Speculation: AMD's response to Intel's 8-core i9-9900K

[epsilon84]

Thats where all my E5-2683v3 Xeons came from, until Ryzen hit. They were $350 on ebay and great performers. But now my 8 core Ryzens are stomping their 14 cores, so Ryzen it is. Oh, and with less power draw.

Would you consider a 9900K for your purposes if it 'stomps' a 2700X? Genuine question.

 

[PeterScott]

Many, you have one M2.0 drive and then you have a Secondary large SATA/USB 3.0/1 drive for storage. The moment you are going to transfer/copy from one to the other or to a third drive, your primary M2.0 on Intel Z370/390 will have half the speed.

That doesn't add up, copying to a Sata/USB3.0 drive wont saturate half of Intels DMI PCI lanes.

DMI 3.0 has 3.9 GB/s bandwidth. The Maximum SATA Bandwidth is 600 MB/s.

Even a full maximum speed SATA Copy between two devices would not use half the DMI 3.0 bandwidth.

So, no you haven't lost half your Bandwidth.

But now look at a Drive connected to the second NVMe Slot on AMD MBs:

You lose more than half your throughput. Not on some rare, occasional event. But all the time.

 

[Markfw]

Would you consider a 9900K for your purposes if it 'stomps' a 2700X? Genuine question.

Well, actually, maybe. But right now the price is supposed to be about 50% more, so they would need to beat it by 50% for me to consider it, and I highly doubt that.

But thanks for an honest question, and I think I gave an honest answer.

 

[epsilon84]

Well, actually, maybe. But right now the price is supposed to be about 50% more, so they would need to beat it by 50% for me to consider it, and I highly doubt that.

But thanks for an honest question, and I think I gave an honest answer.

So you only buy based on CPU price/performance and not platform or system price/performance? I would have thought that would be more appropriate metric given you can't run a CPU in isolation - at a minimum, you also need a motherboard and a decent amount of RAM, particularly for your usage. This is assuming you already have a case, PSU, SSD etc in place. If you don't then it further dilutes the value proposition.

A $150 price difference might seem like a lot in isolation strictly comparing two CPUs, but part of a $1000 system a 9900K could actually be the better value compared to an $850 2700X system, for example, as it is likely to be more than 15% faster than a 2700X. If you run a high end GPU as well then the system costs blow out to $1500 vs $1350 respectively. Assuming a 9900K overclocked to 5.0GHz is ~30% faster than a 2700X @ 4.2GHz, wouldn't that make the 9900K system the one with the better value?

This is what I don't understand when people compare value based on a singular component - it's like buying a car based on the cost of its engine alone, rather than the entire car itself. Using such logic, wouldn't something like a lowly Pentium for $60 be the ultimate price/performance chip? 2C/4T for $60 compared to 8C/16T for $330 (2700X). You get 25% the cores/threads for 18% the price... until you add up the cost of the other components. I know that is an extreme example but I'm hoping you get my point here.

 

[Markfw]

So you only buy based on CPU price/performance and not platform or system price/performance? I would have thought that would be more appropriate metric given you can't run a CPU in isolation - at a minimum, you also need a motherboard and a decent amount of RAM, particularly for your usage. This is assuming you already have a case, PSU, SSD etc in place. If you don't then it further dilutes the value proposition.

A $150 price difference might seem like a lot in isolation strictly comparing two CPUs, but part of a $1000 system a 9900K could actually be the better value compared to an $850 2700X system, for example, as it is likely to be more than 15% faster than a 2700X. If you run a high end GPU as well then the system costs blow out to $1500 vs $1350 respectively. Assuming a 9900K overclocked to 5.0GHz is ~30% faster than a 2700X @ 4.2GHz, wouldn't that make the 9900K system the one with the better value?

This is what I don't understand when people compare value based on a singular component - it's like buying a car based on the cost of its engine alone, rather than the entire car itself. Using such logic, wouldn't something like a lowly Pentium for $60 be the ultimate price/performance chip? 2C/4T for $60 compared to 8C/16T for $330 (2700X). You get 25% the cores/threads for 18% the price... until you add up the cost of the other components. I know that is an extreme example but I'm hoping you get my point here.

Also power usage. Note I said "maybe". Until all the facts are in I stay by that. Oh, and platform aging. AM4 is supposed to work for years. The Intel platform is like dead.

Edit: Not all price/performance. I (myself) look at the top of the performance ladder. Then I see what makes sense. If I didn't already have 4 threadrippers, in the last 2 weeks to get them for $620-$700, I would have gotten several. The new one with 32 cores does not make much sense, since 2 of the 16 cores would be $1400, and its $1800. But I only need one PSU and case. And Ryzen has way more PCIE lanes

So a lot to consider.

 

[tamz_msc]

But now look at AMD Drive connected to the second NVMe Slot:

That's great, but where is the graph for 2 NVMes connected to the PCH on the Intel platform?

I bet it's no worse than what happens to the NVMe drive in the second slot on AMD X470:

Why don't you bring some evidence to the table for a change, since you implore others to do the same?

 

[Zucker2k]

That's great, but where is the graph for 2 NVMes connected to the PCH on the Intel platform?

Why don't you bring some evidence to the table for a change, since you implore others to do the same?

The onus is on the one making the argument to provide proof, just in case you needed reminding

 

[tamz_msc]

The onus is on the one making the argument to provide proof, just in case you needed reminding

Who is making the argument that two NVMe drives running off the Intel PCH don't bottleneck it? Certainly not me.

 

[Zucker2k]

Who is making the argument that two NVMe drives running off the Intel PCH don't bottleneck it? Certainly not me.

Who is arguing they do?

 

[dlerious]

Who is arguing they do?

If you have 2 PCIe 3.0 x4 drives connected to a single PCIe 3.0 x4 connection, do you really expect both drives to run at their max speed?

 

[ChiefBigFeather]

I'm not sure if my old mind is playing tricks on me or if I remember right, but I think connecting your m.2 via PCH used to gimp it. And if I got this right, the bigger problem for the average end user was not bandwidth but latency. 4k random read latency on a PCH used to be in SATA territory.

So I asked over at storage review and continuum found this review which is miles better then that techspot review:
https://www.pcper.com/reviews/Storage/Quick-Look-AMD-Ryzen-X470-NVMe-Storage-Performance

Whatever m.2 via PCH used to be, on the z270 it is actually faster then 2600x cpu lanes! At least at every day enduser tasks. The wrong description of the chart aside, 4k random reads at qd1 have quite a lot lower latency on z270. I hope AMD can improve here with zen 2.

 

[tamz_msc]

I'm not sure if my old mind is playing tricks on me or if I remember right, but I think connecting your m.2 via PCH used to gimp it. And if I got this right, the bigger problem for the average end user was not bandwidth but latency. 4k random read latency on a PCH used to be in SATA territory.

So I asked over at storage review and continuum found this review which is miles better then that techspot review:
https://www.pcper.com/reviews/Storage/Quick-Look-AMD-Ryzen-X470-NVMe-Storage-Performance

Whatever m.2 via PCH used to be, on the z270 it is actually faster then 2600x cpu lanes! At least at every day enduser tasks. The wrong description of the chart aside, 4k random reads at qd1 have quite a lot lower latency on z270. I hope AMD can improve here with zen 2.

That test only concludes that the RAIDXpert driver is causing a bottleneck on the AMD platform. With one NVMe attached to the chipset, you get half the bandwidth, for sure, but the other one connected directly to the CPU still gets full bandwidth. On Intel there isn't a way to connect NVMe directly to the CPU without a PCIe adapter. The PCPer article has the "CPU-connected" NVMe on the Z270 platform using an adapter.

 

[moinmoin]

@PeterScott Don't know why you feel like wasting your time and energy defending the indefensible. Putting two NVMe through Intel's PCH is inane, and nothing about AMD's chipset makes that more defensible. In the end for both systems the best way to connect a second NVMe is through a PCIe 3.0 x4 adapter card.

 

[JoeRambo]

I did the following test on Z370:

Copied large disk image to external USB3 Sam T5 drive from SAM M.2 that is connected to chipset
And ran Aida64 linear test on a different Sam M.2 drive during the same time. I guess the results speak for themselves:

During:

After USB copy was done:

 

[tamz_msc]

I did the following test on Z370:

Copied large disk image to external USB3 Sam T5 drive from SAM M.2 that is connected to chipset
And ran Aida64 linear test on a different Sam M.2 drive during the same time. I guess the results speak for themselves:

During:

After USB copy was done:

So Max read throughput(w/o copying) = Max read throughput while copying + copy transfer rate. Thanks for the confirmation.

 

[JoeRambo]

So Max read throughput(w/o copying) = Max read throughput while copying + copy transfer rate. Thanks for the confirmation.

Yeah, ~2500MB/s are being read from M.2, 400MB/s are written to USB3 external drive, and those 400MB/s come from another M.2 drive. So at least 3400MB/s total load on DMI, before we consider overheads of USB3, SATA would probably inflict even less reduction.

EDIT: and system was in use at the time, perfectly usable, sound was coming out of USB headphones, playing youtube video in background.

 

[ChiefBigFeather]

With one NVMe attached to the chipset, you get half the bandwidth, for sure, but the other one connected directly to the CPU still gets full bandwidth. On Intel there isn't a way to connect NVMe directly to the CPU without a PCIe adapter. The PCPer article has the "CPU-connected" NVMe on the Z270 platform using an adapter.

My statement was about latency, not bandwidth. As far as I know, latency is much more important for every day use, app load times and overall system responsiveness then bandwidth (unless you take it to extremes ofc). I'm not entirely sure though.

 

[ub4ty]

I did the following test on Z370:

Copied large disk image to external USB3 Sam T5 drive from SAM M.2 that is connected to chipset
And ran Aida64 linear test on a different Sam M.2 drive during the same time. I guess the results speak for themselves:

During:

After USB copy was done:

They do speak for themselves and confirm what I stated : In the first case, your simultaneous transfer to the USB cut the speed of the 970 Pro M.2 Nvme Drive because your chipset created a bottleneck as it swamped. In the second case, when you weren't doing the simultaneous transfer, the speeds of the SSD 970 Pro increased roughly by amount of the USB transfer hilariously showing this even further.

Definitive conclusion which was obvious to anyone before the test : Intel's decision to put so much I/O through such a little pipe (DMI 3.0 - PCIE 3.0 x4) was ridiculous especially considering that NVME are first class citizen storage devices that can slam the bandwidth provided by PCIE 3.0 x4. There are a number of use cases where I most definitely don't want my NVME storage hanging off of the chipset where I have a wealth of traffic flowing through.

 

[ub4ty]

So Max read throughput(w/o copying) = Max read throughput while copying + copy transfer rate. Thanks for the confirmation.

Concluding : DMI 3.0 found on Z370 and X299 (for some insane reason) which only has a bandwidth of PCIE 3.0 x4 is a bottleneck if you have anything beyond one NVME drive fully active. What's astonishing to me is that they have this crap on X299 (an HEDT platform). Meanwhile, on my threadripper, I have 3 NVME drives on their own dedicated (PCIE 3.0 x4) lanes + x16/x16/x8/x8 for GPUs + PCIE 3.0 x4 shared for chipset devices. Seriously, this is the kind of crap stretched all over intel's platform that got me infuriated with them. Then they have a literal alphabet soup of chipsets and sockets for an alphabet soup offering of processors that all locked at 4 cores for so long. Then, AMD blows their socks off and they finally start offering higher core counts but with the same gimped chipset nonsense ALL THE WAY UP into HEDT.

I was so thankful when Ryzen came out. The more I dug deeper down into the specs, the more I saw an enterprise processor and platform in disguise. @#@*$! Intel

 

[PeterScott]

They do speak for themselves and confirm what I stated : In the first case, your simultaneous transfer to the USB cut the speed of the 970 Pro M.2 Nvme Drive because your chipset created a bottleneck as it swamped. In the second case, when you weren't doing the simultaneous transfer, the speeds of the SSD 970 Pro increased roughly by amount of the USB transfer hilariously showing this even further.

Definitive conclusion which was obvious to anyone before the test : Intel's decision to put so much I/O through such a little pipe (DMI 3.0 - PCIE 3.0 x4) was ridiculous especially considering that NVME are first class citizen storage devices that can slam the bandwidth provided by PCIE 3.0 x4. There are a number of use cases where I most definitely don't want my NVME storage hanging off of the chipset where I have a wealth of traffic flowing through.

Practically an insignificant reduction compared to using the Second NVMe slot on AMD's X470 platform.

By doing a simultaneous copy during a full transfer speed test, he reduced speed on Z370 drive to ~2600 MB/s.

Using the Second NVMe slot on AMD X470 speed was reduced to ~1300 MB/s, while it was doing nothing at all.

So Z370 was still twice as fast while doing a simultaneous copy. How terrible for Intel.

In practical terms running two fast NVMe drives:

1: On Intel Z370 your drives are almost always running full speed in normal use cases (not benchmarks will copying) and very occasionally you share some bandwidth and still have very high speed, that would almost certainly go unnoticed.

2: On AMD X470, one of your drives is fast all the time, while the other drive is drastically slower all the time.

Only extreme AMD partisans, make the the Z370 bandwidth sharing into some kind of major show stopping drawback, while quietly ignoring how much worse it is for X470.

 

[tamz_msc]

Practically an insignificant reduction compared to using the Second NVMe slot on AMD's X470 platform.

By doing a simultaneous copy during a full transfer speed test, he reduced speed on Z370 drive to ~2600 MB/s.

Using the Second NVMe slot on AMD X470 speed was reduced to ~1300 MB/s, while it was doing nothing at all.

So Z370 was still twice as fast while doing a simultaneous copy. How terrible for Intel.

In practical terms running two fast NVMe drives:

1: On Intel Z370 your drives are almost always running full speed in normal use cases (not benchmarks will copying) and very occasionally you share some bandwidth and still have very high speed, that would almost certainly go unnoticed.

2: On AMD X470, one of your drives is fast all the time, while the other drive is drastically slower all the time.

Only extreme AMD partisans, make the the Z370 bandwidth sharing into some kind of major show stopping drawback, while quietly ignoring how much worse it is for X470.

What happens when you perform the same test using another NVMe instead of an external USB3 SSD?

 

[ub4ty]

Practically an insignificant reduction compared to using the Second NVMe slot on AMD's X470 platform.

By doing a simultaneous copy during a full transfer speed test, he reduced speed on Z370 drive to ~2600 MB/s.

Using the Second NVMe slot on AMD X470 speed was reduced to ~1300 MB/s, while it was doing nothing at all.

So Z370 was still twice as fast while doing a simultaneous copy. How terrible for Intel.

In practical terms running two fast NVMe drives:

1: On Intel Z370 your drives are almost always running full speed in normal use cases (not benchmarks will copying) and very occasionally you share some bandwidth and still have very high speed, that would almost certainly go unnoticed.

2: On AMD X470, one of your drives is fast all the time, while the other drive is drastically slower all the time.

Only extreme AMD partisans, make the the Z370 bandwidth sharing into some kind of major show stopping drawback, while quietly ignoring how much worse it is for X470.

Partitioning...
Primary storage
Secondary storage
Tiering
Using an amazingly fast primary storage device as a cache space

Chipset devices :
SATA/PCIE 2.0/NVME #2/USB/etc/etc

The point is to allow for a clear partition and tiering while realizing that chipset devices are 2nd class citizens.
You wont grasp this nor the use cases that require it. An 8 core processor is not a run of the mill basic desktop platform. It's a power user/server grade core count. I am currently on a dual core 2.5ghz i5. It's what I use when doing mundane tasks like browsing the web. I don't use an 8core based machine to do regular user tasks. If I use SSD as my OS drive and NVME exclusively as a caching drive between ram using my own caching policy [I wrote the program], my nvme is most definitely going to be slamming the PCIE 3.0 x4 interface. I don't want this impacting all of my other 2nd class citizens hanging off of the chipset. I don't want 2nd class citizen data flow activity impacting my high speed caching performance.

You're not thinking like an enterprise user and fail to realize that's what an 8 core count processor is. Intel never wanted people to realize this because they wanted to continue their Xeon meme pricing and sales. There are a number of use cases where someone could have a constant 400MB/s flow through a chipset based device. This could increase by 2 or three more such flows easily. USB 3.1 is 10Gb/s .. ~1.2 GB/s. There is absolutely zero reason to have my #1 class storage device impacted by 2nd rate citizen data flows that can be substantial for a number of users. There is zero reason to let one second class citizen rob all the others.

I've concluded my exchanges with you.
You don't want to get the point and have a sea of excuses every-time someone makes one.
You keep using terms like Partisans. This isn't politics man. This is technology/science/math/data/facts.
There is zero room for politics here. The AMD 2ndary NVME slot has much slower performance because they restrict it to a PCIE 2.0 interface off the chipset treating it like they treat every other device hanging off of it : A 2ndary citizen. No one device should be able to hog the chipset's bandwidth. That's the whole point of a chipset based shared interface : Give slower citizens equal access to a shared lane.

This is actually the very thing that intel were such @(#&@! about .. PCIE lanes. They reserved higher lane counts for enterprise customers that they charged an arm and a leg for via Xeons (even there they kept lane counts artificially low). They do so to force people into these proprietary interfaces that are exclusive to Intel. PCIE was created for the reason of being an open and standard interface to the CPU. Intel fights this constantly with proprietary gimmicks. NVME comes out and they had years of boards to figure out that it should have a dedicated interface to the CPU. They resisted and fought this and hung it off some stupid chipset interface that Used to be PCIE 2.0. Then they come out with Optane...

They engaged in PCIE gimmicks for desktop/enterprise users. Nvidia absolutely hated them for this as this shut them out of sales. Intel did it on purpose because Nvidia's GPU's were becoming first class citizens in data centers.

You don't seem to maintain this broad knowledge/understanding nor a history of understanding that would have formed if you had been historically active and non-partisan. Instead you try to run every sound discussion into the ground w/ these partisan accusations. Someone states they have just as much Intel equipment as AMD... that doesn't register to you. I highlighted the clear issue DMI 3.0 has at an architectural level and you still voiced disbelief of something that is painfully obvious. A user goes out of their way to show you a benchmark demonstrating the very thing I stated and you still have excuses.

Get your head out of your butt man. No one's buying this charade whose a serious systems builder.
When even Linus has negative things to say and is 100% right, you know you've messed up :
https://www.youtube.com/watch?v=TWFzWRoVNnE

Intel's platform for years has been centered on artificial segmentation and a land mind laden maze to trap users and exploit them. AMD caught them with their pants down and delivered a far better platform. Until Intel gets their head out of their butts (not going to happen when apologist keep making excuses and buying into their madness), they are non-existent to me.

https://www.amazon.com/Intel-VROC-STD-RAID-VROCSTANMOD/dp/B077PXMDMZ
https://www.pcworld.com/article/319...le-crazy-raid-configurations-for-a-price.html
FFS !

 

[PeterScott]

You're not thinking like an enterprise user...

Because we are talking about a desktop chip, for desktop users.

You need to adjust your thinking to be more appropriate for desktop users, not the other way around.

 

[ub4ty]

Because we are talking about a desktop chip, for desktop users.

You need to adjust your thinking to be more appropriate for desktop users, not the other way around.

Or buy AMD that designed a more forward thinking architecture...

KEK and /thread

 

[PeterScott]

Or buy AMD that designed a more forward thinking architecture...

KEK and /thread

You are really making the case that you just view everything through the lens of AMD partisanship.