Now that I think of it.... why aren't they making HBM DIMMs?

Because latency is too high for GPU performance and currently CPU's aren't really bottlenecked by memory performance, not to mention that HBM costs a lot more to manufacture. Stacking isn't really a problem, they can easily just reconfigure the heatspreader and make it work with a little extra z height. I don't know if the initial Zen release will have HBM on it, but I definitely could see them doing that in the future. It will give a huge increase to iGPU performance. Plus they are going to need something that eventually goes up against 3D Xpoint, which I assume Intel will start using as L4 cache eventually.

Good news , this means i will finally retire my old , outdated , without support any longer , first generation APU FM1 platform i currently use for htpc when the new APUs comes out. DDR4 high bandwidth will bring what it is needed for the integrated graphics , bring it on and i will jump on it the first day it is release 🙂.

I'm not sure if that would be cost effective, let alone physically possible (remember, HBM is stacked. there isn't much vertical space in a CPU die). IMO, it would make more sense for them to just alter the memory controller to support HBM DIMMs. For CPU performance, this wouldn't accomplish much but it would make the IGP run so much better that getting something like an A10 crossfired with a discrete GPU could probably qualify as an upper-mid range gaming system. Now that I think of it.... why aren't they making HBM DIMMs?

If it was physically possible for the huge 28nm die of the Fury X, it will be for a smaller and much cooler 14nm CPU. It won't be on die, it would be on an interposer. As for the performance, it would accomplish a ton actually. Broadwell wasn't really popular, but the model with the 128MB eDRAM that could be used as an L4, was faster in gaming than a 700Mhz higher clock Haswell i7. It also had 50-80% lower latencies/stutter while gaming. Integrating a bit of HBM could do miracles for Zen, and AMD has both the know how and the patents to do it. Look at the latencies and performance here. Keep in mind that the CPU is much lower clocked than the other top i7. They say constantly how the eDRAM helps with performance.

Seeing as AMD's IGPs have always been crippled by memory, I think the extra channels will make a pretty big difference (though, 8 might be overkill...). Without the IGP, 8 channels just sounds stupid, even for 32 cores. For the average CPU task, you can hardly ever notice a performance difference between 1 and 2 channels. I'm not talking about synthetic benchmarks here, I'm talking about real-world performance.

Oh yeah, look at the memory scaling on the APU's! They actually do not get a whole lot of performance gain from overclocking the cpu cores or the GPU cores even, but the memory is the best for them! Hoping once the Zen APU's are released, my Alpha will be replaced by one.

Hoping once the Zen APU's are released, my Alpha will be replaced by one.

Your alpha? Do you mean your main rig? Do you really see amd besting an Intel 5820k?

That's just an assumption man, there's zero information on how AMD will price anything.

True enough, but if it's no cheaper, nobody will buy a previously totally unknown AMD CPU over an Intel architecture that has already been proven to be good in every other way but the high price. Up until now AMD has tried to compete with lower prices. Of course it's in deep trouble financially due to being in the red all the time, but it seems doubtful expensive first generation would help them. A 4-core i7 K price is still a lot of money. More than AMD has been asking for any consumer CPU for a long time. But like you said, it's nothing but my wishful thinking.

Your alpha? Do you mean your main rig? Do you really see amd besting an Intel 5820k?

I would hope that AMD can best a 22nm Haswell-E part that came out 2 years prior to Zen. If they can't AMD might as well just close it's doors now.

I would hope that AMD can best a 22nm Haswell-E part that came out 2 years prior to Zen. If they can't AMD might as well just close it's doors now.

The problem is I can guarantee you that Zen will not outperform a Haswell-E in every test, and as a result, people will automatically dismiss it as a disappointment or failure. There's a good chance it will outperform Intel in many, maybe even most tests, but not 100% of them, and for some people around here that just won't be good enough. People need constant reminders that no CPU is a one-size-fits-all and we're reaching the physical limits of silicon. Outperforming Intel in terms of benchmarks is hardly an objective worth pursuing, but nobody will realize that because they have the mentality that "moar iz better". But [Intel] CPUs are and have been "good enough" for a long time; what we need to focus on is power efficiency. That being said, even if Zen is slower than Haswell-E, if it can outperform Skylake in performance-per-watt, then it's a winner.

These days more than 4 cores might start to help...despite Intel's best efforts to tell people nobody needs more than four cores.

Yes! Intel's conspiratorial campaign to misinform people must end! Those 20 core Xeons are just to throw people off the truth! AMD certainly hasn't tried to compete by offering 8 core CPU's that are missing 1/2 the FPU's and perform worse than i3's in most real world applications.

Yes! Intel's conspiratorial campaign to misinform people must end! Those 20 core Xeons are just to throw people off the truth!

Though I don't know where Intel was quoted saying people only need 4 cores, they're not wrong. Those 20 core Xeons you speak of are not "for people", they're for high-end servers and mainframes. I'd like to see you buy one of those processors without taking out a loan first.

AMD certainly hasn't tried to compete by offering 8 core CPU's that are missing 1/2 the FPU's and perform worse than i3's in most real world applications.

Irrelevant. For most real world applications, the average person would be content with something worse than either of those. In real word environments, the speed difference in applications where an i3 is faster than an 8-core FX isn't noticeable, but where the i3 isn't faster, the performance gap is noticeable.

The problem is I can guarantee you that Zen will not outperform a Haswell-E in every test, and as a result, people will automatically dismiss it as a disappointment or failure. There's a good chance it will outperform Intel in many, maybe even most tests, but not 100% of them, and for some people around here that just won't be good enough. People need constant reminders that no CPU is a one-size-fits-all and we're reaching the physical limits of silicon. Outperforming Intel in terms of benchmarks is hardly an objective worth pursuing, but nobody will realize that because they have the mentality that "moar iz better". But [Intel] CPUs are and have been "good enough" for a long time; what we need to focus on is power efficiency. That being said, even if Zen is slower than Haswell-E, if it can outperform Skylake in performance-per-watt, then it's a winner.

Even though it would render my 1000$ cpu worthless, I really hope AMD does infact achieve a better performance than it except im a little sceptical that it will. I dont think they really need to either, if amd can have an ipc nearly as good as Haswell, it would be enough. Intel has left a very wide margin of space in pricing, it'll be nice to have amd topple it down like a house of Cards.

Everyone but everyone forget one simple thing:ze Chipset. Maybe Bulldozer was good,maybe Excavator was good,maybe Steamroller was good in Labs,but in real life they are not because ze Chipset. All of you talk about Cpu but the main culprit in bad performance is ze Chipset,without him and the chipset software to push Cpu at work then all is for nothing in this case. So Zen without a good ze Chipset and without good chipset software is Nothing. 🙂

Your alpha? Do you mean your main rig? Do you really see amd besting an Intel 5820k?

Alienware Alpha. It's my rig in my living room. <3

I only care about the APU's, an APU with ddr4 support would be amazing!!!

That ddr 4 support won't mean jack if Zen underperforms. I really hope AMD can pull that magic rabbit out of their hat.

That ddr 4 support won't mean jack if Zen underperforms. I really hope AMD can pull that magic rabbit out of their hat.

Faster memory for the GPU, that's what I am interested in the most.

Zen is AMD's take on next-gen frying pans.

I can guarantee you

Aww, i love how you can guarantee stuff

Aww, i love how you can guarantee stuff

Lol that's hilarious. I wish I could guarantee stuff, at best I can certify

Seeing as AMD's IGPs have always been crippled by memory, I think the extra channels will make a pretty big difference (though, 8 might be overkill...). Without the IGP, 8 channels just sounds stupid, even for 32 cores. For the average CPU task, you can hardly ever notice a performance difference between 1 and 2 channels. I'm not talking about synthetic benchmarks here, I'm talking about real-world performance.

There will not be APU with 8 ch DDR4. 8 channels makes perfect sense for the 32-core processor. Remember, it is two 16 core dies that each have a quad channel memory controller. Sure desktop apps don't really need memory b/w that much, but with 32 cores + HT, thats 64 threads, all competing for memory b/w. AMD did this right.

Now that I think of it.... why aren't they making HBM DIMMs?

Because latency is too high for GPU performance and currently CPU's aren't really bottlenecked by memory performance, not to mention that HBM costs a lot more to manufacture. Stacking isn't really a problem, they can easily just reconfigure the heatspreader and make it work with a little extra z height. I don't know if the initial Zen release will have HBM on it, but I definitely could see them doing that in the future. It will give a huge increase to iGPU performance. Plus they are going to need something that eventually goes up against 3D Xpoint, which I assume Intel will start using as L4 cache eventually.

It has nothing to do with latency. it has to do with the bus width. Remember the silicon interposer and TSV's and micro bumps are what made HBM possible. A single DDR chip is 8 bits wide, yes just 8. They use 8 of them on a DIMM to make a 64-bit wide DDR stick. This has been the same since PC66 SDRAM. HBM is 1024-bit wide PER CHIP. That's how Fiji has 4096-bit bus, it has 4 chips. You need literally thousands and thousands of pins, probably at least 2 per bit (for differential signalling) plus control and power pins. That would mean, that even if you had a SINGLE CHIP on DIMM, it would need over 2,000 pins. A current DDR4 stick has 288 pins. This is basically not possible. That's why you need the interposer is because it can connect up all those thousands of pins in a really small space. So any CPU with HBM will need to have an interposer below it, just like Fiji. This would make it a LOT more expensive. Don't expect to see HBM on a cpu any time soon. You could theoretically stack the HBM directly on the CPU die, but then you would not be able to cool the cpu die directly. Basically a non-starter. You need the interposer in this case, just like Fiji needs it. EDIT: Also 3D XPOINT is not a good idea for L4 cache. It is MUCH slower than SRAM. SRAM will always be used as cpu cache. Plus 3D XPOINT has limited r/w cycles, much higher than flash, but still not infinite like SRAM. 3D XPOINT makes zero sense for CPU cache. There is no need for non volatile memory for cache anyways. 3D XPOINT is for applications where the non-volatile aspect is important.

EDIT: Also 3D XPOINT is not a good idea for L4 cache. It is MUCH slower than SRAM. SRAM will always be used as cpu cache. Plus 3D XPOINT has limited r/w cycles, much higher than flash, but still not infinite like SRAM. 3D XPOINT makes zero sense for CPU cache. There is no need for non volatile memory for cache anyways. 3D XPOINT is for applications where the non-volatile aspect is important.

I'm too off now to discuss the rest, but high speed HBM is close to SRAM in any respect. Also Broadwell had eDRAM, not even SRAM and it managed to get faster performance for gaming latencies and in general than a 700MHz Haswell.