Xeon Phi might be dead

[patrick3027]

Quote

 

While vendors are busy announcing new HPC offerings at this week’s Supercomputing Conference (SC17), Intel announced it is removing its next-generation “Knights Hill” Xeon Phi product from its roadmap. And that might just be the beginning.

In a blog penned by Intel’s Data Center Group GM Trish Damkroger describing the company’s exascale strategy and other topics they are talking about at the SC17 conference, she offhandedly mentioned that the Knights Hill product is dead. More specifically she said that the chip will be replaced in favor of “a new platform and new microarchitecture specifically designed for exascale.”

 

Full article: https://www.top500.org/news/intel-dumps-knights-hill-future-of-xeon-phi-product-line-uncertain/

 

Might have to do something with AMD's increased competition?

[Bananasplit_00]

well they killed the Atom already, and these chips were basically just a shitload of Atoms with glue so in a way im not surprised

[porina]

This category of product is all about double precision floats. AMD aren't competitive in that area. Nvidia are. Atom is neither here nor there, it is basically just a carrier to enable the FPUs to be used.

[Master Disaster]

23 minutes ago, patrick3027 said:

Full article: https://www.top500.org/news/intel-dumps-knights-hill-future-of-xeon-phi-product-line-uncertain/

 

Might have to do something with AMD's increased competition?

AMD had nothing that competed with Phi, actually nothing even close to it.

 

You wanna know my opinion? Intel are planning to use Phi technology as a base for their new GPUs. I mean Phi was basically like CUDA/OpenCL in the way it handled workloads, makes sense for Intel to stick it in a GPU rather than sell it as a CPU Add On.

[Guest]

13 minutes ago, porina said:

This category of product is all about double precision floats. AMD aren't competitive in that area. Nvidia are. Atom is neither here nor there, it is basically just a carrier to enable the FPUs to be used.

Especially in the AI supercomputing segment, Nvidia provides Tesla V100 Volta GPUs which have 16GB of VRAM each and are implemented in big clusters in data centres.

[Castdeath97]

Well I kinda thought it was neat, shame but I guess they couldn't find a good use case perhaps.

[Guest]

I'd say that AMD is more competitive in the server/workstation grade hardware with the 32 core EPYCs and opterons falling into the server category and the threadripper CPUs falling into the workstation/enthusiast level grade hardware. Xeon processors are known for having lots of cores (hence why they are used in the supercomputing segment) and the workloads that they deal with consists of spreading the workloads amoung lots of cores.

[Bananasplit_00]

21 minutes ago, Castdeath97 said:

Well I kinda thought it was neat, shame but I guess they couldn't find a good use case perhaps.

They were only useful in tasks that were extremely parelalised and can't be GPU accelerated, which ends up being not a lot of things 

[leadeater]

1 hour ago, porina said:

This category of product is all about double precision floats. AMD aren't competitive in that area. Nvidia are. Atom is neither here nor there, it is basically just a carrier to enable the FPUs to be used.

They used to be, AMD had to give up FP64 performance to focus on other areas as they weren't selling any cards based on that decent FP64 performance so no real loss removing it.

 

As for FP64 for Nvidia their product stack is actually very limited as to which has good performance, the majority has 1/32 FP64 even on the Quadro P6000.

 

AMD FirePro S9000 (2012): 806 GFLOPS FP64

AMD FirePro S9150 (2014): 2530 GFLOPS FP64

AMD FirePro S9170 (2015): 2620 GFLOPS FP64

https://en.wikipedia.org/wiki/List_of_AMD_graphics_processing_units#Workstation_GPUs

 

Nvidia K80 (dual GPU, 2014): 1864–2912 GFLOPS FP64

Nvidia M60 (dual GPU, 2015): 230.1–301.6 GFLOPS

Nvidia Quadro P6000 (2016): 375 GFLOPS FP64

Nvidia P100 (2016): 4036‒4670 GFLOPS FP4

Nvidia V100 (2017): 7014 GFLOPS

https://en.wikipedia.org/wiki/Nvidia_Tesla

 

AMD actually had better FP64 performance than Nvidia did for a rather long time and Nvidia's FP64 performance has been all over the place, only came good since the P100 and they limit it to the Gx100 dies.

 

Edit:

Xeon Phi 7120X (2013): 1208 GFLOPS FP64

Xeon Phi 7290 (2016): 3456 GFLOPS FP64

 

For reference to compare against GPUs.

[DocSwag]

Random thought... you think Intel might be replacing Phi with that gpu they are making?

[BuckGup]

15 minutes ago, leadeater said:

snip

Didn't know the GV100 was so fast. wow

[DocSwag]

1 hour ago, Master Disaster said:

AMD had nothing that competed with Phi, actually nothing even close to it.

 

You wanna know my opinion? Intel are planning to use Phi technology as a base for their new GPUs. I mean Phi was basically like CUDA/OpenCL in the way it handled workloads, makes sense for Intel to stick it in a GPU rather than sell it as a CPU Add On.

I don't think they're gonna evolve Phi for GPUs. I think Phi is dead.

 

Phi originally rose from the ashes of Larrabee which was Intel's take in the late 2000s on a gpu made of lots of smaller less powerful cores (I believe originally they planned to make them P4 based). When that failed they ended up adapting Xeon Phi from that and it actually kind of worked out.

 

GPUs need parallelism. Yes, Intel could stuff a hundred or two hundred smaller cores on a pcie add in card and potentially make a graphics driver that allows it to act as a gpu, but it still probably won't be as good at it as an actual gpu that has more fixed function hardware more suitable for the task.

 

If you ask me HPC makes the most sense for lots of smaller CPU cores.

 

What I really think is going to happen is Intel is going to replace Phi with something built off the GPUs they are making.

[porina]

39 minutes ago, leadeater said:

They used to be, AMD had to give up FP64 performance to focus on other areas as they weren't selling any cards based on that decent FP64 performance so no real loss removing it.

 

As for FP64 for Nvidia their product stack is actually very limited as to which has good performance, the majority has 1/32 FP64 even on the Quadro P6000.

 

AMD actually had better FP64 performance than Nvidia did for a rather long time and Nvidia's FP64 performance has been all over the place, only came good since the P100 and they limit it to the Gx100 dies.

DP float performance is an interesting area for me, and in a similar way that hardware changes over time, so have software.

 

DP performance has always been a niche, important for those that need it, not for those that don't. AMD are betting on the forward looking money following higher rates of lower precision, which seem to suit machine learning. So in AMD's current product range (CPU or GPU), I'm not aware of anything really useful in that area. nvidia only offer it on the professional end, and it is pretty much gone from consumer level kit. Intel are the odd ball, even their consumer CPUs still offer strong performance in that area, and with cherry picking, an i3-8350k would have a good chance of exceeding Threadripper 1900X in this kinda use case.

 

I still have an R9 280X card, chosen as it had ~1 TFLOP of 64 bit performance. It was the last "fast" consumer DP card from red or green. From memory, nvidia's 500 series whatever that was, was the last one they did with strong DP before they started cutting it to the near non-existent levels in today's cards. However I found my GPU DP need to have gone, as software has adapted to the consumer market cards. It isn't ideal, but the throughput of a much higher SP rate is faster than using what little remains of the DP rate, so from an individual basis, consumer cards remain viable without paying for the much more expensive "professional" cards.

 

On the CPU side, it doesn't look like Intel are giving up, with AVX-512 offering a potential doubling of throughput over AVX 2, and rumours of AVX-512 will be making its way into future generation consumer level CPUs. Currently it is most easily obtained through Skylake-X. Still waiting for software to catch up with it though. If I were to describe my ideal DP powerhouse CPU based on existing technologies and concepts, it would essentially be Xeon Phi. You just want a great FPU with enough stuff around it to feed it. It is still closer to CPU than GPU in terms of flexibility. My dream would be to have compute devices natively handling data sizes far greater than 64-bit, say megabits, but this would be far too niche for anything commercially viable, so I doubt I'll see it in my lifetime. I would settle for a conceptual "x64 GPU", but that would be a very different beast from GPUs optimised for graphical uses.

[Dabombinable]

Perhaps the phi will transition back to what it was with Larabee.....

[Coaxialgamer]

The cancelation of xeon phi has nothing to do with amd. AMD doesn't even have a competing product. Zen core is too big and too power hungry for that. Jaguar is the closest they have, and it's old and inadequate. 

Rather, xeon phi was competing against Nvidia, Cuda and gpus for the AI and GPGPU market. It was doing poorly in most cases, and intel’s gpu push might be the reason it finally bit the dust. 

[mr moose]

This doesn't surprise me,  The whole (and I mean whole) of the tech industry is in a very dynamic place right now,  not only are we pushing the boundaries of silicon and lithography, but future demand and processing requirements can change.  If someone had of said GPUs would be powering data centres 7 years ago you'd be laughed at in many circles (who needs cuda to retrieve a db cell).  If it takes 2-3 years to develop a GPU then you could be almost finished your next one when you discover that the industry is calling for something your card doesn't do well.    Given Intel have been buying companies with IP in the AI sector for a few years now I dare say they have a new plan.

 

 

[AnonymousGuy]

They killed Knight's Hill because they found that it wouldn't have really been feature/performance/whatever competitive by the time it would be released.  It's not a sign that Phi is completely dead.  It could go that way but more likely they'll just rejigger together a better product.

[Trixanity]

Atom is dead so they need a new foundation for Xeon Phi. Seems to be a smart move to make something new that is low power without sacrificing performance like Atom did. They might have had something in the works for a while without us knowing. Atom's demise have been going on for years.

[SpaceGhostC2C]

On 11/15/2017 at 1:20 PM, patrick3027 said:

Might have to do something with AMD's increased competition?

I think it may rather have something to do with a former AMD employee and the unit he will join to

 

It would be sad, though: GPGPU is a pain compared to having x86-based, making Phi very attractive for scientific computations - except Phi was always behind tons of locks that tied it to excessively expensive platforms, so not exactly the thing your average academic can hope to get his/her hands on  

I guess we can always hope for scrapyard hunting a few years down the road...