Intel Chips With “Vega Inside” Coming Soon?

[IntelUser2000]

Stop all the anti-intel FUD. It's all you do in these forums and it's always completely wrong. Intel has a AI chip from a company they bought. This one will release soon and estimates are it will be in the range of Volta V100. We will see soon.

I'm not sure if that matters a lot. Lake Crest is said to deliver 50 TOPS, and the next generation coming late 2018 is said to deliver almost double that: https://syncedreview.com/2017/04/15/what-does-it-take-for-intel-to-seize-the-ai-market/

That said, early Volta performance shows that 100TOPS tensor doesn't do much as expected, because in the real world, its a fraction of processing. At the best case it did 2x over Pascal, and average turned out to be 30% in some scenarios, with good ones reaching 50%. That is only slightly better than the number of shaders between Volta and Pascal.

So while Volta still is the best DL solution, its not 10x better and a complete revolution as Tensor unit throughput suggests. We should have known, real performance falls far short of looking at one number on a spec sheet.

 

[Dayman1225]

I'm not sure if that matters a lot. Lake Crest is said to deliver 50 TOPS, and the next generation coming late 2018 is said to deliver almost double that: https://syncedreview.com/2017/04/15/what-does-it-take-for-intel-to-seize-the-ai-market/

That said, early Volta performance shows that 100TOPS tensor doesn't do much as expected, because in the real world, its a fraction of processing. At the best case it did 2x over Pascal, and average turned out to be 30% in some scenarios, with good ones reaching 50%. That is only slightly better than the number of shaders between Volta and Pascal.

So while Volta still is the best DL solution, its not 10x better and a complete revolution as Tensor unit throughput suggests. We should have known, real performance falls far short of looking at one number on a spec sheet.

Any idea where they pulled the 80-90 TOPs number for Springs Crest?

 

[IntelUser2000]

Any idea where they pulled the 80-90 TOPs number for Springs Crest?

It sounds like the person who wrote the article heard it from a legitimate source. Look for the Innovation Symposium which Intel was a part of. Apparently representatives from DCG were there and talked about DL and other topics. Companies like Intel often host events in which they reveal information about current and future products.

Again, if TOPS is same TOPS as the one in Volta, then numbers are really not significant. In Volta the tensor units are probably responsible for 10-20% gains in performance. That is significant in itself, but not 10x as people expect. To get 10x gain with a unit that's 10x faster, that means the code consists of nothing but the type of instructions run by the unit.

 

[Dayman1225]

It sounds like the person who wrote the article heard it from a legitimate source. Look for the Innovation Symposium which Intel was a part of. Apparently representatives from DCG were there and talked about DL and other topics. Companies like Intel often host events in which they reveal information about current and future products.

Again, if TOPS is same TOPS as the one in Volta, then numbers are really not significant. In Volta the tensor units are probably responsible for 10-20% gains in performance. That is significant in itself, but not 10x as people expect. To get 10x gain with a unit that's 10x faster, that means the code consists of nothing but the type of instructions run by the unit.

50 Teraops seems about right for NNP, Naveen quoted 55 Teraops back in august'16

https://www.nextplatform.com/2016/08/08/deep-learning-chip-upstart-set-take-gpus-task/

Nvidia also has a half-precision mode built in as well, which can yield approximately 20 teraops per second, at least based on the numbers he says are gleaned from users that are actively putting Pascal through the paces for deep learning workloads. “On our chip, which is a TSMC 28 nanometer chip, we’re seeing around 55 teraops per second.”

 

[FIVR]

I'm not sure if that matters a lot. Lake Crest is said to deliver 50 TOPS, and the next generation coming late 2018 is said to deliver almost double that: https://syncedreview.com/2017/04/15/what-does-it-take-for-intel-to-seize-the-ai-market/

That said, early Volta performance shows that 100TOPS tensor doesn't do much as expected, because in the real world, its a fraction of processing. At the best case it did 2x over Pascal, and average turned out to be 30% in some scenarios, with good ones reaching 50%. That is only slightly better than the number of shaders between Volta and Pascal.

So while Volta still is the best DL solution, its not 10x better and a complete revolution as Tensor unit throughput suggests. We should have known, real performance falls far short of looking at one number on a spec sheet.

From what I've heard V100 loses to Vega FE in some workloads. There is a lot of hype around this "Tensor" core stuff that nvidia is peddling right now and much of it is not backed up by real-world performance numbers.


And you can only imagine how much of this Intel stuff is hype if nvidia is doing the same thing. I'm pretty sure the Intel hardware won't even come close to being performant enough. At this point it's vaporware.

 

[beginner99]

From what i read you are correct INTC's Nervana is an entry into DL,AI ect - but there is no software or community there which will be their challenge -

Agree. Software is the biggest issue. Strangely enough the open-source / scientific crowd jumped on CUDA like no tomorrow. So you have this great open-source projects but can only really use it with specific hardware and proprietary software (cuda). The irony...

Well all they need to to is write a cross-compiler like AMD is doing or compile tensorflow and co. for their own hardware. I guess AMD cross-compiler approach seems the smarter idea, if it works.

 

[IntelUser2000]

Agree. Software is the biggest issue. Strangely enough the open-source / scientific crowd jumped on CUDA like no tomorrow. So you have this great open-source projects but can only really use it with specific hardware and proprietary software (cuda). The irony...

That's why I believe Intel's best bet would likely end up being built around their core business(or should I say Core business?).

For example:
Knights Crest, a bootable Intel Xeon chip with Nervana IP.
Xeon SP version coming later with integrated FPGAs.

Those are the real deals.

There is a lot of hype around this "Tensor" core stuff that nvidia is peddling right now and much of it is not backed up by real-world performance numbers.

I think Tensor core does help, its just another specialized accelerator, rather than a general purpose boost for DL. Not understanding what it really did caused many of us to look at the 100TOPS number and translated that into 100TFLOPS.

 

[stockolicious]

Agree. Software is the biggest issue. Strangely enough the open-source / scientific crowd jumped on CUDA like no tomorrow

yup - NVDA was the first to invest heavily in this market - good move by them. I just dont know about their "go it alone" strategy though. AMD and INTC are already working together on special projects and I would not be surprised if they collaborated on something that would drive open source agenda in AI.

 

[Karnak]

https://www.intel.in/content/www/in...ntel-processors.html?iid=subhdr-IN+game_power

 

[Dayman1225]

https://www.intel.in/content/www/in...ntel-processors.html?iid=subhdr-IN+game_power

Guess that settles it, they're using Vega and not Polaris, interesting to say the least.

 

[Glo.]

https://www.intel.in/content/www/in...ntel-processors.html?iid=subhdr-IN+game_power

So, the i7 8809G is basically an 7920HQ with Radeon Vega 24 GPU with 4 GB of HBM2.

100W TDP target means that CPU has 45W TDP, and the GPU - the rest. If it really can be fit in this thermal envelope the Radeon graphics part of AMD may not be as doomed as people say they are...

 

[Glo.]

Guess that settles it, they're using Vega and not Polaris, interesting to say the least.

The work done for Intel in MI space with those GPUs will also be done for AMD, so it has more sense.

 

[itsmydamnation]

Guess that settles it, they're using Vega and not Polaris, interesting to say the least.

It also lines up perfectly with two other interesting data points:
1. Intel saying that this chip wasn't designed specifically for them ( paraphrasing)
2. The AMD chip that pop'd up a while ago with a very large number of compute units.

 

[Glo.]

It also lines up perfectly with two other interesting data points:
1. Intel saying that this chip wasn't designed specifically for them ( paraphrasing)
2. The AMD chip that pop'd up a while ago with a very large number of compute units.

You mean the one with Intel CPU, or Ryzen CPU?

Intel CPU version had 4 GB's of HBM2, and 1536 GCN cores, and Ryzen GPU had 1792 GCN cores, and 2 GB's of HBM2.

 

[Dayman1225]

Just noticed, this is under the overclocking section on Intel's website, so it seems you can OC it? Interesting.

 

[Glo.]

One more thing guys. When do you think Apple will announce new Mac mini with those SoCs, and in similar trash can shape as last gen Mac pro, just smaller?

 

[itsmydamnation]

You mean the one with Intel CPU, or Ryzen CPU?

Intel CPU version had 4 GB's of HBM2, and 1536 GCN cores, and Ryzen GPU had 1792 GCN cores, and 2 GB's of HBM2.

I mean I think they are both the same style setup with the same Dgpu, just different configurations.

 

[IntelUser2000]

100W TDP target means that CPU has 45W TDP, and the GPU - the rest. If it really can be fit in this thermal envelope the Radeon graphics part of AMD may not be as doomed as people say they are...

I think the CPU part may be running 35W nominally. 55W is too little for the performance they seem to be targeting.

Even Polaris parts can go into low power envelopes. Use of HBM2 saves significant amount of power on the memory side. If they bin it for low enough voltage the whole chip may do it readily in the 65W left.

 

[witeken]

Guess that settles it, they're using Vega and not Polaris, interesting to say the least.

This was already known for quite a while, since November at least!

 

[Dayman1225]

This was already known for quite a while, since November at least!

Always a rumor until confirmed and the fact that Geekbench said GFX804, pointing us towards Polaris.

 

[IntelUser2000]

Always a rumor until confirmed and the fact that Geekbench said GFX804, pointing us towards Polaris.

I wonder how long they have been planning for this then. Perhaps the Polaris versions were early testing chips.

 

[tential]

If only they had put this into the Xbox One X!

 

[NTMBK]

If only they had put this into the Xbox One X!

...then it would have a much less powerful GPU? Rumours suggest that this is a 24CU part, compared to the 40CU X1X.

 

[ksec]

One more thing guys. When do you think Apple will announce new Mac mini with those SoCs, and in similar trash can shape as last gen Mac pro, just smaller?

At first this all seems to be a special chip for Apple's Macbook range.

Now it is a Desktop Part, and listed in the OC section, I am not so sure if it is still Apple.

100W TDP doesn't fit into Mac Mini, and doesn't seems to be any reason why they would want this chip in 21" iMac, there is enough space for CPU + GPU combination.

So I am not entirely sure this is for Apple.

As a matter of fact, I dont understand why i want a Desktop Chip that has this combination and not a normal CPU + Vega.

 

[Glo.]

At first this all seems to be a special chip for Apple's Macbook range.

Now it is a Desktop Part, and listed in the OC section, I am not so sure if it is still Apple.

100W TDP doesn't fit into Mac Mini, and doesn't seems to be any reason why they would want this chip in 21" iMac, there is enough space for CPU + GPU combination.

So I am not entirely sure this is for Apple.

As a matter of fact, I dont understand why i want a Desktop Chip that has this combination and not a normal CPU + Vega.

Apple's Industrial Design is all about efficiency. Using as low parts as possible, using as low material as possible, using as low amount of power as possible... etc.

If you have single package, you can use one idea that was actually pretty brilliant: Thermal Core, and even make it smaller than before.

Nobody have said that new Mac mini with the Kaby Lake-G will resemble anything that has Apple designed before in Mac Mini space, maybe it will resemble the Mac Pro 6.1(Trash can)

Imagine something in this shape, but smaller, with SO-DIMM RAM fittings. It will make perfect sense.

And funniest part: Pike, from PikerAlpha has said in December 2016(!), that next generation Mac Mini will be bigger than current one. And more expensive . He, in the same post on his blog, said that next gen. iPhone will have wireless charging(and was pretty darn correct on this matter).

So lets say, that next gen. Mac Mini will resemble Mac Pro: 999$, 100W TDP package, with 24 CU Vega and 4C/8T CPU with 2.9 GHz base clock, 8 GB of DDR4 2400 MHz, 256 GB of NVMe SSD. It will make sense after all. For Apple ... .