Broadwell GT3 48EUs? TDP range 4.5W-47W

[Homeles]

Nope, the reason iGPUs are scaling faster than dGPU today is because the performance was at the very bottom. There will be a time soon, when iGPU performance scalling will start to slow down and will catch that of dGPUs.

I don't disagree that the rate that the gap between APUs and dGPUs is closing will slow down, but the gap is still closing regardless.

Even next year, Kaveri will not have the same performance as 7750 (GDDR5).

You're still missing the point. Does Kaveri not close a larger portion of that gap? The point is that the gap is closing, and with every passing generation, dGPUs become increasingly less relevant.

Broadwell GT2 might actually pass Kaveri's IGP performance, and could match a 7750 in performance.

And dont forget that HD7750 will be already in the market for 2+ years.

That is only relevant because GloFo is so behind on process tech. AMD's APUs are currently a node and a half behind any upcoming 20nm dGPUs.

Intel's immune to that issue -- Skylake will be competing with whatever 20nm successors AMD and Nvidia have. The problem is exacerbated even more when you realize that GPUs refresh less often than CPUs/APUs.

 

[Lonbjerg]

I don't disagree that the rate that the gap between APUs and dGPUs is closing will slow down, but the gap is still closing regardless.

You're still missing the point. Does Kaveri not close a larger portion of that gap? The point is that the gap is closing, and with every passing generation, low end GPU

Broadwell GT2 might actually pass Kaveri's IGP performance, and could

That is only relevant because GloFo is so behind on process tech. AMD's APUs are currently a node and a half behind any upcoming 20nm dGPUs.

Intel's immune to that issue -- Skylake will be competing with whatever 20nm successors AMD and Nvidia have. The problem is exacerbated even more when you realize that GPUs refresh less often than CPUs/APUs.

GPU's offer a much lager performance jump between generations than with CPU's.

Hence why my i7 920 went through 3 GPU's.
My 990x will also get 3 GPU upgrades.

When eg. Maxwell hits...I could upgrade my GPU...and get WAY more performance.
There is no CPU on the market that would give me a similar CPU performance upgrade.

Those are facts...like them or not.

 

[Homeles]

GPU's offer a much lager performance jump between generations than with CPU's.

Hence why my i7 920 went through 3 GPU's.
My 990x will also get 3 GPU upgrades.

When eg. Maxwell hits...I could upgrade my GPU...and get WAY more performance.
There is no CPU on the market that would give me a similar CPU performance upgrade.

Those are facts...like them or not.

Funnily enough, APUs have a GPU in them. That's a fact; like it or not.

 

[Lonbjerg]

Funnily enough, APUs have a GPU in them. That's a fact; like it or not.

You mean those little IGP's that offer console level "performance"?
No thanks...might as well buy a console if image quality/framerates have no interest.

Why do you need a PC?
Shouldn't you be looking at a console?

 

[AtenRa]

I don't disagree that the rate that the gap between APUs and dGPUs is closing will slow down, but the gap is still closing regardless.

No it is not, Kaveri which will be the highest performance mainstream APU in 2014 both from AMD and Intel, will only have HD7750(DDR3) dGPU performance. Even if Kaveri had 2x the Stream Processors it would gain no tangible extra performance due to lack of Memory Bandwidth. Same applies to Intel iGPUs, HD5000(GT3) is not that Faster than HD4600(GT2) although it has double the EUs.

You're still missing the point. Does Kaveri not close a larger portion of that gap? The point is that the gap is closing, and with every passing generation, dGPUs become increasingly less relevant.

Kaveri will only manage to get close to HD7750(DDR3) performance. dGPUs have that performance for 2+ years now at that price/TDP point.

Broadwell GT2 might actually pass Kaveri's IGP performance, and could match a 7750 in performance.

Without eDRAM it will not have a chance. Broadwell GT2 could only catch HD5000 at the most which is way lower than HD7750.

That is only relevant because GloFo is so behind on process tech. AMD's APUs are currently a node and a half behind any upcoming 20nm dGPUs.

20nm dGPUs are one year away. But when they will arive, mainstream performance will double. That means we will get ~7850 performance at $150 or less price point. Broadwell in early 2015 will be pressed to compete against mainstream 20nm dGPUs and it will loose big time.

Intel's immune to that issue -- Skylake will be competing with whatever 20nm successors AMD and Nvidia have. The problem is exacerbated even more when you realize that GPUs refresh less often than CPUs/APUs.

Intel has the same problem AMD and NVIDIA has, Memory Bandwidth and cost of die size. Even with 14nm Intel will not be able to have a mainstream iGPU at HD7750 GDDR5 level of performance in 2014.

But, AMD and Intel dont throw all that money to APUs for Gaming performance. Compute is the name of the game and that will be seen with Kaveris hUMA and HSA.

 

[Zodiark1593]

You're not getting it, which is clearly demonstrated by that straw man. The only wall that IGPs have right now that dGPUs do not is memory bandwidth. Once that wall is removed with the bandwidth provided by stacked memory, IGPs will make low end GPUs irrelevant.

After stacked memory drops in, the only things holding back GPUs are manufacturing process and architecture. Since Intel has a lead here on the former, you'll likely see them catch up to all but the highest end dGPUs as they improve on the latter.

This is false.

Do take a look at the die budget given to dGPUs and iGPUs. Even if iGPUs somehow gain superior bandwidth, the die space allotted for mid-high end dGPUs allow for vastly superior compute and shader performance. Bandwidth alone will not bring Intel to parity due to the gap in sheer compute power as Intel only has a fraction of a die to spend on GPU, while Nvidia and AMD have dies several times larger than Intel's entire cpu die, the entirety of which is spent on the GPU.

For Intel to pull ahead of high end dGPUs would require a vast process advantage (no, a 14nm vs 28nm is not vast) so as to fit the required compute units needed to overtake AMD and Nvidia in compute performance.

I suppose for your next argument, you will attempt to downplay the importance of compute power. Amirite?

 

[ShintaiDK]

This is false.

Do take a look at the die budget given to dGPUs and iGPUs. Even if iGPUs somehow gain superior bandwidth, the die space allotted for mid-high end dGPUs allow for vastly superior compute and shader performance. Bandwidth alone will not bring Intel to parity due to the gap in sheer compute power as Intel only has a fraction of a die to spend on GPU, while Nvidia and AMD have dies several times larger than Intel's entire cpu die, the entirety of which is spent on the GPU.

For Intel to pull ahead of high end dGPUs would require a vast process advantage so as to fit the required compute units needed to overtake AMD and Nvidia in compute performance.

Remember Intels node is 1 ahead, closing in on 2.

A 14nm GK104 for example would be less than 100mm2. Maybe as low as 75mm2. And parts of its is redundant due to being tied with the CPU.

 

[Ajay]

Until we run into the diesize.

I know I would rahter have a dedicated CPU die and GPU die...than 2 "universal dies"

I don't want my CPU trying to be a midget GPU...I want it to be a CPU...if I need a GPU...I will buy one.

IGP's are overhyped if you look at their performance...

Cache used to be on a separate chip.
The FPU used to be on a separate chip.
The MMU used to be on a separate chip.
Etc., etc., etc....

There is this funny concept call "integrated circuits" that started to become popular in the 1950's "/sarcasm"

 

[Homeles]

No it is not, Kaveri which will be the highest performance mainstream APU in 2014 both from AMD and Intel, will only have HD7750(DDR3) dGPU performance. Even if Kaveri had 2x the Stream Processors it would gain no tangible extra performance due to lack of Memory Bandwidth. Same applies to Intel iGPUs, HD5000(GT3) is not that Faster than HD4600(GT2) although it has double the EUs.

You clearly haven't been paying attention to this discussion. Stacked DRAM solves the bandwidth issue, which has been mentioned probably about a dozen times in this thread.

Why don't you sit this thread out, since you seem to be so far out of the loop?

Kaveri will only manage to get close to HD7750(DDR3) performance. dGPUs have that performance for 2+ years now at that price/TDP point.

Yeah, because AMD has fallen severely behind with their APU time to market.

Intel, on the other hand, is on a clear path to intersect that level of performance.

Without eDRAM it will not have a chance. Broadwell GT2 could only catch HD5000 at the most which is way lower than HD7750.

Once again, there's this thing called stacked memory that you should take a look at. AMD will be implementing it too, which I'm sure you'll love to hear since that's your agenda and all.

20nm dGPUs are one year away. But when they will arive, mainstream performance will double. That means we will get ~7850 performance at $150 or less price point. Broadwell in early 2015 will be pressed to compete against mainstream 20nm dGPUs and it will loose big time.

7850s already hover not far from the $150 price point.

And once again, you fail to see the big picture. What comes after 20nm? A node with virtually nonexistent areal savings. Do you not see how non-Intel GPUs will stagnate?

Intel has the same problem AMD and NVIDIA has, Memory Bandwidth and cost of die size. Even with 14nm Intel will not be able to have a mainstream iGPU at HD7750 GDDR5 level of performance in 2014.

They'll be quite close though. Once again, the gap is closing. Very obviously too -- do I need to make a chart or something for you?

But, AMD and Intel dont throw all that money to APUs for Gaming performance. Compute is the name of the game and that will be seen with Kaveris hUMA and HSA.

Ah, there's the shameless AMD plug. I was waiting for it.

 

[Homeles]

This is false.

Do take a look at the die budget given to dGPUs and iGPUs. Even if iGPUs somehow gain superior bandwidth, the die space allotted for mid-high end dGPUs allow for vastly superior compute and shader performance. Bandwidth alone will not bring Intel to parity due to the gap in sheer compute power as Intel only has a fraction of a die to spend on GPU, while Nvidia and AMD have dies several times larger than Intel's entire cpu die, the entirety of which is spent on the GPU.

As Shintai pointed out, it'd be a good idea for you to analyze the foundry situation.

For Intel to pull ahead of high end dGPUs would require a vast process advantage (no, a 14nm vs 28nm is not vast) so as to fit the required compute units needed to overtake AMD and Nvidia in compute performance.

You can't be serious. Even Intel's 32nm process blew TSMC's 28nm process out of the water in terms of switching speed.

 

[mrmt]

Do take a look at the die budget given to dGPUs and iGPUs. Even if iGPUs somehow gain superior bandwidth, the die space allotted for mid-high end dGPUs allow for vastly superior compute and shader performance. Bandwidth alone will not bring Intel to parity due to the gap in sheer compute power as Intel only has a fraction of a die to spend on GPU, while Nvidia and AMD have dies several times larger than Intel's entire cpu die, the entirety of which is spent on the GPU.

What is happening to the delta between the die size of the dGPU chip and the die size devoted to the iGPU? It is shrinking. Once you move memory to the die then the dGPU market will be wiped out for everyone but Lonbjergs that need huge GPU power.

 

[AtenRa]

You clearly haven't been paying attention to this discussion. Stacked DRAM solves the bandwidth issue, which has been mentioned probably about a dozen times in this thread.

dGPUs will use stacked memory as well, and since they can use higher TDP than APUs they will maintain their performance advantage.

Why don't you sit this thread out, since you seem to be so far out of the loop?

Personally attacking me will not change the fact that dGPUs can have higher performance due to larger die size and higher TDP limits.

Yeah, because AMD has fallen severely behind with their APU time to market.

Intel, on the other hand, is on a clear path to intersect that level of performance.

Intel is making progress but they heavily relying on manufacturing process to make up for there lack in Architecture design. AMD with kaveri for the first time will have the same dGPU architecture in their APUs. From now on iGPUs will be on par with dGPUs both in Architecture/Designs as well as manufacturing process.

Once again, there's this thing called stacked memory that you should take a look at. AMD will be implementing it too, which I'm sure you'll love to hear since that's your agenda and all.

As i have said before, dGPUs will also use Stacked Memory. You seam to forget that.

7850s already hover not far from the $150 price point.

Yes my bad, I meant $99.00

And once again, you fail to see the big picture. What comes after 20nm? A node with virtually nonexistent areal savings. Do you not see how non-Intel GPUs will stagnate?

GloFos 14XM and TSMC 16nm is only meant for low power mobile/phones chips. No HighEnd dGPUs will use those two process. Next process for HighEnd dGPUs will be 10nm.

They'll be quite close though. Once again, the gap is closing. Very obviously too -- do I need to make a chart or something for you?

HD7750 (GDDR5) is more than 2x faster than Haswell GT3, how they'll be close ??? Intel will only catch HD7750 with Broadwell GT4 and eDRAM maybe. But Broadwell GT4 will not be priced at $99.

Ah, there's the shameless AMD plug. I was waiting for it.

Do you actually believe that Intel is investing money, time, resources, die size and more for Games ??? Nope, they invest in Compute. Games only benefiting from the higher performance of the APUs.

 

[Fjodor2001]

Cache used to be on a separate chip.
The FPU used to be on a separate chip.
The MMU used to be on a separate chip.
Etc., etc., etc....

There is this funny concept call "integrated circuits" that started to become popular in the 1950's "/sarcasm"

The difference is that you only need one MMU, but GPU cores you'll always benefit from having more of (as long as there is no memory bandwidth bottleneck etc). So GPU cores can always benefit from being given more die area to consume. But there is no point in adding e.g. additional MMUs, or more cache once the cache hit rate is sufficient.

 

[Homeles]

dGPUs will use stacked memory as well, and since they can use higher TDP than APUs they will maintain their performance advantage.

While the TDP issue is a good point, you should understand that dGPUs, particularly ones on the lower end of the spectrum, will benefit less from stacked memory. APUs will benefit most, because they're quite starved at this point in time.

Intel is making progress but they heavily relying on manufacturing process to make up for there lack in Architecture design. AMD with kaveri for the first time will have the same dGPU architecture in their APUs. From now on iGPUs will be on par with dGPUs both in Architecture/Designs as well as manufacturing process.

Intel's addressing that shortcoming with Broadwell.

GloFos 14XM and TSMC 16nm is only meant for low power mobile/phones chips. No HighEnd dGPUs will use those two process. Next process for HighEnd dGPUs will be 10nm.

So the process gap will worsen? That only makes my argument stronger.

HD7750 (GDDR5) is more than 2x faster than Haswell GT3, how they'll be close ??? Intel will only catch HD7750 with Broadwell GT4 and eDRAM maybe. But Broadwell GT4 will not be priced at $99.

Perhaps my prediction was a bit optimistic, but at the end of the day, the gap is closing.

Really, you should have just taken me up on the chart offer.

Do you actually believe that Intel is investing money, time, resources, die size and more for Games ??? Nope, they invest in Compute. Games only benefiting from the higher performance of the APUs.

Rather than resorting to a straw man, why don't you show me where I said anything about games? It doesn't matter where they're putting their resources -- both applications benefit mutually.

 

[mikk]

Intel is making progress but they heavily relying on manufacturing process to make up for there lack in Architecture design.

I wouldn't count on it for Gen8 and Gen9. Intel brings two new GPU generations within the next 2 years. They should close the gap or could maybe overtake AMD/Nvidia. Broadwell is per EU obviously much faster than Gen7.5. Also twice the sampler per EU should make better use of Intels Gflops. I wouldn't be so sure that Broadwell-K GT3e with extrapolated 1,2-1,3 Tflop can't match a HD7750. Gen8 is too different. If Intel can close the architectural deficit they are in a very good position with the huge process advantage. In the long-term AMD lost the APU battle to Intel, at the latest with Skylake/Goldmont Gen9 GPU they lost it over the whole APU segment range.

 

[AtenRa]

What is happening to the delta between the die size of the dGPU chip and the die size devoted to the iGPU? It is shrinking.

Really ???

Haswell GT3 iGPU(40 EUs) at 22nm is close to 80-100mm2. Iris Pro HD5200 (eDRAM) is slower than GT640.

HD7750 at 28nm is 123mm2 including the 128-bit memory controller. It will be close to 100mm2 without the memory controller. HD7750 is 50% faster than GT640.

That means at almost the same die size and with a process advantage Intel iGPUs are 50% slower than dGPUs.

I wouldn't call that shrinking

 

[Fjodor2001]

I wouldn't count on it for Gen8 and Gen9. Intel brings two new GPU generations within the next 2 years. They should close the gap or could maybe overtake AMD/Nvidia. Broadwell is per EU obviously much faster than Gen7.5. Also twice the sampler per EU should make better use of Intels Gflops. I wouldn't be so sure that Broadwell-K GT3e with extrapolated 1,2-1,3 Tflop can't match a HD7750. Gen8 is too different. If Intel can close the architectural deficit they are in a very good position with the huge process advantage. In the long-term AMD lost the APU battle to Intel, at the latest with Skylake/Goldmont Gen9 GPU they lost it over the whole APU segment range.

But it's not like AMD will be sitting still on the GPU front either.

Also, AMD is closing the CPU performance gap to Intel, and is delivering CPU performance advances at a faster pace currently. Looking forward to Broadwell vs Excavator that trend looks to continue.

 

[Sweepr]

HD7750 (GDDR5) is more than 2x faster than Haswell GT3, how they'll be close ??? Intel will only catch HD7750 with Broadwell GT4 and eDRAM maybe. But Broadwell GT4 will not be priced at $99.

I dont think you have enough knowledge about Gen 8 Graphics Architecture or Broadwell specs to make such a statement.

 

[mikk]

But it's not like AMD will be sitting still on the GPU front either.

Also, AMD is closing the CPU performance gap to Intel, and is delivering CPU performance advances at a faster pace currently. Looking forward to Broadwell vs Excavator that trend looks to continue.

AMDs GPU progress is slow, Kaveri continues this trend and from what we know they are still coming with 28nm APUs 2015 which can't be a match for Skylake on 14nm. Also Intel is still using DDR3-1600, AMD is currently on DDR3-2133 and is going to DDR3-2400 with Kaveri. Once Intel switches to DDR4-2400 with Skylake they get a healthy bandwidth boost. AMD can't get the same RAM boost.

 

[Zodiark1593]

I dont think you have enough knowledge about Gen 8 Graphics Architecture or Broadwell specs to make such a statement.

Agreed. There are points from both sides of the debate, and I still stand by my points, another internet fight is pointless when no one here knows what will happen in the future.

Best I can hope for is the future chips, be it dGPU and/or iGPU are suitable for both the average joe, and those that want for vast amounts of compute power.

 

[AtenRa]

While the TDP issue is a good point, you should understand that dGPUs, particularly ones on the lower end of the spectrum, will benefit less from stacked memory. APUs will benefit most, because they're quite starved at this point in time.

This point in time, next year with 20nm they will not because $99.00 dGPU segment will have HD7870 performance. And they will need more memory bandwidth.

Intel's addressing that shortcoming with Broadwell.

Yes and dGPUs will get a full node process shrink at 20nm at the same time Broadwell will arrive. The performance Delta will remain the same.

So the process gap will worsen? That only makes my argument stronger.

Nope, 20nm will be used one year earlier than GloFo14XM and 16nm TSMC. 10nm will come on time to replace the 20nm process, not the 14/16nm.

Perhaps my prediction was a bit optimistic, but at the end of the day, the gap is closing.

See my post above (Haswell GT3 die size) and you will see that the gap is not closing.

Really, you should have just taken me up on the chart offer.

By all means, make one.

Rather than resorting to a straw man, why don't you show me where I said anything about games? It doesn't matter where they're putting their resources -- both applications benefit mutually.

You were debating with Lonbjerg about HighEnd Gaming Discrete GPUs, did you not ??

 

[Haserath]

Really ???

Haswell GT3 iGPU(40 EUs) at 22nm is close to 80-100mm2. Iris Pro HD5200 (eDRAM) is slower than GT640.

HD7750 at 28nm is 123mm2 including the 128-bit memory controller. It will be close to 100mm2 without the memory controller. HD7750 is 50% faster than GT640.

That means at almost the same die size and with a process advantage Intel iGPUs are 50% slower than dGPUs.

I wouldn't call that shrinking

Intel's 22nm is equal to TSMC's 28nm area wise. 20nm is looking to be much less of a gain for TSMC & especially their customers. HD 7790 level GPU looks like it will be the next $100 GPU.

 

[Homeles]

This point in time, next year with 20nm they will not because $99.00 dGPU segment will have HD7870 performance. And they will need more memory bandwidth.

I wasn't arguing otherwise, in regards to the first statement. As for the second, that only benefits my argument.

Yes and dGPUs will get a full node process shrink at 20nm at the same time Broadwell will arrive. The performance Delta will remain the same.

And how about the year after that? You are thinking too short-term.

Nope, 20nm will be used one year earlier than GloFo14XM and 16nm TSMC. 10nm will come on time to replace the 20nm process, not the 14/16nm.

Do you have a source on when TSMC/CP will have 10nm?

See my post above (Haswell GT3 die size) and you will see that the gap is not closing.

You're basing your argument on a single data point?

You were debating with Lonbjerg about HighEnd Gaming Discrete GPUs, did you not ??

Did you not realize that Lonbjerg is all alone in his little world?

Compute performance or gaming performance -- doesn't matter which in regards to this conversation. APUs will catch dGPUs on both metrics.

APUs have more potential than dGPUs for compute anyway. That potential is being slowly tapped into as time progresses.

Intel's 22nm is behind TSMC's 28nm area wise. 20nm is looking to be much less of a gain for TSMC & especially their customers. HD 7790 level GPU looks like it will be the next $100 GPU.

Intel's 22nm is acutally denser than TSMC's 28nm. It's hard do see, but it's there on that graph.

But yes, Intel's 14nm and 10nm will pull much farther ahead.

 

[AtenRa]

Intel's 22nm is equal to TSMC's 28nm area wise. 20nm is looking to be much less of a gain for TSMC & especially their customers. HD 7790 level GPU looks like it will be the next $100 GPU.

Heh, nobody show it or was unwilling to say anything about it.

If TSMC 28nm is equal to Intel 22nm in density, then 20nm TSMC that will have almost 2x density over 28nm will be on par with Intels 14nm.

Unless Intel 14nm has 4x the density of 22nm then the above diagram is wrong and way misleading. Purely PR BS,

 

[AtenRa]

I wasn't arguing otherwise, in regards to the first statement. As for the second, that only benefits my argument.

And how about the year after that? You are thinking too short-term.

Do you have a source on when TSMC/CP will have 10nm?

You're basing your argument on a single data point?

Did you not realize that Lonbjerg is all alone in his little world?

Compute performance or gaming performance -- doesn't matter which in regards to this conversation. APUs will catch dGPUs on both metrics.

APUs have more potential than dGPUs for compute anyway. That potential is being slowly tapped into as time progresses.

Intel's 22nm is acutally denser than TSMC's 28nm. It's hard do see, but it's there on that graph.

But yes, Intel's 14nm and 10nm will pull much farther ahead.

Why dont you make that graph of yours with both AMD and Intel APUs and GPUs and see what happens the next 1-2 years. ??? You will find that APUs will not overcome dGPUs performance in the mainstream market (up to A10 and Core i3).