Something New from AMD on the 23rd?

[monstercameron]

 

 

 

APU's were developed with the idea of coming closer to confining the desktop. They've been on the market for years without anyone ever thinking of utilizing the iGPU for compute. AMD has been selling them to the consumer market ever since Llano based on how well they game as they did not feature a heterogeneous architecture back then. Sure they do offer a great amount of compute power now that HSA exists and can finally be utilized to its fullest. Tho a new addition wouldn't change anything for the consumer desktop market. It would have to be packing something extra special in order for it to top Kaveri. Even with Excavator cores and a GPU context switch, it still wouldn't warrant enough of a reason to launch a new flagship APU. Let's also understand that consumer chips differ from commercial chips (you won't find a A10-7850k in a server). This PR clearly came straight out of AMD's GPU department. So I am willing to bet its Volcanic Islands 2.0 (aka Tonga).

http://wccftech.com/amd-berlin-server-apu-glimpse-upcoming-kaveri-apu-4-steamroller-cores-512-gcn-sps/

 

whether or not they get design wins is another matter.

[patrickjp93]

APU's were developed with the idea of coming closer to confining the desktop. They've been on the market for years without anyone ever thinking of utilizing the iGPU for compute. AMD has been selling them to the consumer market ever since Llano based on how well they game as they did not feature a heterogeneous architecture back then. Sure they do offer a great amount of compute power now that HSA exists and can finally be utilized to its fullest. Tho a new addition wouldn't change anything for the consumer desktop market. It would have to be packing something extra special in order for it to top Kaveri. Even with Excavator cores and a GPU context switch, it still wouldn't warrant enough of a reason to launch a new flagship APU. Let's also understand that consumer chips differ from commercial chips (you won't find a A10-7850k in a server). This PR clearly came straight out of AMD's GPU department. So I am willing to bet its Volcanic Islands 2.0 (aka Tonga).

And it shocks me AMD has not released a workstation platform for professional users to utilize APUs with HSA software.

[monstercameron]

And it shocks me AMD has not released a workstation platform for professional users to utilize APUs with HSA software.

its getting there!

[patrickjp93]

its getting there!

It better have at least 24 PCIe 3.0 lanes lol.

[DigitalHermit]

With this thread still going, I'm getting pretty hyped on the release... (Aug 21)

[Opcode]

http://wccftech.com/amd-berlin-server-apu-glimpse-upcoming-kaveri-apu-4-steamroller-cores-512-gcn-sps/

 

whether or not they get design wins is another matter.

Berlin is pretty much Kaveri, tho I was referring to not seeing a consumer branded chip in a server.

 

And it shocks me AMD has not released a workstation platform for professional users to utilize APUs with HSA software.

I don't think its necessary as you can just go buy a Kaveri chip. Plus if you do that you will know exactly how to optimize the software for what hardware is available. Unless you're referring to on a larger scale.

[TriBeard]

As much as I'd like it to be something new for the FX line, or a new direction and something competitive for the high end, it's probably just Tonga. Which, is, by all accounts a great bang for the buck GPU, but not as exciting to me.

[Vacsol]

The picture only shows their flagship cards, so it has to be something big.

although if it was a huge release, they should have had tons of teasers for the past month

 

Maybe it will come with your first month's supply of LN₂ to keep it cool

[Humbug]

It seems really unlikely given how well Maxwell has proven itself. I know Nvidia is overpriced (though they say the 800s cards will be cheaper than the 700s), but that power efficiency and overclockability look sweet.

Honestly I doubt that Maxwell is going to give AMD any nightmares. Nvida is already producing bigger dies with more transistors (more expensive) in order to achieve the same performance that AMD does. And the only Maxwell part we have seen so far the 750ti did not show any indication that they can catch up with AMD in the perfomance/die size department. Most of the improvements were with power consumption not performance. I hope the rest of the Maxwell range is more impressive in terms of performance improvements.

 

I think it will be similar to the previous generation with AMD leading on performance/die size and hence price/performance.

But Nvida winning on power consumption/performance.

[MarvinKMooney]

I wish people would cut the crap about heat jokes <_<

. If only they weren't jokes! Don't worrie about it I still like amd.

[patrickjp93]

Berlin is yes pretty much Kaveri, tho I was referring to not seeing a consumer branded chip in a server.

 

I don't think its necessary as you can just go buy a Kaveri chip. Plus if you do that you will know exactly how to optimize the software for what hardware is available. Unless you're referring to on a larger scale.

No ECC*

[patrickjp93]

Honestly I doubt that Maxwell is going to give AMD any nightmares. Nvida is already producing bigger dies with more transistors (more expensive) in order to achieve the same performance that AMD does. And the only Maxwell part we have seen so far the 750ti did not show any indication that they can catch up with AMD in the perfomance/die size department. Most of the improvements were with power consumption not performance. I hope the rest of the Maxwell range is more impressive in terms of performance improvements.

 

I think it will be similar to the previous generation with AMD leading on performance/die size and hence price/performance.

But Nvida winning on power consumption/performance.

You forgot overclockability. The scaling on the 750TI was way better than anyone expected. Also, the new flagship chip is smaller than GK110, while fitting more cores.

[vm'N]

And it shocks me AMD has not released a workstation platform for professional users to utilize APUs with HSA software.

HSA is not really going into workstation. They use a dedicated GPU.

HSA is not for heavy workloads as you would see in the workstation environment.

Serial SIMD workloads favour the integrated SIMD cluster, meanwhile heavy parallel workload will favour the GPU.

Some parallel SIMD workload will favour HSA because of a lower latency.

This is also why HSA WONT replace the integrated SIMD cluster. Some workloads are more serial, and therefore are more optimised for a smaller SIMD cluster.

[patrickjp93]

HSA is not really going into workstation. They use a dedicated GPU.

HSA is not for heavy workloads as you would see in the workstation environment.

Serial SIMD workloads favour the integrated SIMD cluster, meanwhile heavy parallel workload will favour the GPU.

Some parallel SIMD workload will favour HSA because of a lower latency.

This is also why HSA WONT replace the integrated SIMD cluster. Some workloads are more serial, and therefore are more optimised for a smaller SIMD cluster.

facepalm* There are workstation tasks which need a balance of large parallelism and low latency which HSA would be great for. Also, don't forget AMD allows its iGPU to Crossfire with dedicated cards. Obviously right now that scaling is limited, and memory bandwidth is such a problem the 7850k can actually wound performance when in XFire with a 250X.

 

That said, the system is going to get better, and there is a discernible need for APU workstation chips. Intel seems to agree given some of its Xeons have iGPUs and some don't, and Skylake is a Kaveri copycat.

[vm'N]

facepalm* There are workstation tasks which need a balance of large parallelism and low latency which HSA would be great for.

Most parallel workload are NOT very depended on latency. By far most parallel workload are non-affected by the latency.

HSA would only hit a VERY small market of workstations (will change when a better integrated solution is available (+5 years)). EDIT: When the integrated GP will be strong enough. By this time lower/middle-end GPUs will be gone due to APUs, and in the end you cannot buy a gaming GPU (due to integrated graphics been good enough). However this is not within the next 7 years.

 

That said, the system is going to get better, and there is a discernible need for APU workstation chips. Intel seems to agree given some of its Xeons have iGPUs and some don't

No, there really are not any need for APU workstation chips. Do you even realise why Intel ship a IGP on SOME of their xeons? It is in most cases irrelevant to its workloads and more relevant to having a output for the display (smaller workstations).

[patrickjp93]

Most parallel workload are NOT very depended on latency. By far most parallel workload are non-affected by the latency.

HSA would only hit a VERY small market of workstations (will change when a better integrated solution is available (+5 years)). EDIT: When the integrated GP will be strong enough. By this time lower-end GPUs will be gone due to APUs, and in the end you cannot buy a gaming GPU (due to integrated graphics been good enough). However this is not within the next 7 years.

 

No, there really are not any need for APU workstation chips. Do you even realise why Intel ship a IGP on SOME of their xeons? It is in most cases irrelevant to its workloads and more relevant to having a output for the display (smaller workstations).

Vs. simply having the display go through a more useful quadro/firepro used in the big calculations? I don't think so. There are a number of tasks small enough where the removed latency is great. Real-time simulation I happen to think is one of them.

[vm'N]

Vs. simply having the display go through a more useful quadro/firepro used in the big calculations? I don't think so. There are a number of tasks small enough where the removed latency is great. Real-time simulation I happen to think is one of them.

Xeons with integrated GP is normally the lower-end xeons. Where you normally won't have those quadro/firepros.

But those numbers are still considered extremely small overall. Parellel is by nature not as depended on latency as serial.

[patrickjp93]

Xeons with integrated GP is normally the lower-end xeons. Where you normally won't have those quadro/firepros.

But those numbers are still considered extremely small overall. Parellel is by nature not as depended on latency as serial.

It is still benefits by having low latency. As per low-end, yes, that's the B9x/H9x platform on socket 1150, but what does that matter? Intel knows socket 2011 is dead in a couple years. That's why the top Xeons had FPGAs attached to them: more performance capabilities at lower power additions to stave off competition from Qualcomm, AMD, and other ARM chip makers.

[vm'N]

It is still benefits by having low latency. As per low-end, yes, that's the B9x/H9x platform on socket 1150, but what does that matter? Intel knows socket 2011 is dead in a couple years. That's why the top Xeons had FPGAs attached to them: more performance capabilities at lower power additions to stave off competition from Qualcomm, AMD, and other ARM chip makers.

Most SIMD workloads actually don't benefit from lower latency. Because what we currently do it copy everything over to the GPU and let it do the calculations, and it will send all the answers back. Workloads that are more latency depended, is generally due to having a more serial approach, where the integrated SIMD cluster would be more than fine (as it is optimised for serial SIMD workloads).

Intel implementing FPGA is more related to microsoft.

[patrickjp93]

Most SIMD workloads actually don't benefit from lower latency. Because what we currently do it copy everything over to the GPU and let it do the calculations, and it will send all the answers back. Workloads that are more latency depended, is generally due to having a more serial approach, where the integrated SIMD cluster would be more than fine (as it is optimised for serial SIMD workloads).

Intel implementing FPGA is more related to microsoft.

The FPGA thing has to do with scientific computing. Nothing like circuitry dedicated to just doing FFTs, or dedicated to compressing data, or x86-ARM translation on the way from the Xeon to a neighboring low-power ARM chip.

 

Copying everything over to the GPU goes out with window with Skylake as it did for Kaveri.

[vm'N]

The FPGA thing has to do with scientific computing. Nothing like circuitry dedicated to just doing FFTs, or dedicated to compressing data, or x86-ARM translation on the way from the Xeon to a neighboring low-power ARM chip.

 

Copying everything over to the GPU goes out with window with Skylake as it did for Kaveri.

I simply said that microsoft is the reason why Intel implemented the FPGA.

When microsoft presented their 'catapult', Intel had to do something. This was their reaction.

There are actual benefits of doing copying.

Skylake will still copy to the GPU, unless you want the GPU to access main memory constantly. So no. It stays.

[patrickjp93]

I simply said that microsoft is the reason why Intel implemented the FPGA.

When microsoft presented their 'catapult', Intel had to do something. This was their reaction.

There are actual benefits of doing copying.

Skylake will still copy to the GPU, unless you want the GPU to access main memory constantly. So no. It stays.

It is accessing main memory. You don't keep up with news much do you? It's been confirmed it's using the unified memory model.

[vm'N]

It is accessing main memory. You don't keep up with news much do you? It's been confirmed it's using the unified memory model.

You have no idea of how GPGPU works? Do I really need to explain the whole copy thing, and why it will continue?

[patrickjp93]

You have no idea of how GPGPU works? Do I really need to explain the whole copy thing, and why it will continue?

I know exactly how it works and I know AMD and Intel both are getting rid of requiring copying to reduce workload time for their APUs. Do YOU know how unified memory works?

[monstercameron]

You make it seem like unified memory is a new concept.