You're not thinking clearly. The workstation market itself is fairly small from a unit perspective and I would imagine that the vast majority of workstation workloads require some sort of reasonable GPU power.
If you look at clock for clock comparisons, they are very closely matched as well. It's only due to the very conservatively clocked i7 920's that there was a big gap.
Would the HSA initiative increase the value of an IGP even if weaker in flops as compared to a discrete GPU?
I imagine access to main memory without the PCIe interface should be of some benefit, at least for some workloads.
Would the HSA initiative increase the value of an IGP even if weaker in flops as compared to a discrete GPU?
I imagine access to main memory without the PCIe interface should be of some benefit, at least for some workloads.
I was not even thinking of gaming but HPC and productivity uses.there would have to be game engine and driver tweaks to take advantage of the lower latency mem operations
All quite correct, but do you see AMD as being a bit more focused lately?HSA is supposed to be much more than bypassing the PCIe link for communication between GPU and CPU, and yes, the iGPU would become more relevant if HSA took off. The thing is that since AMD has close to zero relevance for software companies it will never take off.
All quite correct, but do you see AMD as being a bit more focused lately?
I'm getting the the impression that they are more carefully picking their target markets.
With that said, some high end modelling and HPC applications should greatly benefit from the use of HSA systems as the software is partially/wholly custom and these clients are often willing to pay well for higher performance. You don't need to start with winning everyone.
No, I don't, AMD is anything but a focused company: They still have two wars to wage against much bigger and resourceful competitors, and the stakes on both markets only get higher, not smaller. In fact, on top of waging the two wars they have the semi-custom distraction because they need to somehow make profits.
AMD strategy as of lately has been a reversal and dismantling of the strategy Rory Read tried to develop for the company. Rory tried to take AMD outside of the competition against Intel for new markets, once this strategy backfired he was fired and now Lisa Su reverted back trying to reassert their place on the x86 market while still competing against Nvidia. Not really radical, not really intelligent, not really worth their investor's money. They seem to think that with improved execution they will be able to survive and even thrive in the medium term. I don't think they understand one of their most fundamental problem today, which is that they lack resources to do everything they are supposed to do on the markets they are supposed to compete and because of that they won't be able to execute well. By executing well I mean developing good products, manufacturing, marketing and supporting it on the supply chain.
I don't think they will get any traction with their HSA approach, they don't have the software muscle to support such a venture and even if it did the high end isn't moving towards general purpose hardware like AMD is advocating, but dedicated accelerators on the MPU and FPGAs (Altera alone was almost as big as Nvidia in terms of market cap).
What is it here?No, I don't, AMD is anything but a focused company: They still have two wars to wage against much bigger and resourceful competitors, and the stakes on both markets only get higher, not smaller. In fact, on top of waging the two wars they have the semi-custom distraction because they need to somehow make profits.
AMD strategy as of lately has been a reversal and dismantling of the strategy Rory Read tried to develop for the company. Rory tried to take AMD outside of the competition against Intel for new markets, once this strategy backfired he was fired and now Lisa Su reverted back trying to reassert their place on the x86 market while still competing against Nvidia. Not really radical, not really intelligent, not really worth their investor's money. They seem to think that with improved execution they will be able to survive and even thrive in the medium term. I don't think they understand one of their most fundamental problem today, which is that they lack resources to do everything they are supposed to do on the markets they are supposed to compete and because of that they won't be able to execute well. By executing well I mean developing good products, manufacturing, marketing and supporting it on the supply chain.
I don't think they will get any traction with their HSA approach, they don't have the software muscle to support such a venture and even if it did the high end isn't moving towards general purpose hardware like AMD is advocating, but dedicated accelerators on the MPU and FPGAs (Altera alone was almost as big as Nvidia in terms of market cap).
For workstation use you just need a decent iGPU, not Iris Pro class iGPU. The latter is waste.
a workstation APU would become the jack of all trades, master of none, it would become the broke student choice for building a workstation.
Would the HSA initiative increase the value of an IGP even if weaker in flops as compared to a discrete GPU?
I imagine access to main memory without the PCIe interface should be of some benefit, at least for some workloads.
All quite correct, but do you see AMD as being a bit more focused lately?
This problem is not specific to workstation CPUs though. If anything it's even worse on mainstream desktop CPUs.Actually some workstations do not require GPU power at all, while others do quite good with an accelerator like Xeon Phi or Tesla, and others need graphics grunt.
All in all PPB is quite correct in what he is saying, whenever you need GPU or acceleration power in a workstation it's "Go big or go home", a workstation APU would become the jack of all trades, master of none, it would become the broke student choice for building a workstation.
With that said, some high end modelling and HPC applications should greatly benefit from the use of HSA systems as the software is partially/wholly custom and these clients are often willing to pay well for higher performance.
A bit more focused = a focused company, is your understanding of what I wrote. Heaven help us, what's up with all this either or thinking.
edit:
By the way, unlike you, I consider the custom SOCs for consoles, etc, to be an example of focused development. No other company, right now, can hit those targets.
This problem is not specific to workstation CPUs though. If anything it's even worse on mainstream desktop CPUs.
This problem is not specific to workstation CPUs though. If anything it's even worse on mainstream desktop CPUs.
On a 4 core Skylake CPU the iGPU occupies ~50% of the die area, on a 2 core ~60-70%. All potential waste for those that really require high GPU performance (e.g. for gaming), since they'll get a discrete GPU anyway. And for those that don't require high GPU performance, the iGPU will be overkill.
On an 8 core workstation CPU with the same iGPU it would only occupy ~25% of the die area. Cut that in half to get a decent iGPU good enough for workstation use requiring mostly excellent MT CPU performance, and we're down to 10-15%. Not nearly as problematic as the 50-70% on mainstream desktop CPUs.
AMD in 2010 used to have:
Big core CPU, with Server and APU line up
Small core APU
GPUs
Toda's AMD:
Big core CPU, with server and APU line up
ARM core server and embedded line up
Server interconnect
Memory technology
GPUs
Software tools for HSA
Middleware for gaming industry
Semi-custom MPU business
The problem of potentially wasted die area by always including an iGPU is still much less than on mainstream desktop, as clarified in my previous post. So if it is ok to include an iGPU on mainstream desktops, it should certainly be ok on workstation CPUs too where the percentage of die area penalty is much much less.You still have 10-15% wasted die space in a iGP no one would use (in the workstation world) Considering the low die size per core hypotesis some people are running, you could get 2 more cores/4 threads in that space budget. For workstations running fully multithreaded apps that is a net 25% perf gain (if we are coming from a 8c/16t chip).
The only possible configuration workstations would see a good use is probably a totally anemic iGP that is only needed to give you video image in the workstation. That would be 1-2CU for AMD and still it would make the now-APU innecesarily complex just to have an iGP.
I would say that from a debugging, process and time to market oriented POV, if you iGP doesnt even take 25% o your die, you should reconsider having an iGP altogether.
The problem of potentially wasted die area by always including an iGPU is still much less than on mainstream desktop, as clarified in my previous post. So if it is ok to include an iGPU on mainstream desktops, it should be ok on workstation CPUs too where the percentage of die area penalty is much much less.
Both. If you can reduce platform cost by including a decent iGPU on die, why not do it?The main reason for integration is to shrink platform costs. Is the workstation market demanding less costs, like the consumer market, or is this market demanding more performance?
The problem of potentially wasted die area by always including an iGPU is still much less than on mainstream desktop, as clarified in my previous post. So if it is ok to include an iGPU on mainstream desktops, it should be ok on workstation CPUs too where the percentage of die area penalty is much much less.
For most workstation usage you're going to need an iGPU anyway. For typical workstation task like SW development, editing videos, and similar you're going to need to connect it to a display. And often those same computers are also used for web surfing, editing documents and other typical office tasks. It doesn't require much GPU performance, but still and only a decent one.
I can see why AMD would not want to include an iGPU in the first generations of the Zen CPU, since it would require additional design and testing as you say, delaying the products TTM. But later once they have a Zen based APU ready, they have all the bits and pieces ready to create a workstation variant with 8-16 cores and a small iGPU. So then it should not require much effort.You are still not thinking this from the perspective of AMD. Why add validation and debugging time into their already tight schedule if the iGP:
This is where we disagree. Because I do think the iGPU will be used. Typical tasks like SW development, video editing, and office-type side-activities like document editing and web surfing requires it. A decent iGPU that is, nothing more, nothing less. You can of course always add a discrete graphics card, but then the total platform cost will be higher except for the cases where a really powerful GPU is needed. Same reasoning as for mainstream desktop CPU/APUs.- Will ocuppy only 15% of the die
- Will seldom be used in the HEDT market (even an 1150/1 i7 K proccesor's iGP is seldom used).
It's not either or. You can have an iGPU on a 10C/20T CPU too.You are far off better adding more cores which have a lot higher of a chance of being used in HEDT/workstation scenarios, making ToM stay almost the same and have a higher chance to compete against Intel's HEDT products. AMD will sell more 10/20t dies if they are priced the same as they would sell 8c/16t plus 2CU if they are stacked (and will be) against Intel's HEDT.
As I mentioned above, I agree this makes sense in the first generation, before there are Zen APUs ready. But after that it's a different situation.To make it simpler, why do you think the first Zen APU (higher volume, less margin but potential for more revenue) is coming later than the first Zen CPU (lower volume, more margins but less potential of revenue)? It's very simple: The Zen APU has a lot more of complexity on it due to the iGP, thus more validation/debugging/binning that the CPU variant. Having both products be actual APUs is no use because it would only cover corner cases for AMD, while delaying their product and leaving even more sales to Intel.
The faster AMD gets Zen out of the door and the more competitive in pure CPU performance is against Intel's HEDT, the better. Adding super-anemic iGPs to HEDT products is a no-go.
Both. If you can reduce platform cost by including a decent iGPU on die, why not do it?