AMD Zen - Key Dates and Information

[SpaceBeer]

The thing is - AMD can also offer power saving and innovative form factors. Take a look at Carrizo and Bristol Ridge. If AMD had Zen for those APUs, they would actually had better offer for laptops, SFF builds, etc. since their IGP is lot faster (except expensive Iris Pro CPUs).

 

[Nereus77]

The thing is - AMD can also offer power saving and innovative form factors. Take a look at Carrizo and Bristol Ridge. If AMD had Zen for those APUs, they would actually had better offer for laptops, SFF builds, etc. since their IGP is lot faster (except expensive Iris Pro CPUs).

Zen will have an APU range, but we haven't heard much about it. Hopefully it also sips power and spits performance.

 

[bjt2]

The thing is - AMD can also offer power saving and innovative form factors. Take a look at Carrizo and Bristol Ridge. If AMD had Zen for those APUs, they would actually had better offer for laptops, SFF builds, etc. since their IGP is lot faster (except expensive Iris Pro CPUs).

Iris Pro was faster for the 128MB of L4 and also the 14nm advantage. With 14nm the clocks can go up even on AMD GPUs and if they use HBM2 as cache...

 

[CatMerc]

Iris Pro was faster for the 128MB of L4 and also the 14nm advantage. With 14nm the clocks can go up even on AMD GPUs and if they use HBM2 as cache...

Pretty much. AMD's GPU tech is miles ahead of whatever Intel has. Even with all the advantages in the world like the L4 cache and far higher transistor budget, they still can only JUST beat AMD's APU's. IIRC A12 9800 should be a bit faster than Iris Pro. On 14nm, and given HBM or some other form of L4 cache, AMD's APU's would be far far faster.

 

[itsmydamnation]

Zen APU is Vega GPU as well so hopefully we will see a lot less bandwidth required vs Excavator IGP per unit of performance, better mem compression, bigger L2 and better culling in general equals less bandwdith.

 

[Atari2600]

Zen APU is Vega GPU as well so hopefully we will see a lot less bandwidth required vs Excavator IGP per unit of performance, better mem compression, bigger L2 and better culling in general equals less bandwdith.

Are there versions of Vega that don't use HBM?

Agreed on what you say BTW, much of the modifications to Vega are driving at minimising the workload involved in producing 3D graphics as opposed to just increasing the work capacity.
[While this may not not necessarily help compute, it should make graphics performance more palatable.]

 

[SpaceBeer]

There is no way and consumer APU will have HBM2 in near (like 3 years or so) future. Please don't expect laptops with HBM. Not even with highly clocked DDR4. The point is - APUs will replace Bristol Ridge and will be great for home/office/media use.

 

[bjt2]

There is no way and consumer APU will have HBM2 in near (like 3 years or so) future. Please don't expect laptops with HBM. Not even with highly clocked DDR4. The point is - APUs will replace Bristol Ridge and will be great for home/office/media use.

How much does Fiji cost, with its 4 HBM chips and the interposer? Less than 400$ with board, liquid cooling and so on. So the chip alone probabily costs less than 350$. How much does a top bin ULV Mobile APU cost? The prices are comparable even if you put 4 HBM(2) stacks. Considering that they can also put only one stack, a top bin mobile APU 4c/8t with 2/4/8GB used as L4 cache or even plain RAM (along with some optional DDR4 module to increase capacity) could be feasible even at 200-250$...

 

[Atari2600]

There is no way and consumer APU will have HBM2

If you said entry level, I'd tend to agree with you.

If you said top-level, I'd disagree.

http://techreport.com/review/31224/the-curtain-comes-up-on-amd-vega-architecture


The Vega GPU:

Note, AMD don't term the memory controller a memory controller any more! Its now cache controller, and integrates with both the onboard HBM (now termed HBC), and the system memory. This approach to memory system architecture allows for somewhat decoupling the memory control of the GPU from the CPU (a change in direction from previous AMD roadmaps) - allowing a subset of HBM without it having to comprise all the system memory. Assuming that this generation of APU won't have to run games on anything beyond a HD screen, 2GB should be sufficient HBM to feed the GPU.

 

[SpaceBeer]

Ok, yes. There might be APU with HBM but probably more expensive than combo of CPU+GPU of same power. Plus, IGP can't have the same number of CUs as descrete GPU cards. Could they fit more than 14-16 CUs in those chips? And cards such as RX 460 certainly don't need HBM. Such APU would actually be more usefull for other (compute) tasks than gaming

 

[Qwertilot]

Still a market for such things - look how pricey the iris pro NCU is. Honestly if AMD were in a rather better financial position they'd probably already have done one using HBM1, just to have a genuinely class leading part.

Not sure if/when they'll actually do one as things stand. The thing that would make it very likely to appear would be HBM becoming due to appear in the consoles.

 

[Atari2600]

Ok, yes. There might be APU with HBM but probably more expensive than combo of CPU+GPU of same power. Plus, IGP can't have the same number of CUs as descrete GPU cards. Could they fit more than 14-16 CUs in those chips? And cards such as RX 460 certainly don't need HBM. Such APU would actually be more usefull for other (compute) tasks than gaming

Yes, well no one is pretending its going to be capable of running a 4k screen at 60 fps!

If it can run pretty much any game at 1080 at 60 fps on good settings, that will do.


Moving away from games - if AMD can build an APU that can function sufficiently for CATIA/Autocad/Maya etc, then professionals will eat it up. If they could build a hardware+network+software* solution that would allow for designers to use during the day, then analysts to use for distributed HPC overnight, they'd be onto a massive winner.

*which would need to be able to add/remove nodes as users log out and log in.

 

[Ancalagon44]

If you think about it, such an SKU (CPU + GPU + 2GB of HBM) is perfect for a company like Apple. It ticks all of their boxes - fast and power efficient CPU, fast and power efficient GPU, fast and power efficient memory. The HBM is just for the GPU.

 

[SpaceBeer]

Fast CPU + fast GPU + fast, low latency, high bandwidth memory is perfect for any compute-intensive task. Especially if it's power efficient and not expensive

 

[MrTeal]

Ok, yes. There might be APU with HBM but probably more expensive than combo of CPU+GPU of same power. Plus, IGP can't have the same number of CUs as descrete GPU cards. Could they fit more than 14-16 CUs in those chips? And cards such as RX 460 certainly don't need HBM. Such APU would actually be more usefull for other (compute) tasks than gaming

Cards like the 460 don't need HBM to support 14 CUs because they have a fast 128-bit GDDR5 bus to dedicated memory giving >100GB/s. An APU on the other hand gets to share the 34GB/s (for dual channel DDR4-2133) with the CPU, and hopefully the laptop vendor has bothered to populate both channels. Even a single stack of HBM1 would go a long way to improving GPU performance on an APU laptop.

Cost and packaging are another issue though, as that solution is going to be quite a bit more expensive than a traditional APU, with a larger package and greater z height. Total size will be much smaller than a CPU and GPU combo though, so an HBM solution could find a place in a high performance premium light product.

 

[Thunder 57]

Would it make sense to offer a Zen APU with some sort of high speed memory to act as a cache, backed by dual channel DDR4? The XB1 does a good job at matching the PS4 with much less memory bandwidth this way. Intel has used eDRAM. I would imagine a small amount of GDDR5 or HBM could be used as well.

I think the reason we haven't seen anything like this is because AMD isn't exactly floating in money, and the CPU performance on current APU's just isn't there for it to make a whole lot of sense.

EDIT

Personally I wouldn't expect to see eSRAM or HBM. Perhaps HBM2 at some point.

How much would say, 2GB 128 bit 7-8GHz GDDR5 help? I can't imagine it costing too much but all of a sudden you have a pool of memory with 3-4x the bandwidth.

 

[TemjinGold]

Fast CPU + fast GPU + fast, low latency, high bandwidth memory is perfect for any compute-intensive task. Especially if it's power efficient and not expensive

But that's the point. As an Apple machine, it doesn't need to be inexpensive.

 

[Glo.]

Ok, yes. There might be APU with HBM but probably more expensive than combo of CPU+GPU of same power. Plus, IGP can't have the same number of CUs as descrete GPU cards. Could they fit more than 14-16 CUs in those chips? And cards such as RX 460 certainly don't need HBM. Such APU would actually be more usefull for other (compute) tasks than gaming

299$ for 4C/8T plus 16 CU GPU and 2 HBM2 stacks in 95W TDP.

35W for the same setup in mobile APUs, made specifically for Apple for example.

 

[ShintaiDK]

299$ for 4C/8T plus 16 CU GPU and 2 HBM2 stacks in 95W TDP.

35W for the same setup in mobile APUs, made specifically for Apple for example.

So you add 408/512GB/sec memory bandwidth for something whos TDP would at best use 80-100GB/sec?

And the mobile one? Oh that was funny!

Sounds like another fairy tale and unicorns product.

 

[Atari2600]

So you add 408/512GB/sec memory bandwidth for something whos TDP would at best use 80-100GB/sec?

1 HBM stack is enough.

It would be quite the idiot that couldn't see that a combination of Zen+Vega having a single stack of HBM (2GB or less) would be a potent mix.

 

[crashtech]

A Zen APU could definitely be a threat to Intel in the mobile space, it looks like the CPU may rival Intel in IPC, and AMD GPUs are pretty obviously better. The thing that has hobbled APUs has been memory bandwidth and TDP, both these problems seem on the verge of a solution.

 

[coffeemonster]

Pretty much. AMD's GPU tech is miles ahead of whatever Intel has. Even with all the advantages in the world like the L4 cache and far higher transistor budget, they still can only JUST beat AMD's APU's. IIRC A12 9800 should be a bit faster than Iris Pro. On 14nm, and given HBM or some other form of L4 cache, AMD's APU's would be far far faster.

this youtube channel has been showing off Bristol Ridge gaming. https://www.youtube.com/channel/UCV_FbbkkWz4KHNzMlmYO04A/videos fairly impressive, even compared with Kaveri APU's

 

[Rifter]

Yeah if AMD can get the IPC/TDP under control then their APU should destroy anything intel can come up with. This will likely be their big moneymaker in 2017/2018 will be a ZEN APU.

 

[DrMrLordX]

Ok, so MSI is the worst, but their AMD boards have lower failure rate than those with Intel chipsets

Throttling may not be recorded as motherboard failure.

 

[AtenRa]

There is no way and consumer APU will have HBM2 in near (like 3 years or so) future. Please don't expect laptops with HBM. Not even with highly clocked DDR4. The point is - APUs will replace Bristol Ridge and will be great for home/office/media use.

I believe HBM2 will be mainstream available from H2 2018 and onward. We may have second gen ZEN APUs with 4-8GB of Single stack HBM2 memory in late 2018.