AMD ZEN dieshot and structure

[Trixanity]

3 minutes ago, wcreek said:

Ah okay, well still I think AMD needs to work on their single threaded IPC. But yeah if an 8 Core Zen FX CPU  had the near the single core score of like an i5 or i7 from Z97/Z170 but a multicore score of something from X99 around $300 to $400, I think then AMD has got something pretty good.

That's what Zen is. An architecture with IPC to match Intel (hypothetically speaking for now). IPC basically means  the performance per core per clock. So a high IPC (or similar IPC to Intel) would mean similar single threaded performance and from there it will scale with cores meaning that multi threaded work loads will also be similar to Intel.

 

Oh and there shouldn't be any IPC difference between Haswell and Haswell-E processor. They're based on the same architecture. There are of course big differences but in many ways, they are the same. 

 

Only thing you got wrong was the price. An 8 core SKU will not launch at a discount when there is no reason to. 

[That Norwegian Guy]

Awesome, Ivy Bridge IPC was what I hoped for. With 8 cores at a fair price with hopefully better cache than an actual Ivy Bridge CPU and plenty of overclocking headroom. DDR4 is uneccessary but a bonus I guess. Hopefully some good motherboards with up to date technology.

 

I'm excited. Bring on an mATX premium board.

[handymanshandle]

1 minute ago, Trixanity said:

That's what Zen is. An architecture with IPC to match Intel (hypothetically speaking for now). IPC basically means  the performance per core per clock. So a high IPC (or similar IPC to Intel) would mean similar single threaded performance and from there it will scale with cores meaning that multi threaded work loads will also be similar to Intel.

 

Oh and there shouldn't be any IPC difference between Haswell and Haswell-E processor. They're based on the same architecture. There are of course big differences but in many ways, they are the same. 

 

Only thing you got wrong was the price. An 8 core SKU will not launch at a discount when there is no reason to. 

I know what IPC means, iirc Haswell-E was a little slower or the same single threaded performance than the regular i7 4790K...

 

How is $300 to $400 a discount? That's right around the price of the low to mid-ranged Haswell-E CPUs... (Well the 5930K is like $564 but still...)

[That Norwegian Guy]

1 hour ago, Mr.Meerkat said:

Well the 8350 is sub 150 so you've already got it

8350 is not a true 8-core, it only has 4 floating point units so acts as a 4-core in any floating point calculation.

 

It also has 60% the IPC of Sandy Bridge and an astronomical TDP with little OC headroom that will scorch the PCB and VRMs of 94% of available motherboards if you want to OC it hard.

[Mr.Meerkat]

1 minute ago, That Norwegian Guy said:

8350 is not a true 8-core, it only has 4 floating point units so acts as a 4-core in any floating point calculation.

 

It also has 60% the IPC of Sandy Bridge.

mehhhhh  I know

[Trixanity]

9 minutes ago, wcreek said:

I know what IPC means, iirc Haswell-E was a little slower or the same single threaded performance than the regular i7 4790K...

 

How is $300 to $400 a discount? That's right around the price of the low to mid-ranged Haswell-E CPUs... (Well the 5930K is like $564 but still...)

Then you also know the reason why the 4790k is faster, right? Much higher clock speed. Pure and simple. 

 

$564 for a 6 core and you don't think even $400 is a discount for 8?

[That Norwegian Guy]

5 minutes ago, ONOTech said:

Or better yet, mITX. That was a huge flaw with Bulldozer's design. High TDP and bad efficiency meant weak VRMs for SFF.

Well the die size and heatsink mounting bracket was in itself the most major hindrance for m-ITX. Looking at these shots, the die is shaped like an Opteron die, (oblong) which could repeat those issues depending on size (since there's no size noted)

 

Hopefully they change the heatsink mounting bracket, as it takes up too much space both on AM3 and FM2.

[DocSwag]

Please stop taking that Ivy Bridge IPC claim for fact peoples..... It's only a rumor and not even confirmed.... We don't have any good guess as to Zen's IPC right now except it's probably between Ivy Bridge and a little above Skylake.

2 hours ago, TheRandomness said:

Sub £200 octocore please. 

You can dream, but I'm personally doubting that. Profits margins would just be too low. They wouldn't do that unless Zen's IPC is purty horrible like Bulldozer, in which case there's no point in buying it anyways....

I'd be expecting something more in the £500 range.

[handymanshandle]

1 minute ago, Trixanity said:

Then you also know the reason why the 4790k is faster, right? Much higher clock speed. Pure and simple. 

 

$564 for a 6 core and you don't think even $400 is a discount for 8?

Not really, at least if it's suppose to be more in line with performance of the mainstream i5s or i7s. Though if it's suppose to compete with the Intel CPUs then sure, I suppose $500 to $600 would give an Intel 8 core a good run for its money if the performance was there to match.

[Eroda]

well its simple AMD stated 40%+ increase over Excavator  Amd has fully fledged excavator APU's launching soon check those benchmarks and extrapolate from that

[Master Disaster]

1 minute ago, Eroda said:

well its simple AMD stated 40%+ increase over Excavator  Amd has fully fledged excavator APU's launching soon check those benchmarks and extrapolate from that

40% over Excavator is just about the most ambiguous term I've ever heard, which Excavator? Toronto or Carizzo? Which SKU exactly? I reckon I could prolly find a couple of benchmarks from a low end and high end SKU with a huge difference between them.

 

They might as well just say Zen will be 40% faster than the colour purple, that would be more useful.

[NumLock21]

51 minutes ago, That Norwegian Guy said:

8350 is not a true 8-core, it only has 4 floating point units so acts as a 4-core in any floating point calculation.

 

It also has 60% the IPC of Sandy Bridge and an astronomical TDP with little OC headroom that will scorch the PCB and VRMs of 94% of available motherboards if you want to OC it hard.

It has 8 cores, its 2 in a module, so 4 modules, thus 8 cores.

Zen with ivb ipc, is not exciting, no matter how cheap their cpu will be.

[tbake0155]

1 hour ago, That Norwegian Guy said:

8350 is not a true 8-core, it only has 4 floating point units so acts as a 4-core in any floating point calculation.

 

It also has 60% the IPC of Sandy Bridge and an astronomical TDP with little OC headroom that will scorch the PCB and VRMs of 94% of available motherboards if you want to OC it hard.

No, that's not exactly correct!  Each core has 128-bits of FPU bandwidth, which means unless they need all 256-bits of bandwidth, there is no FPU limitation.  And only specific AVX instructions need that much FPU bandwidth.

 

And a lot of older CPU's can't even perform those instructions, at all!

 

@patrickjp93  Maybe you can clarify this a bit for all us!! Might be that I'm wrong.

 

[Trixanity]

1 hour ago, wcreek said:

Not really, at least if it's suppose to be more in line with performance of the mainstream i5s or i7s. Though if it's suppose to compete with the Intel CPUs then sure, I suppose $500 to $600 would give an Intel 8 core a good run for its money if the performance was there to match.

Well, here it is again. The IPC of either is the same. It's the other bits that are different (core count, cache sizes, platform etc). And Summit Ridge is aimed squarely at the enthusiast segment meaning X99. So the IPC increase is not relevant as to where it'll go in the market as it'll be the same whether it's 2 cores or 8. Since it's a product that will go from 4 to 8 cores with accompanying SMT, it's in X99 territory. I at least think AMD will have 4 core Summit Ridge products so that leaves it as an odd man out compared to Intel but those will probably be reasonably cheap and will go up against a mainstream core i7 such as 6700K. Despite not being the same kind of product. Raven Ridge will come later and will have 4 cores (plus SMT) and Polaris integrated graphics and will be a more direct competitor to the 6700K and 6600K for that matter (well and whatever Kaby Lake introduces)

1 hour ago, Master Disaster said:

40% over Excavator is just about the most ambiguous term I've ever heard, which Excavator? Toronto or Carizzo? Which SKU exactly? I reckon I could prolly find a couple of benchmarks from a low end and high end SKU with a huge difference between them.

 

They might as well just say Zen will be 40% faster than the colour purple, that would be more useful.

Excavator is Excavator is Excavator. 40% is 40%. The IPC will go up regardless of SKU. What you're asking is what the actual performance is as in how will it compare to a particular SKU and that's of course impossible to tell as there's a lot of factors to consider such as clock speed, cache performance and size, core count, TDP etc. but in either case it means it will be at least 40% faster clock for clock and core for core at the same TDP although that's a very layman's way of putting it I'd say. The 40% is basically architectural improvements alone and not related to the process node and/or changes in clock speeds achievable. 

[handymanshandle]

6 minutes ago, Trixanity said:

Well, here it is again. The IPC of either is the same. It's the other bits that are different (core count, cache sizes, platform etc). And Summit Ridge is aimed squarely at the enthusiast segment meaning X99. So the IPC increase is not relevant as to where it'll go in the market as it'll be the same whether it's 2 cores or 8. Since it's a product that will go from 4 to 8 cores with accompanying SMT, it's in X99 territory. I at least think AMD will have 4 core Summit Ridge products so that leaves it as an odd man out compared to Intel but those will probably be reasonably cheap and will go up against a mainstream core i7 such as 6700K. Despite not being the same kind of product. Raven Ridge will come later and will have 4 cores (plus SMT) and Polaris integrated graphics and will be a more direct competitor to the 6700K and 6600K for that matter (well and whatever Kaby Lake introduces)

Then what do you think the Zen FX CPUs will be priced at.

[Shahnewaz]

4 hours ago, Lawliet93 said:

It should be around IVY BRIDGE level of performance (according to a Czech website called pctuning.cz )

You sure? Is that source reliable?

As far as I've heard, Zen is aiming to be on par with Haswell IPC.

[handymanshandle]

1 minute ago, valdyrgramr said:

Depends, imo.  I feel like they will range from 150-320ish depending on the power of each.  The Opterons will probably be around 600-5 grand.  I doubt that 32 core is gunna be cheap.

agreed, I was thinking an 8 core, 8 thread Zen FX CPU would be around 300 to 400 maybe 500 but @Trixanity must think that they'll be more if not a lot more than that...

[handymanshandle]

2 minutes ago, wcreek said:

agreed, I was thinking an 8 core, 8 thread Zen FX CPU would be around 300 to 400 maybe 500 but @Trixanity must think that they'll be more if not a lot more than that...

I just couldn't see AMD bringing in a CPU close to the $1000 mark that the 5960X is at unless it was like hugely faster than the 5960X in single and multithreaded workloads with a far higher base clock and a good OC margin. 

$600 or so might be the most AMD might charge for an 8 core 8 thread CPU but that seems way too high already.

[Trixanity]

1 minute ago, wcreek said:

Then what do you think the Zen FX CPUs will be priced at.

Tricky question to answer. It'll be priced based on the segment first and then the actual performance later. Similar to how graphics cards are priced as well. Since we're talking enthusiast grade segment (according to AMD, who have also stated they don't want to be the budget choice but a legitimate competitor in performance) that means prices close to Intel if not downright the same and then adjust from there to get the sales going. They're not pricing it any lower than they have to, that's for sure.

 

8 core is a premium product right now. So I'm not imagining it less than $800 but I think AMD would try to sell it for $999 if they could (for the absolute top SKU that is). Depends on what their market research tells them.

As for 6 cores? Probably $350 at least for the lowest SKU. I'm assuming more like $400-450 but with full PCIE-lanes to give Intel's lowest 6 core SKU the middle finger by promising better value for your money.

 

The problem is that we also don't know what and how many SKUs there are. Will they all have SMT enabled? Or will they have SKUs with it disabled on some like Intel? How many 8 core SKUs will there be? How many 6 core SKUs? How many 4 core SKUs? 

The more SKUs you have with X number of cores, the less value that segment will have, so if they flood the market with good 6 cores for example, some will have to be cheaper than the others, so you'll end up with prices dropping for the bottom SKUs, so it's a tricky business.

 

So it's all guessing games, just like performance numbers. But know this: AMD will try to price as close to Intel as they possibly be can but will have to adjust for performance discrepancies and also probably have to undercut a little bit to get both consumers and OEMs to bite. They will not, absolutely not, hand out 8 cores for $250 unless something has gone completely wrong, and just like with Bulldozer, it'll only drop to such a level after launch when performance numbers and sales numbers start to fail. 

 

So tl;dr: Expect within range of similar Intel offerings and/or the same prices but with some value added to compensate.

 

At best, expect up to a $100-150 discount on an AMD product with similar specs to Intel but even that is probably wishful thinking.

As someone else put it, if 15% slower, expect a similar % in discount. Seems fair but the sales & marketing at AMD will be trying to sell you the product for top dollar if they can.

[handymanshandle]

2 minutes ago, Trixanity said:

Tricky question to answer. It'll be priced based on the segment first and then the actual performance later. Similar to how graphics cards are priced as well. Since we're talking enthusiast grade segment (according to AMD, who have also stated they don't want to be the budget choice but a legitimate competitor in performance) that means prices close to Intel if not downright the same and then adjust from there to get the sales going. They're not pricing it any lower than they have to, that's for sure.

 

8 core is a premium product right now. So I'm not imagining it less than $800 but I think AMD would try to sell it for $999 if they could (for the absolute top SKU that is). Depends on what their market research tells them.

As for 6 cores? Probably $350 at least for the lowest SKU. I'm assuming more like $400-450 but with full PCIE-lanes to give Intel's lowest 6 core SKU the middle finger by promising better value for your money.

 

The problem is that we also don't know what and how many SKUs there are. Will they all have SMT enabled? Or will they have SKUs with it disabled on some like Intel? How many 8 core SKUs will there be? How many 6 core SKUs? How many 4 core SKUs? 

The more SKUs you have with X number of cores, the less value that segment will have, so if they flood the market with good 6 cores for example, some will have to be cheaper than the others, so you'll end up with prices dropping for the bottom SKUs, so it's a tricky business.

 

So it's all guessing games, just like performance numbers. But know this: AMD will try to price as close to Intel as they possibly be can but will have to adjust for performance discrepancies and also probably have to undercut a little bit to get both consumers and OEMs to bite. They will not, absolutely not, hand out 8 cores for $250 unless something has gone completely wrong, and just like with Bulldozer, it'll only drop to such a level after launch when performance numbers and sales numbers start to fail. 

 

So tl;dr: Expect within range of similar Intel offerings and/or the same prices but with some value added to compensate.

 

At best, expect up to a $100-150 discount on an AMD product with similar specs to Intel but even that is probably wishful thinking.

As someone else put it, if 15% slower, expect a similar % in discount. Seems fair but the sales & marketing at AMD will be trying to sell you the product for top dollar if they can.

Well I wasn't saying that it would be $250, I think for an 8 core 8 thread that doesn't perform with in small margins ahead or below the 8 core 16 thread from Intel then I don't see possibly why AMD should charge $1000 for their 8 core, 8 thread units...

 

[Lawliet93]

3 minutes ago, wcreek said:

Well I wasn't saying that it would be $250, I think for an 8 core 8 thread that doesn't perform with in small margins ahead or below the 8 core 16 thread from Intel then I don't see possibly why AMD should charge $1000 for their 8 core, 8 thread units...

 

Well, there might not be any 8c/8t CPUs from AMD, because they are going to use SMT to effectively double the thread count (like HT does for intel), so 8c/16t it is and they might even go the intel route with 4c/4t and 4c/8t CPUs (note, that HT or this level of SMT only gives 30% more performance in multithreaded applications).

[Trixanity]

4 minutes ago, wcreek said:

Well I wasn't saying that it would be $250, I think for an 8 core 8 thread that doesn't perform with in small margins ahead or below the 8 core 16 thread from Intel then I don't see possibly why AMD should charge $1000 for their 8 core, 8 thread units...

 

I said top SKU. The top SKU will have 8 cores/16 threads. It is two-way SMT after all - just like Intel. The hyper threading moniker is just a marketing term. But again, they will probably have to lower it to compensate for performance discrepancies. So let's just say $899 or $799 if we're lucky. If it was $499 for such a SKU, I'll assume something has gone horribly wrong.

[Lawliet93]

1 minute ago, Trixanity said:

I said top SKU. The top SKU will have 8 cores/16 threads. It is two-way SMT after all - just like Intel. The hyper threading moniker is just a marketing term. But again, they will probably have to lower it to compensate for performance discrepancies. So let's just say $899 or $799 if we're lucky. If it was $499 for such a SKU, I'll assume something has gone horribly wrong.

Well the price difference will reflect the performance difference, like it has been in the past. (they might drop the price a little bit more just to be that much more appealing to the customers)

[Trixanity]

2 minutes ago, Lawliet93 said:

Well the price difference will reflect the performance difference, like it has been in the past. (they might drop the price a little bit more just to be that much more appealing to the customers)

Yup, that's exactly my point. Performance is a huge factor in pricing but when we're delving into enthusiast or enterprise segments, there will be diminishing returns in performance gains per dollar hence you see huge price jumps from SKU to SKU in Intel's enthusiast line-up, even though the price does not reflect the gains. But that's also a reflection of supply and demand. As I mentioned previously, there are very few SKUs and they are in a segment and price range where the units shipped are much smaller, so it drives prices up.

[patrickjp93]

1 hour ago, tbake0155 said:

No, that's not exactly correct!  Each core has 128-bits of FPU bandwidth, which means unless they need all 256-bits of bandwidth, there is no FPU limitation.  And only specific AVX instructions need that much FPU bandwidth.

 

And a lot of older CPU's can't even perform those instructions, at all!

 

@patrickjp93  Maybe you can clarify this a bit for all us!! Might be that I'm wrong.

 

You're correct, but the fighting over the FPU still requires the cores check with each other to see which half of it is in use by the other, hence the horrible FP performance. In terms of integer instructions, for Piledriver and beyond, they are true cores. For Bulldozer, the two halves would share a scheduler too, making them far less than true cores.

 

43 minutes ago, Shahnewaz said:

You sure? Is that source reliable?

As far as I've heard, Zen is aiming to be on par with Haswell IPC.

I ran some simulations a long time ago using QEMU and an integer linear program where the constraints were the total IPC increase (40% at the time), integrality of clock cycle latencies, and a few theoretical barriers such as no instruction can take 0 cycles and some things like integer division require more cycles for more bits and there is a lower bound. In the best case scenario (used 40 different benchmarks including the Linpack Suite), Zen would end up with 3% higher IPC than Haswell. Accounting for reality, it'll be closer to Ivy Bridge.