Most test results shown by Intel are well bellow the 200% to 400% performance improvement demonstrated by #AMD EPYC2 .
Add to that that Intel's CPUs will use a huge load of power and will also need new infrastructure with motherboards and everything and you get a disastrous 2019 year for Intel.
Not "disastrous" because AMD will take 90% of the server market , not at all, but because that 5% "red line" imposed by Intel's former CEO on the server market will likely be pushed back towards 15% which is a truly disastrous result comparing with the initial goal of the company.
What makes it even worse for Intel is that AMD will also have the revenues & profit to keep investing and developing new tech which will put it in a completely different position than the cash-strapped scenario it has been into, during the past 10 years.
You forget what is Rome. And half 7nm SKU, so with halved power savings over 14nm. In this condition Cascade AP on 14nm++ is well enough. Rome was a delusion under many espects.
You picked the wrong comment to reply to given that IGTrading never mentioned power, but I'll take that bait.
Realistically (as in, without reeking of bias) we can assume that AMD are targeting increased performance at roughly the same ISO power consumption for their new chips, because that's how server designs work. Even if your baseless half-and-half claim is correct (and that's a big if) AMD's accordingly modest performance gains for Rome would still make them competitive with CLAP, and with a less-expensive and more-mature platform to boot. In reality there's no reason to doubt AMD's claims of 2-4x performance boost depending on the application (twice as many cores, twice the FP width, similar power draw).
Put simply, your stupid FUD comment is stupid FUD.
You seem to miss understand the issues with design limits. The I/O interfaces of all processors regardless of manufacturer have the same issue. Regardless of their design size (3nm,5,10,14,22...) they have to accept, switch and deliver the current and voltage required by the interface they support (DRAM, PCI, USB, SATA....). So even once Intel has its 10nm process fully operational they will not gain much advantage over AMD even if their interface logic is also 10nm in size.
True. The difference is infinity fabric was a seriously thought out long term technology designed to be used in both cpus and gpus. It has its own set of disadvantages so multi die microprocessors have to be designed correctly from start to finish.
Intel established a trend of doing whatever they want for 8 years and got caught with their pants down in so many ways. This is a rushed response by a company freaked out over losing market share, 10nm could still be a huge problem for them (they would never admit it until they have to). Not really the right conditions to launch a technology like this.
I think you are forgetting that by time EPYC2 is out - Intel will likely have it based on Ice Lake and will significantly improved it even based on this process. It just so how interesting when AMD talks vaporware people cheer - but Intel talks they are doing and people don't believe it.
This is probably why companies do not trust AMD. Big difference between Intel and AMD is Intel is scalable - meaning that it not just 2P systems, but also 4P and 8P and even more. So core count is not really that important when you can just add more cpu's on same system.
One thing I am curious about if a system have 4 16 core cpus each with their own dedicated memory or single 64 core cpu - with memory that is shared - which is actually better especially when taking to larger systems with say 128 plus CPU and one of them fails. For me scalable sounds more reliable.
Please remember that AMDs numbers are also FUD, as those are also a vendor supplied numbers and just 1 test. But yes, that test did give a glimpse of what is coming as this Intel system, in reality, is just nerfed 28core 4 socket system.
Also, i am worried of lack of glue in this article. So, let the GLUE wars begin. Who has better GLUE - AMD or INTEL or AMD?
Intel's 'benchmarks' are worthless without power consumption figures to go along with them. Which I'm sure Intel would be happy (NOT!) to supply. And I'm sure they continue to use their whole toolbox of shenanigans. Maybe anandtech should formally ask Intel for this information and make it crystal clear to readers that Intel refuses to provide any of it.
Being fair, AMD didn't disclose exact Zen 2 power figures either, but AMD did say 'half the power at the same clock', on a per-core basis. That is fairly definitive. That means the 64-core AMD Zen 2 EPYC is going to consume roughly the same power as their 32-core part does right now, for twice the performance. At a minimum. AMD is still holding back really describing how the doubled FP performance per core (yielding 4x the performance when cores are also doubled) fits into the power consumption specs.
You're right on the money. Power consumption and efficiency are what really matter for Data Centers. If they have the same crew runn those benchmarks as they did at CES its probably running at 20kwatts with a 20kwatt phase change cooler and a hacked motherboard. AMD at least has a process shrink to help drive efficiency a bit how much remains to be seen but their claims of 25% aren't off the wall.
Don't expect half power consumption, considering that the I/O chip is still at 14nm. But it will be that much lower that it will make any Intel AMD power consumption comparisons laughable at best. Think what is happening with Nvidia/AMD GPUs but this time with AMD having the clear advantage.
Don t be so confident, Xeon AP will be low clocked like Rome, in these conditions 14nm++ is pretty in line with 7nm (early) power savings. Rome will be hot unfortunately. Intel is right now selling Xeon AP because customers have the exact power figures of Rome (AMD always share the data six/eight months before shipment to have an idea of the volume to order to foundry)
In other words, “hey, please delay your next order until we can get out act together!”
EPYC 2 sounds like it will have more IPC and better clocks, not to mention memory bandwidth gains. We won’t be asking how Intel’s next CPU compares to AMDs old CPU.
So lets put these benchmarks into context. They compare a 48 core CPU with a 64 core system made from 2 CPUs each 32 cores each. Fair enough, they do manage to get better results with lesser cores, but I am pretty confident that these are scenarios where inter-CPU communications matter and we all know that the latency is high on a dual socket system. Second, AMD has their new CPU line-up just around the corner and they double pretty much everything. They also quadruple FP resources. If things go well, then it could very be that one Rome CPU could keep up with 2S Cascale Lake AP system, which would be disastrous for Intel.
If I understand correctly, the figures refer to a system with two sockets each occupied by a 48 Core Cascade Lake AP. That means that they get 1.5-1.6x the performance with 1.5x the number of cores of Epyc 7601 in most scenarios and maintain a performance lead on some hyper-optimized AVX-workloads. And this is comparing an unreleased system to AMD's line of processors, which have been available for a year or so. No power figures are provided and only few benchmarks are shown. I think, it is safe to say that Zen 2 will be much faster based on the number of cores alone. That is not even considering the expected improvements in IPC, clocks, and power efficiency.
Intel disabled SMT on the Epyc chip in their "benchmarks". You should keep that mind. They probably did this because their own chip won't have hyper-threading, anything you can do to show your own processor in the best light.
The test on how scared intel is will be what they do with prices. If they drop prices where they've planned to raise them 20% we'll be seeing AMD taking more of the cloud operators.
Well, on the footnotes is showing 2x48c against a 2x32c, so its 96c against 64c. Seems that everyone is thinking that one CPU of 48c is beating a 64c. This turn the things worst for Intel.
They also didn't say whether SMT was enabled on the Intel system or not. On several of the benchmarks SMT was disabled on the Epyc and then enabled on other benchmarks, however, no mention of what it is for Intel.
So, AMD and Intel are pushing parallel FP benchmarks.
Someone tell them that CPUs already lost this battle (despite AVX-512 for Intel) as HPC users almost always use accelerators for these use cases like CFD and weather forecast. Just look at the TOP500...
There are still some cases where splitting the data out to separate accelerators takes more time (either in programming effort or just in terms of waiting around) than just performing the calculations on CPU, but you're right that the number of scenarios these performance comparisons apply to is shrinking rapidly.
This does not bode well for Intel. 2x48 only manages to beat out 2x32 by 1.5x to 1.6x for non-AVX heavy workloads and for AVX-heavy workloads, the lead jumps to 3.4x. The fact that Intel is pitting 50% more cores against AMD in this comparison already accounts for that 1.5x to 1.6x performance uplift... With EPYC 2, AMD touts 2x the perf/W and 2x the cores AND with AT's interview with Mark Papermaster, AMD will support AVX2 without backing off on clocks. In a 2S head to head, Intel's 2x48 vs AMD's 2x64 will show AMD having the performance lead across the board at a power consumption level similar to a 2S EPYC 1 system.
The interesting slides are not the performance numbers, but the last two: vendors designing systems around the chips. Checking the latest TOP500, there are not yet AMD design wins listed (for Epyc, much less Rome), and none that I am aware of under construction or design (though that could have changed).
@edzieba..."Checking the latest TOP500, there are not yet AMD design wins listed (for Epyc..."
#38 on the newest TOP500 list contains 5120 - EPYC 7501 (32 core) processors...note the processors are listed as Hygon Dhyana Epyc 7501 processors, which comes from a JV between AMD and a company created by the Chinese government.
- Intel is having to commit what in many ways its future 4-way server design to compete with AMD's current 2-way server. This is not a great story for the majority of deployments that are 1 or 2 socket systems and 4 socket systems are meant to be a high-profit area for Intel.
- Intel disabled SMT on the AMD system for the first test and many of the other tests indicate that the number of threads running was set at 1 per core. So SMT was not used where possible.
- Intel used their C/C++ compiler to compile the test suites where possible. There is no indication that they used the AMD Optimized compiler for the AMD systems.
- The header text states that microprocessor-dependent optimizations and non-specific optimizations were only carried out for the Intel processors.
All in all the info seems to have come from the same marketing team that release complete spin for the i9-990K and before that the consumer version of the 28 chore Xeon Platinum 8180.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
52 Comments
Back to Article
The Hardcard - Sunday, November 11, 2018 - link
No “Cash Out Of Your Pocket” benchmark? What’s the multiplier versus EPYC on that?boeush - Sunday, November 11, 2018 - link
Translation: Intel is getting worried, like, for reals this time.Good.
Urthor - Sunday, November 11, 2018 - link
I mean they hired Jim Keller how long ago? They knew exactly what was comingFritzkier - Sunday, November 11, 2018 - link
7 months ago is not "long" if you are talking about development process...Lolimaster - Monday, November 12, 2018 - link
Jim Keller mumbo jumbo is 2022 at the earliest.Gondalf - Monday, November 12, 2018 - link
Too much, 2021 is more in line with a cpu developmentLolimaster - Monday, November 12, 2018 - link
1st they need a proper 10 Ice-Lake family then sub10nm Tiger Lake, then JK-lake.quadrivial - Monday, November 12, 2018 - link
If Jim Keller said ANYTHING about upcoming AMD products, he would be fiscally ruined and would never work anywhere again.Alexvrb - Monday, November 12, 2018 - link
He wouldn't need to say anything. He could just say "Look we're taking two of these and cramming together."*bangs two expensive chips into each other for emphasis*
"Doesn't that seem like... overkill?"
"No."
FredWebsters - Wednesday, December 26, 2018 - link
I think that the products that AMD do - are very useful products!FredWebsters - Wednesday, December 26, 2018 - link
I think that there are only cool products!IGTrading - Monday, November 12, 2018 - link
Most test results shown by Intel are well bellow the 200% to 400% performance improvement demonstrated by #AMD EPYC2 .Add to that that Intel's CPUs will use a huge load of power and will also need new infrastructure with motherboards and everything and you get a disastrous 2019 year for Intel.
Not "disastrous" because AMD will take 90% of the server market , not at all, but because that 5% "red line" imposed by Intel's former CEO on the server market will likely be pushed back towards 15% which is a truly disastrous result comparing with the initial goal of the company.
What makes it even worse for Intel is that AMD will also have the revenues & profit to keep investing and developing new tech which will put it in a completely different position than the cash-strapped scenario it has been into, during the past 10 years.
Gondalf - Monday, November 12, 2018 - link
You forget what is Rome. And half 7nm SKU, so with halved power savings over 14nm. In this condition Cascade AP on 14nm++ is well enough. Rome was a delusion under many espects.Meaker10 - Monday, November 12, 2018 - link
The I/O does not account for half the power, the cores consume the large part of it.Spunjji - Monday, November 12, 2018 - link
You picked the wrong comment to reply to given that IGTrading never mentioned power, but I'll take that bait.Realistically (as in, without reeking of bias) we can assume that AMD are targeting increased performance at roughly the same ISO power consumption for their new chips, because that's how server designs work. Even if your baseless half-and-half claim is correct (and that's a big if) AMD's accordingly modest performance gains for Rome would still make them competitive with CLAP, and with a less-expensive and more-mature platform to boot. In reality there's no reason to doubt AMD's claims of 2-4x performance boost depending on the application (twice as many cores, twice the FP width, similar power draw).
Put simply, your stupid FUD comment is stupid FUD.
RogerAndOut - Tuesday, November 13, 2018 - link
You seem to miss understand the issues with design limits. The I/O interfaces of all processors regardless of manufacturer have the same issue. Regardless of their design size (3nm,5,10,14,22...) they have to accept, switch and deliver the current and voltage required by the interface they support (DRAM, PCI, USB, SATA....). So even once Intel has its 10nm process fully operational they will not gain much advantage over AMD even if their interface logic is also 10nm in size.Opencg - Monday, November 12, 2018 - link
True. The difference is infinity fabric was a seriously thought out long term technology designed to be used in both cpus and gpus. It has its own set of disadvantages so multi die microprocessors have to be designed correctly from start to finish.Intel established a trend of doing whatever they want for 8 years and got caught with their pants down in so many ways. This is a rushed response by a company freaked out over losing market share, 10nm could still be a huge problem for them (they would never admit it until they have to). Not really the right conditions to launch a technology like this.
HStewart - Monday, November 12, 2018 - link
I think you are forgetting that by time EPYC2 is out - Intel will likely have it based on Ice Lake and will significantly improved it even based on this process. It just so how interesting when AMD talks vaporware people cheer - but Intel talks they are doing and people don't believe it.This is probably why companies do not trust AMD. Big difference between Intel and AMD is Intel is scalable - meaning that it not just 2P systems, but also 4P and 8P and even more. So core count is not really that important when you can just add more cpu's on same system.
HStewart - Monday, November 12, 2018 - link
One thing I am curious about if a system have 4 16 core cpus each with their own dedicated memory or single 64 core cpu - with memory that is shared - which is actually better especially when taking to larger systems with say 128 plus CPU and one of them fails. For me scalable sounds more reliable.Haawser - Sunday, November 11, 2018 - link
Seeing as AMD just showed a prototype EPYC2 beating two top of the line 28C Xeons, I'm not sure that two 24C ones are going to stand much of a chance.Karmena - Monday, November 12, 2018 - link
Please remember that AMDs numbers are also FUD, as those are also a vendor supplied numbers and just 1 test. But yes, that test did give a glimpse of what is coming as this Intel system, in reality, is just nerfed 28core 4 socket system.Also, i am worried of lack of glue in this article. So, let the GLUE wars begin. Who has better GLUE - AMD or INTEL or AMD?
danjw - Sunday, November 11, 2018 - link
Nothing but FUD, until outsiders actually get silicon to test!MattZN - Sunday, November 11, 2018 - link
Intel's 'benchmarks' are worthless without power consumption figures to go along with them.Which I'm sure Intel would be happy (NOT!) to supply. And I'm sure they continue to use their whole toolbox of shenanigans. Maybe anandtech should formally ask Intel for this information and make it crystal clear to readers that Intel refuses to provide any of it.
Being fair, AMD didn't disclose exact Zen 2 power figures either, but AMD did say 'half the power at the same clock', on a per-core basis. That is fairly definitive. That means the 64-core AMD Zen 2 EPYC is going to consume roughly the same power as their 32-core part does right now, for twice the performance. At a minimum. AMD is still holding back really describing how the doubled FP performance per core (yielding 4x the performance when cores are also doubled) fits into the power consumption specs.
-Matt
CBeddoe - Sunday, November 11, 2018 - link
You're right on the money.Power consumption and efficiency are what really matter for Data Centers.
If they have the same crew runn those benchmarks as they did at CES its probably running at 20kwatts with a 20kwatt phase change cooler and a hacked motherboard.
AMD at least has a process shrink to help drive efficiency a bit how much remains to be seen but their claims of 25% aren't off the wall.
mode_13h - Monday, November 12, 2018 - link
Actually, AMD did say they doubled the width of their vector pipelines to 256-bit.yannigr2 - Monday, November 12, 2018 - link
Don't expect half power consumption, considering that the I/O chip is still at 14nm. But it will be that much lower that it will make any Intel AMD power consumption comparisons laughable at best. Think what is happening with Nvidia/AMD GPUs but this time with AMD having the clear advantage.Gondalf - Monday, November 12, 2018 - link
Don t be so confident, Xeon AP will be low clocked like Rome, in these conditions 14nm++ is pretty in line with 7nm (early) power savings. Rome will be hot unfortunately.Intel is right now selling Xeon AP because customers have the exact power figures of Rome (AMD always share the data six/eight months before shipment to have an idea of the volume to order to foundry)
MonkeyPaw - Sunday, November 11, 2018 - link
In other words, “hey, please delay your next order until we can get out act together!”EPYC 2 sounds like it will have more IPC and better clocks, not to mention memory bandwidth gains. We won’t be asking how Intel’s next CPU compares to AMDs old CPU.
Spunjji - Monday, November 12, 2018 - link
This is what's making me laugh. They insist on comparing to a product that will be old news when theirs releases, assuming it releases on time.I wonder how much they paid out in "marketing development funds" to get all those partner logos on their slides. :D
shabby - Sunday, November 11, 2018 - link
Did they mention if they're using Elmer's glue or gorilla glue? 🤔yeeeeman - Monday, November 12, 2018 - link
So lets put these benchmarks into context. They compare a 48 core CPU with a 64 core system made from 2 CPUs each 32 cores each. Fair enough, they do manage to get better results with lesser cores, but I am pretty confident that these are scenarios where inter-CPU communications matter and we all know that the latency is high on a dual socket system.Second, AMD has their new CPU line-up just around the corner and they double pretty much everything. They also quadruple FP resources.
If things go well, then it could very be that one Rome CPU could keep up with 2S Cascale Lake AP system, which would be disastrous for Intel.
GruenSein - Monday, November 12, 2018 - link
If I understand correctly, the figures refer to a system with two sockets each occupied by a 48 Core Cascade Lake AP. That means that they get 1.5-1.6x the performance with 1.5x the number of cores of Epyc 7601 in most scenarios and maintain a performance lead on some hyper-optimized AVX-workloads. And this is comparing an unreleased system to AMD's line of processors, which have been available for a year or so. No power figures are provided and only few benchmarks are shown. I think, it is safe to say that Zen 2 will be much faster based on the number of cores alone. That is not even considering the expected improvements in IPC, clocks, and power efficiency.yeeeeman - Monday, November 12, 2018 - link
Then they are doomedFitzgeraldH - Monday, November 12, 2018 - link
Gruensein is indeed correct. Slide 6 states 2x48 cores Cascade Lake (it's the last sentence in each paragraph).Spunjji - Monday, November 12, 2018 - link
Solid summary. +1rahvin - Monday, November 12, 2018 - link
Intel disabled SMT on the Epyc chip in their "benchmarks". You should keep that mind. They probably did this because their own chip won't have hyper-threading, anything you can do to show your own processor in the best light.The test on how scared intel is will be what they do with prices. If they drop prices where they've planned to raise them 20% we'll be seeing AMD taking more of the cloud operators.
lordtux - Monday, November 12, 2018 - link
Well, on the footnotes is showing 2x48c against a 2x32c, so its 96c against 64c. Seems that everyone is thinking that one CPU of 48c is beating a 64c. This turn the things worst for Intel.schujj07 - Monday, November 12, 2018 - link
They also didn't say whether SMT was enabled on the Intel system or not. On several of the benchmarks SMT was disabled on the Epyc and then enabled on other benchmarks, however, no mention of what it is for Intel.Rukur - Monday, November 12, 2018 - link
It's a rainbow unicorn CPU. Where is my 5GHz 28 core workstation and abomination of a motherboard Intel ?.vodka - Monday, November 12, 2018 - link
Shortly after the public demo, the one and only sample that could do 5GHz on all 28 cores crashed and burned.Spunjji - Monday, November 12, 2018 - link
What a bargain - the product provided its own viking funeral at no additional cost. :Dmuziqaz - Monday, November 12, 2018 - link
This is 64 AMD cores vs 96 intel cores, since SMT is disabled in 7601.yannigr2 - Monday, November 12, 2018 - link
In Linpack testing, Intel decided that disabling SMT on EPYC was the correct thing to do.brakdoo - Monday, November 12, 2018 - link
So, AMD and Intel are pushing parallel FP benchmarks.Someone tell them that CPUs already lost this battle (despite AVX-512 for Intel) as HPC users almost always use accelerators for these use cases like CFD and weather forecast. Just look at the TOP500...
Spunjji - Monday, November 12, 2018 - link
There are still some cases where splitting the data out to separate accelerators takes more time (either in programming effort or just in terms of waiting around) than just performing the calculations on CPU, but you're right that the number of scenarios these performance comparisons apply to is shrinking rapidly.saylick - Monday, November 12, 2018 - link
This does not bode well for Intel. 2x48 only manages to beat out 2x32 by 1.5x to 1.6x for non-AVX heavy workloads and for AVX-heavy workloads, the lead jumps to 3.4x. The fact that Intel is pitting 50% more cores against AMD in this comparison already accounts for that 1.5x to 1.6x performance uplift... With EPYC 2, AMD touts 2x the perf/W and 2x the cores AND with AT's interview with Mark Papermaster, AMD will support AVX2 without backing off on clocks. In a 2S head to head, Intel's 2x48 vs AMD's 2x64 will show AMD having the performance lead across the board at a power consumption level similar to a 2S EPYC 1 system.edzieba - Monday, November 12, 2018 - link
The interesting slides are not the performance numbers, but the last two: vendors designing systems around the chips. Checking the latest TOP500, there are not yet AMD design wins listed (for Epyc, much less Rome), and none that I am aware of under construction or design (though that could have changed).ilt24 - Monday, November 12, 2018 - link
@edzieba..."Checking the latest TOP500, there are not yet AMD design wins listed (for Epyc..."#38 on the newest TOP500 list contains 5120 - EPYC 7501 (32 core) processors...note the processors are listed as Hygon Dhyana Epyc 7501 processors, which comes from a JV between AMD and a company created by the Chinese government.
edzieba - Tuesday, November 13, 2018 - link
Ah, I'd done a ctrl+f for AMD. I'd forgotten about the Hygon and THATIC legal dodges.RogerAndOut - Tuesday, November 13, 2018 - link
WOW, Intel is spinning in just so many ways.- Intel is having to commit what in many ways its future 4-way server design to compete with AMD's current 2-way server. This is not a great story for the majority of deployments that are 1 or 2 socket systems and 4 socket systems are meant to be a high-profit area for Intel.
- Intel disabled SMT on the AMD system for the first test and many of the other tests indicate that the number of threads running was set at 1 per core. So SMT was not used where possible.
- Intel used their C/C++ compiler to compile the test suites where possible. There is no indication that they used the AMD Optimized compiler for the AMD systems.
- The header text states that microprocessor-dependent optimizations and non-specific optimizations were only carried out for the Intel processors.
All in all the info seems to have come from the same marketing team that release complete spin for the i9-990K and before that the consumer version of the 28 chore Xeon Platinum 8180.
FredWebsters - Wednesday, December 26, 2018 - link
I think that there are only cool products!joana - Monday, April 13, 2020 - link
Do you think it will be <a href=”https://www.romexsoft.com/blog/aws-hipaa-complianc...”>good</a>?