AMD Unveils Zen Microarchitecture, Demos Summit Ridge Performance

[elbert]

Zen+ i already posted its core count and the motherboard is backward compatible to the 8 and 4 cores talked about here in the forum's zen thread. The core count on what I expect to be AM4+ is 12 and 16 cores or basically what is used in a 2 chip melt for Naples.

 

[SylentVyper]

Intel last 3 Gen of CPU have no impact on games and have barely seen any improvement in calculation power. It's kind of sad. If AMD can provide similar performance for less, guess what, they will keep their leadership in the console business. That's about 225 millions consoles every 5 years.

https://www.youtube.com/watch?v=VDo-j00vUtw

No impact? Ok.

 

[Dags]

That broadwell-E CPU is way out of my budget. That's a very expensive chip and way beyond mainstream. So If AMD can bring performance close to that super expensive intel CPU-Mobo combo (remind you that 2011 mobos are more expensive that 4c i7 CPUs) for considerably less money, then I am sold!

 

[elbert]

If AMD delivers Broadwell performance or more for fraction of what Intel counterpart cost, they got my money.

If Zen manages to match Broadwell, it will likely get priced half-way to the next step below. Ex.: if the top-end Zen matches the i7-6900k, it'll get priced between the 6850k and 6900k, which will still set you back more than $700.

System wise tho cheaper then even the 6850K. Zen moves alot more motherboard parts into the core thus sub 100 motherboards is to be expected. This compared to $200 2011v3.

 

[ComputerSecurityGuy]

Yeah, board cost is what keeps most from getting the i7-5820K over the i7-6700K.

And Kaby Lake, I believe, is pretty much an invention of Intel's marketing department coupled with the new chipset release like Haswell. Kaby Lake, I am pretty sure, is really Skylake Refresh.

 

[blppt]

Sigh....if only this was released in 2016 like it was originally going to. I still believe that by the time Zen actually launches, AMD will have given Intel extra time to counter with their next gen, and that is not good for an AMD CPU resurgence.

 

[littleleo]

Zen doesn't have to beat Intel, just be close at a better cost many will give them a chance. I'm just not sure when they will ever be ready to launch Zen. The date keeps getting pushed back worse then the Brown's offensive line.

You know before AMD bought ATI they were in talks with Nvidia, but the CEO of Nvidia wanted to run the new company and the AMD CEO wanted control so they went after ATI instead. If there had been an Nvidia AMD merger I wonder what the GPU and CPUs we would be looking at today. Nvidia with an AMD CPU division would have defiantly made Intel take notice.

 

[ComputerSecurityGuy]

Not as much as you think. They are probably finalizing Cannonlake now. Kaby Lake is a marketing creation. And Skylake-E will only be marginally faster than Broadwell-E if Broadwell-Skylake IPC growth repeats itself. So I'm not too concerned about Intel's new lines immediately in terms of Zen competition.

 

[TJ Hooker]

@Xenol

I sort of agree with him though. They're saying a 40% increase over Excavator, which has been around for 5 years. Intel doing 10% a year for 5 years is 50%, so until actual reviews and benchmarks come out - at least from a numbers standpoint - they're no closer to Intel than they were 5 years ago.

You're thinking of Bulldozer, which came out in 2011. Excavator is the last of a series of revisions to the Bulldozer architecture, and came out in 2015. I don't know the numbers on how much better Excavator is than Bulldozer in terms of IPC, but it's not accurate to say that Zen is just 40% better than a 5 year old architecture.

 

[InvalidError]

They are probably finalizing Cannonlake now. Kaby Lake is a marketing creation.

Kaby Lake is an in-between bench-warmer just like Haswell Refresh because Intel expects to need more time to get 10nm up to speed as it did with Broadwell. If there are any execution mishaps, Cannonlake may very well slip into 2018.

 

[joex444]

So when they underclock the i7-6900K to 3.0GHz, an 8C/16T part, and put it head-to-head with a 3.0GHz 8C/16T Zen they get about the same performance. That's great that they have, effectively, the same IPC as Broadwell-E. The i7-6900K is a $1000 CPU; if Zen comes out significantly under that then it's going to spur a small price war. If Zen is priced close to Broadwell-E then we'll probably see AMD making money again. Either one is good for everyone except Intel. Of course Zen and AM4 would need to come in with similar technology as X99, namely, 40 PCIe 3.0 lanes, quad channel memory (or bandwidth > 50GB/s without insanely fast DDR4), and PCIe 3.0 x4 NVMe M.2 slots/U.2 ports. With the lack of SATA Express devices and the meaningless improvement to 10Gbps with NVMe being 32Gbps AMD should've skipped implementing SATA Express entirely.

Downside is we already know Broadwell-E and the i7-6900K can overclock from 3.2GHz to 4.3GHz from the THG review, and 4.0GHz is apparently trivial. That's a 33% improvement AMD needs to make from "early silicon" to production. People aren't paying $1000 for a 3.2GHz 8C, they're paying $1000 for a 4.0GHz+ 8C/16T.

 

[RedJaron]

I sort of agree with him though. They're saying a 40% increase over Excavator, which has been around for 5 years. Intel doing 10% a year for 5 years is 50%, so until actual reviews and benchmarks come out - at least from a numbers standpoint - they're no closer to Intel than they were 5 years ago.

Wrong. Excavator is not five years old. You're confusing that with Bulldozer. The first desktop Excavator chip was the Athlon 845, released earlier this year. You're forgetting the improvements that came with Piledriver and Steamroller.

Your math is all kinds of screwed up. First, you're making a fence-post error; you're counting an extra data point after the end of the set. Actually, you're doing it twice since you're counting from before your set as well ( Sandy Bridge should be your base 0, not the first 10% increase in your methodology ). You actually only have three increases thus far in the last five years: Sandy Bridge to Ivy Bridge, IB to Haswell, HW to Skylake. Kaby Lake hasn't launched yet. It might be fair to compare Zen to KL since they'll hopefully launch around the same time, but we don't have any KL numbers yet.

Second, Intel hasn't been increasing 10% in each generation. It's actually between to 5% - 8%. So even assuming the higher 8% figure, SL is about 26% faster than SB, or half of your assumed 50%. Even using the 10% figure, it's only 33%. Conversely, Zen's IPC is 40% better than EX, which was ~5% better SR, which was ~6% better than PD, which was ~6% better than BD. Which means they're claiming Zen's IPC is 66% higher than BD's

Under-clocking the Intel chip to boost stats on "comparable" systems and Zen was not even noticeably faster? That's sad.

i stopped reading at the part where they had to underclock an intel I7 to match the speed of the test chip. top notch marketing as always by AMD

You both should have actually read the article and details before jumping on the comments to rant. The i7 was underclocked to from 3.2 GHz to 3.0 GHz so that it was running at the same speed as the demo Zen CPU. That way you get a true clock-for-clock comparison. If this demo is to be believed, Zen has similar IPC to Broadwell-E. That's a huge jump over their previous architectures.

 

[elbert]

So when they underclock the i7-6900K to 3.0GHz, an 8C/16T part, and put it head-to-head with a 3.0GHz 8C/16T Zen they get about the same performance. That's great that they have, effectively, the same IPC as Broadwell-E. The i7-6900K is a $1000 CPU; if Zen comes out significantly under that then it's going to spur a small price war. If Zen is priced close to Broadwell-E then we'll probably see AMD making money again. Either one is good for everyone except Intel. Of course Zen and AM4 would need to come in with similar technology as X99, namely, 40 PCIe 3.0 lanes, quad channel memory (or bandwidth > 50GB/s without insanely fast DDR4), and PCIe 3.0 x4 NVMe M.2 slots/U.2 ports. With the lack of SATA Express devices and the meaningless improvement to 10Gbps with NVMe being 32Gbps AMD should've skipped implementing SATA Express entirely.

Downside is we already know Broadwell-E and the i7-6900K can overclock from 3.2GHz to 4.3GHz from the THG review, and 4.0GHz is apparently trivial. That's a 33% improvement AMD needs to make from "early silicon" to production. People aren't paying $1000 for a 3.2GHz 8C, they're paying $1000 for a 4.0GHz+ 8C/16T.

I haven't seen the specs for AM4 but leaks say that Naple has 128(64) PCI lanes. I put (64) because half will be used for 2 and 4 way connections. The math would suggest the AM4 has atleast 32 PCI lanes and next years enthusist AM4+ 64 PCI lanes. The Naple server leaks suggest 8 channel DDR4. The AM4 has dual channel DDR4 memory thus one can conclude next years AM4+ enthusists board will have quad channel.
This said i'm not sure you can compare the mobo's raw specs as Zen takes up much more on the AM4 tasks.

 

[shrapnel_indie]

This looks promising. Further testing will be required to know if the results hold true under other workloads and in production chips. Underclocking, and matching clock speeds is perfect for IPC comparisons. These were said to be pre-production chips and that they'd be shipping faster units later, so the underclocking comparison doesn't concern me the least bit.

 

[RedJaron]

Downside is we already know Broadwell-E and the i7-6900K can overclock from 3.2GHz to 4.3GHz from the THG review, and 4.0GHz is apparently trivial.

This is correct, but it's important to understand the context. Yes, 4.0 GHz on a 6900K and 6950X is somewhat trivial since you can do it on a number of affordable big air coolers. However, hitting 4.2 requires a massive air cooler or sizable liquid loop ( my 6950X on a NH-D15 is stable for the most part, but the CPU is constantly ~95 °C ). Reaching 4.3 needs a huge liquid loop, and even then the CPU is still in the 90s. The problem is that the silicon simply can't transfer heat fast enough to the heat spreader.

 

[RedJaron]

I haven't seen the specs for AM4 but leaks say that Naple has 128(64) PCI lanes. I put (64) because half will be used for 2 and 4 way connections. The math would suggest the AM4 has atleast 32 PCI lanes and next years enthusist AM4+ 64 PCI lanes. The Naple server leaks suggest 8 channel DDR4. The AM4 has dual channel DDR4 memory thus one can conclude next years AM4+ enthusists board will have quad channel.
This said i'm not sure you can compare the mobo's raw specs as Zen takes up much more on the AM4 tasks.

Is AMD separating its server-based boards from the consumer boards, a la X99 with its 40 lanes vs Z170 with only 20?

 

[littleleo]

It would make sense to do so since server boards are more expensive to make. AMD has always tried to have the price advantage. A board with dual desktop and server use would cost more than one just for consumer desktops. I believe it was said the AMD Server ZEN models are targeted for Q2 of 2017, but of course the ZEN was supposed to come out this year in October.

It will be interesting to see how these ZEN CPUs will OC. AMD was more generous with unlocked CPUs in the past then Intel.

 

[blppt]

Not as much as you think. They are probably finalizing Cannonlake now. Kaby Lake is a marketing creation. And Skylake-E will only be marginally faster than Broadwell-E if Broadwell-Skylake IPC growth repeats itself. So I'm not too concerned about Intel's new lines immediately in terms of Zen competition.

Thats only if you assume that Intel isnt holding anything back with these refreshes---lets be honest, since Sandy Bridge, there hasnt been a whole lot of reason for them to push the technology with AMD's CPU advancements as the only real competition.

 

[littleleo]

You actually only have three increases thus far in the last five years: Sandy Bridge to Ivy Bridge, IB to Haswell, HW to Skylake. Kaby Lake hasn't launched yet. .

Actually it was Sandy Bridge to Sandy Bridge V.2 to Ivy Bridge to Haswell to Haswell refresh to Broadwell to Skylake. Though broadwell was delayed until 2 weeks before Skylake drop sealing its fate as a DOA line some still bought it and some notebooks used it as well. If you figure there was some improvement perhaps as low as 2% improvement you still have 6 increase in performance after the introduction of the Sandy Bridge iCore CPUs.

 

[elbert]

I haven't seen the specs for AM4 but leaks say that Naple has 128(64) PCI lanes. I put (64) because half will be used for 2 and 4 way connections. The math would suggest the AM4 has atleast 32 PCI lanes and next years enthusist AM4+ 64 PCI lanes. The Naple server leaks suggest 8 channel DDR4. The AM4 has dual channel DDR4 memory thus one can conclude next years AM4+ enthusists board will have quad channel.
This said i'm not sure you can compare the mobo's raw specs as Zen takes up much more on the AM4 tasks.

Is AMD separating its server-based boards from the consumer boards, a la X99 with its 40 lanes vs Z170 with only 20?

Leaks suggest a thrid socket in 2h 2017 for 12 and 16 cores plus drop in compatible with the 4 and 8 cores.

 

[InvalidError]

AMD has always tried to have the price advantage.

AMD is not trying to take the "price advantage", it is forced into the value position due to lack of performance. Back when AMD had the performance lead while Intel was desperately trying to make Netburst work, AMD had $1000 desktop chips of their own.

 

[filippi]

Now give me 4 unlocked SMT Zen Cores w/ 90% i7-6700k@4GHz performance and Haswell i3<Zen<i5 price range
Go AMD. You can do it again!

 

[bit_user]

The company added a micro-op cache, which helps with better instruction predictions

I'm honestly surprised their previous CPUs didn't have this. I see the main benefit as alleviating the burden on the decoder, and possibly removing it as a bottleneck. Should also save power.

Everything detailed here is pretty boring. It's all stuff Intel & others have been doing for years. Not that there's anything wrong with that, but it would've been nice if they'd come up with something new and innovative.

But the main thing I care about is to see AMD back in the game. The benchmarks definitely sound promising, and I'm pretty impressed by the 32-core server CPUs.

 

[InvalidError]

Now give me 4 unlocked SMT Zen Cores w/ 90% i7-6700k@4GHz performance and Haswell i3<Zen<i5 price range
Go AMD. You can do it again!

They cannot afford to do that with billions of dollars worth of debt due for renewal over the next three years. If they have a chip good enough to match Intel, expect it to almost match Intel in price as well. If Zen 4C8T delivers i7-6700 level performance with no major flaws (such as significantly worse performance per watt), it'll get priced between the i5-6600 and i7-6700.

 

[Virtual_Singularity]

Some time ago, Lisa Su said there'd be "no doubt AMD was back", by the time CES '17 took place. There was, IIRC, something about near-term goals stated/printed around the same time. I'm unable to find the page in question, but as I recall, these goals were... 1. offer strong competition in the dGPU market. And 2. Maintain a strong second to Intel by improving the value, efficiency, performance of their latest CPU's/APU's.

While the article above is appreciated, I'd rather not speculate on it. If AMD intends to "be back" by CES 2017, both Zen _and_ Vega launches must be genuinely successful, and offer real competition to Intel and NV. If they're to remain relevant vs Intel and NV in 2017 and onward, AMD really needs to step up their game and lose the "always 'on the way', too little, too late", stigma that's formed around the company, and frustrated many of their potential customers. I hope they succeed.