can make a large GPU die. I, obviously, do not know what Intel is doing, just that things aren’t adding up.

 

[Dayman1225]

Think Exponential : Xe Architecture - Blog post by David Blythe

“Xe spans multiple computing and graphics market segments and will include a range of implementations that will allow us to address a wide range of markets and workloads, from mainstream notebooks to enthusiast game systems, to powerful computing solutions for the data center”

“We are already deeply engaged with customers and industry partners to develop data center solutions based on Intel Xe compute architecture. We will look forward to sharing more details on the Intel Xe architecture in the months ahead.”

Are probably the most interesting tidbits

 

[jpiniero]

So how are they going to supply the CPUs then?

It'd have to be 14 nm I guess. Or that's being fabbed on a non-Intel node too.

I think it’ll be A-Ok by 2021, when they’re delivering the hardware..

They said it would be up and running by 2021. Setting these things up takes time, so Intel would have to start delivering the hardware in 2020.

 

[Dayman1225]

It'd have to be 14 nm I guess. Or that's being fabbed on a non-Intel node too.



They said it would be up and running by 2021. Setting these things up takes time, so Intel would have to start delivering the hardware in 2020.

It’s being “stood up” in early 2021 and will be operational by the end of 2021. So they have till late 2020 early 2021.

Also LOL! Those CPUs will be 10nm just like the GPUs.

 

[DrMrLordX]

I wonder how Xe, Loihi, and Intel's various FPGA projects will share space in the project lineup?

 

[IntelUser2000]

It'd have to be 14 nm I guess. Or that's being fabbed on a non-Intel node too.

Or its just your bias talking.

There was a guy from twitter that went to the OCP summit and said not only 10nm problems were fixed, it was going better than expected. I guess going better than expected means it improved very rapidly in the last few months.

Historically they announced very low volume chips(high end usually) before ramping volume. Few Cannonlake chips may have played that part. The few chips that ship first are important because they provide real world experience which are better than endless experimentation.

 

[JasonLD]

It'd have to be 14 nm I guess. Or that's being fabbed on a non-Intel node too.



They said it would be up and running by 2021. Setting these things up takes time, so Intel would have to start delivering the hardware in 2020.

Aurora's target performance must be at least 5 times more powerful than Summit, which uses 14nm Power 9 and 12nm V100, it won't be achievable in 14nm unless they make it ridiculously bigger. Being skeptical is one thing, i think it is pretty naive to assume Intel will never find a way to fix their 10nm.