Speculation: Ryzen 4000 series/Zen 3

[scannall]

They were talking 56 cores on one socket, not 2.

They were referring to the Xeon 9200. You could only get it on an Intel motherboard. Very few were made at around $45,000 each. Likely none are still available. It's really a pointless part at that price.

 

[lobz]

They were referring to the Xeon 9200. You could only get it on an Intel motherboard. Very few were made at around $45,000 each. Likely none are still available. It's really a pointless part at that price.

The poster he answered to was not talking about Cascade Lake-AP, he explicitly said 2 CPUs per system.

 

[eek2121]

IMO it is a weakness AMD still has. You can get a 56 core dual CPU Intel system for a price that isn’t too far away from EPYC. While EPYC is superior for many workloads, Intel still wins in some key areas (including our platform, for whatever reason our extremely large web application saw a 40% hit when we tried to test), and between that and brand recognition it ends up being a tough fight.

AMD really needs to increase core counts, although that may not happen before Zen 4. If they could sell a 96-128 core chip for the price of the 2x28 core then their platform would essentially be untouchable for multicore workloads.

I hope we hear more about Zen 3 soon, because honestly I can’t wait.

 

[itsmydamnation]

IMO it is a weakness AMD still has. You can get a 56 core dual CPU Intel system for a price that isn’t too far away from EPYC.

I dont know where you get that idea from ( just CPU price)
but by that logic, who buys just 1 server ( no one) so you can get 1 AMD 128core server with say 512gb memory for less then 2x 56 core 256gb intel servers.


edit: i was actually pricing server builds in cisco ccw/ HPE OCA today

 

[Markfw]

IMO it is a weakness AMD still has. You can get a 56 core dual CPU Intel system for a price that isn’t too far away from EPYC. While EPYC is superior for many workloads, Intel still wins in some key areas (including our platform, for whatever reason our extremely large web application saw a 40% hit when we tried to test), and between that and brand recognition it ends up being a tough fight.

AMD really needs to increase core counts, although that may not happen before Zen 4. If they could sell a 96-128 core chip for the price of the 2x28 core then their platform would essentially be untouchable for multicore workloads.

I hope we hear more about Zen 3 soon, because honestly I can’t wait.

2 28 core Intel=$8000 and one 64 core AMD = $4000. Where on earth does your math come from that its "a price that isn’t too far away from EPYC". Now mind you I get these prices from Amazon/ebay, but still what planet are you posting from ? And where are your benchmarks that show its even close in ANY benchmark, the only ones I have seen its a massacre.

 

[Gideon]

Getting back on topic:

The recent info about Renoir (much higher FCLKs) and leaks (4.9Ghz Boost) make me quite hopeful of Vermeer being a kickass product. Looking forward to upgrading to it + RDNA2 (if the latter indeed delivers vs RTX 3xxx).

There's one thing I'm wishing for (though I'm almost certain it won't happen) is AMD bringing back AMP - which would set all memory subtimings (similar to what currently needs to be done manually with Ryzen DRAM Calculator) rather than just the primary ones, like XMP.

I have no issue manually tuning timings myself, but the reality is that most of the user will never do it themselves (hell a lot aren't even knowledgeable enough to enable XMP) and that just leaves 5-10% perfomance on the table (just look at the linked 3600 XT results with XMP 3200 CL14 vs 1:1 3800 Mhz with a heavy tune).

I've helped a couple of friends set them multiple times (the last time just this weekend) and it has enough caveats that the layman will never do it (or at least do it well). Computer-savy people usually can do it themselves on the second try, but still often miss some dibits, or don't memtest thoroughly ending up with an erratically unstable system.

AMD has the DIY market share now (and most definitely will with Zen 3 launch) to matter enough to memory-companies to pull this off as It would benefit everybody:

• Memory makers can charge a premium from validated AMP modules (That get, say 90-95% of the custom timing performance)
• I know a number of users that would gladly pay a premium to avoid setting these timings themselves (and again after each BIOS update)
• A lot of reviewers (not Anandtech though) do test with XMP, so it would also help AMD get better launch reviews.
• You can still buy the cheaper kits and manual-tune, if you wish.

I can understand if this fails to gain enough traction from mobo/memory companies but IMO AMD is really missing out if they are not even trying. If there is not enough incentive for AMD they could just differentiate new chipsets with it (e.g. X670, B650 with AMP support). I hope not but a thought.

Realistically I don't think they'll do it that far into the socket life-cycle (if at all, it might arrive with Zen4/DDR5, that needs a new XMP-like standard anyhow) but one can hope

 

[coercitiv]

The recent info about Renoir (much higher FCLKs) and leaks (4.9Ghz Boost) make me quite hopeful of Vermeer being a kickass product.

In case you missed it, we already have a reliable leak on current Vermeer clocks thanks to Igor's Lab - engineering sample with 4.9Ghz boost and 3.7Ghz base sounds great for a 16c/32t part, even if only for the top bin.

 

[JoeRambo]

2 28 core Intel=$8000 and one 64 core AMD = $4000. Where on earth does your math come from that its "a price that isn’t too far away from EPYC". Now mind you I get these prices from Amazon/ebay, but still what planet are you posting from ? And where are your benchmarks that show its even close in ANY benchmark, the only ones I have seen its a massacre.

Intel has completely revised lineup and pricing. Again, everyone with actual brain is not buying top of the lines from either of companies. 64 cores and 28 cores are completely irrelevant for most buyers.

Where the value currently is with Intel is 6248R 24C 3Ghz base 4Ghz turbo CPU for $2400 list price. 95% of 28C performance for ~60% of price?
~same money can buy 32C from AMD, with 2.5 Base / 3.3 turbo clock. Not exactly massacre anymore?
Lets replace 2S Intel 6248R with 1S Epyc? 7552/ 7662 offer 48C with ~2.2 base and ~3.3 turbo for ~$4-5K ? Where's massacre here.

For web application servers Intel is currently very compelling in pricing and features. For other servers like web cache/front etc -> not so, but both AMD and Intel will soon get kicked out by ARM there.

Forum warriors can keep talking about massacres while running Cinebench in the background, but Intel will keep making sales due to price/perf even at the list price. And company purchase decision makers sure don't care about Intel or AMD bottom line.

 

[DrMrLordX]

Not exactly massacre anymore?

You're ignoring TCO.

 

[Thunder 57]

Getting back on topic:

The recent info about Renoir (much higher FCLKs) and leaks (4.9Ghz Boost) make me quite hopeful of Vermeer being a kickass product. Looking forward to upgrading to it + RDNA2 (if the latter indeed delivers vs RTX 3xxx).

There's one thing I'm wishing for (though I'm almost certain it won't happen) is AMD bringing back AMP - which would set all memory subtimings (similar to what currently needs to be done manually with Ryzen DRAM Calculator) rather than just the primary ones, like XMP.

I have no issue manually tuning timings myself, but the reality is that most of the user will never do it themselves (hell a lot aren't even knowledgeable enough to enable XMP) and that just leaves 5-10% perfomance on the table (just look at the linked 3600 XT results with XMP 3200 CL14 vs 1:1 3800 Mhz with a heavy tune).

I've helped a couple of friends set them multiple times (the last time just this weekend) and it has enough caveats that the layman will never do it (or at least do it well). Computer-savy people usually can do it themselves on the second try, but still often miss some dibits, or don't memtest thoroughly ending up with an erratically unstable system.

AMD has the DIY market share now (and most definitely will with Zen 3 launch) to matter enough to memory-companies to pull this off as It would benefit everybody:

• Memory makers can charge a premium from validated AMP modules (That get, say 90-95% of the custom timing performance)
• I know a number of users that would gladly pay a premium to avoid setting these timings themselves (and again after each BIOS update)
• A lot of reviewers (not Anandtech though) do test with XMP, so it would also help AMD get better launch reviews.
• You can still buy the cheaper kits and manual-tune, if you wish.

I can understand if this fails to gain enough traction from mobo/memory companies but IMO AMD is really missing out if they are not even trying. If there is not enough incentive for AMD they could just differentiate new chipsets with it (e.g. X670, B650 with AMP support). I hope not but a thought.

Realistically I don't think they'll do it that far into the socket life-cycle (if at all, it might arrive with Zen4/DDR5, that needs a new XMP-like standard anyhow) but one can hope

Chalk me up as one that doesn't bother with the subtimings. Maybe that's just me being a simpleton though. It's not that I can't; Rather it's not wanting to deal with it after every BIOS reset.

 

[Thibsie]

If ARM is bought by Nvidia, ARM is at risk of being kicked rather than the opposite.

 

[Thunder 57]

Intel has completely revised lineup and pricing. Again, everyone with actual brain is not buying top of the lines from either of companies. 64 cores and 28 cores are completely irrelevant for most buyers.

Where the value currently is with Intel is 6248R 24C 3Ghz base 4Ghz turbo CPU for $2400 list price. 95% of 28C performance for ~60% of price?
~same money can buy 32C from AMD, with 2.5 Base / 3.3 turbo clock. Not exactly massacre anymore?
Lets replace 2S Intel 6248R with 1S Epyc? 7552/ 7662 offer 48C with ~2.2 base and ~3.3 turbo for ~$4-5K ? Where's massacre here.

For web application servers Intel is currently very compelling in pricing and features. For other servers like web cache/front etc -> not so, but both AMD and Intel will soon get kicked out by ARM there.

Forum warriors can keep talking about massacres while running Cinebench in the background, but Intel will keep making sales due to price/perf even at the list price. And company purchase decision makers sure don't care about Intel or AMD bottom line.

You do realize that for someone trying to bash "forum warriors", you sort of sound like a "forum warrior"?

 

[JoeRambo]

You do realize that for someone trying to bash "forum warriors", you sort of sound like a "forum warrior"?

Anyone using "massacre" talking about server CPUs that are irrelevant to 99% of users here already is.

And TCO is rather subtle to access. When not running avx256/512 workloads those Intel machines are fine in power usage. Even if we ignore official TDP values or tests at sites like Servethehome and use common sense -> all those fabric links chew into power efficiency budget provided by better process.

 

[Makaveli]

Chalk me up as one that doesn't bother with the subtimings. Maybe that's just me being a simpleton though. It's not that I can't; Rather it's not wanting to deal with it after every BIOS reset.

Subtimings do make a difference and the quality of your memory matters. i've been looking at some issues on Hardocp forum that some users are having with some of the Newer AM4 bios. Users trying to run Ryzen Dram profiles on average memory running into stability issues, while on my system with B die I can run every profile in the DRam Calc and have no issues with sub timings at all.

And I get not wanting to do it after every reset. I actually wish the we were able to load a saved Ryzen Dram profile into the bios directly.

 

[Thunder 57]

Subtimings do make a difference and the quality of your memory matters. i've been looking at some issues on Hardocp forum that some users are having with some of the Newer AM4 bios. Users trying to run Ryzen Dram profiles on average memory running into stability issues, while on my system with B die I can run every profile in the DRam Calc and have no issues with sub timings at all.

And I get not wanting to do it after every reset. I actually wish the we were able to load a saved Ryzen Dram profile into the bios directly.

Never said either did not. I just don't care enough to chase every last gain anymore I guess. I don't even bother with PBO since the difference seems to be rather small. What I have servers me well.

As for the BIOS, yea it would be nice if you could save things like timings and other options. I get that things change but do I really have to turn off RGB all the time?

 

[Kenmitch]

Users trying to run Ryzen Dram profiles on average memory running into stability issues, while on my system with B die I can run every profile in the DRam Calc and have no issues with sub timings at all.

Did you push your b-die beyond the calculators profiles? I went it alone all the way up to 4333MHz stabbing in the dark with mine. It was kind of a pointless adventure paired with a 3700x when the memory divider reared it's ugly head beyond 3800MHz.

 

[Kenmitch]

As for the BIOS, yea it would be nice if you could save things like timings and other options. I get that things change but do I really have to turn off RGB all the time?

Your motherboard doesn't have overclocking profiles? Or they just don't work?

 

[Makaveli]

Did you push your b-die beyond the calculators profiles? I went it alone all the way up to 4333MHz stabbing in the dark with mine. It was kind of a pointless adventure paired with a 3700x when the memory divider reared it's ugly head beyond 3800MHz.

nah i'm never bothered pushing it that far. I'm very happy with the system at the moment its very stable and no random issues. I pretty much just run PBO and a Fast profile from Dram Calc and that's about it.

Not to mention I work from home and this is my primary desktop in the house beside the work laptop so stability matters more for me than pushing for every last percentage.

 

[lobz]

IMO it is a weakness AMD still has. You can get a 56 core dual CPU Intel system for a price that isn’t too far away from EPYC. While EPYC is superior for many workloads, Intel still wins in some key areas (including our platform, for whatever reason our extremely large web application saw a 40% hit when we tried to test), and between that and brand recognition it ends up being a tough fight.

AMD really needs to increase core counts, although that may not happen before Zen 4. If they could sell a 96-128 core chip for the price of the 2x28 core then their platform would essentially be untouchable for multicore workloads.

I hope we hear more about Zen 3 soon, because honestly I can’t wait.

In many cases people like to hide incompetence and laziness behind masks with pretty words, like workload compatibility and the unbearable pain of validation etc. Not always, but many times. This behavior can cost a lot of money to companies. The real incompetence lies among executives who can't be bothered to get a second opinion, as long as they get their compensation either way. I'd be very curious, exactly what kind of large web application sees a 40% hit when going from Cascade Lake to a properly configured Rome.

Until then, please don't forget that the top of the line 2x64 core AMD system is still cheaper than the top of the line 2x28 core Xeon system, while offering much more memory bandwidth and vastly superior I/O, with the one exception of Optane DIMMs (for now).

 

[jpiniero]

Until then, please don't forget that the top of the line 2x64 core AMD system is still cheaper than the top of the line 2x28 core Xeon system,

Not anymore with the "Price Cut". The 24/28C Gold Refresh prices are much closer to the 32 core Epycs.

 

[Hitman928]

Intel has completely revised lineup and pricing. Again, everyone with actual brain is not buying top of the lines from either of companies. 64 cores and 28 cores are completely irrelevant for most buyers.

Where the value currently is with Intel is 6248R 24C 3Ghz base 4Ghz turbo CPU for $2400 list price. 95% of 28C performance for ~60% of price?
~same money can buy 32C from AMD, with 2.5 Base / 3.3 turbo clock. Not exactly massacre anymore?
Lets replace 2S Intel 6248R with 1S Epyc? 7552/ 7662 offer 48C with ~2.2 base and ~3.3 turbo for ~$4-5K ? Where's massacre here.

For web application servers Intel is currently very compelling in pricing and features. For other servers like web cache/front etc -> not so, but both AMD and Intel will soon get kicked out by ARM there.

Forum warriors can keep talking about massacres while running Cinebench in the background, but Intel will keep making sales due to price/perf even at the list price. And company purchase decision makers sure don't care about Intel or AMD bottom line.

The 32 core Epyc offers 33% more cores, twice the maximum memory capacity, better CPU IP security, better memory encryption extensions, 33% more memory channels, pcie4 and more IO in general, 3200 MT/s versus 2933 MT/s memory support, and significant energy savings for the same price as the 6248R with 20% higher base clock which doesn't tell us what the real world clocks are as Epyc works well above base clocks in many work loads, even those that scale to all 32 cores.

Maybe massacre is a bit hyperbolic, but the situations where the Intel solution would be preferred are not many.

 

[Thunder 57]

Your motherboard doesn't have overclocking profiles? Or they just don't work?

TBH I never bothered. I'm fine with things at stock other than XMP.

 

[JoeRambo]

The 32 core Epyc offers 33% more cores, twice the maximum memory capacity, better CPU IP security, better memory encryption extensions, 33% more memory channels, pcie4 and more IO in general, 3200 MT/s versus 2933 MT/s memory support, and significant energy savings for the same price as the 6248R with 20% higher base clock which doesn't tell us what the real world clocks are as Epyc works well above base clocks in many work loads, even those that scale to all 32 cores.

Maybe massacre is a bit hyperbolic, but the situations where the Intel solution would be preferred are not many.

Intel Xeon Gold 6248R Benchmarks and Review Big Refresh Gains

In our Intel Xeon Gold 6248R benchmarks and review we see the massive impact that the new "Refresh" Xeon line has on the competitive landscape

www.servethehome.com

A good review of relative performance vs 32C and 24C AMD options. Perf per dollar included as well. Intel is in no way in lead, but not getting massacred in lower range of deployments either, which is unexpected considering all Intel process debacle.

Question to those commenters pushing 64C AMD options - why does AMD even bother to offer other SKUs, like 24-32-48? And once you answer this question, does Intel offer comparable systems in same price range?

 

[lobz]

Intel has completely revised lineup and pricing. Again, everyone with actual brain is not buying top of the lines from either of companies. 64 cores and 28 cores are completely irrelevant for most buyers.

Where the value currently is with Intel is 6248R 24C 3Ghz base 4Ghz turbo CPU for $2400 list price. 95% of 28C performance for ~60% of price?
~same money can buy 32C from AMD, with 2.5 Base / 3.3 turbo clock. Not exactly massacre anymore?
Lets replace 2S Intel 6248R with 1S Epyc? 7552/ 7662 offer 48C with ~2.2 base and ~3.3 turbo for ~$4-5K ? Where's massacre here.

For web application servers Intel is currently very compelling in pricing and features. For other servers like web cache/front etc -> not so, but both AMD and Intel will soon get kicked out by ARM there.

Forum warriors can keep talking about massacres while running Cinebench in the background, but Intel will keep making sales due to price/perf even at the list price. And company purchase decision makers sure don't care about Intel or AMD bottom line.

You forget something. The $100 cheaper 7502 is much faster than the 6248R in everything outside of AVX512, while still offering vastly superior I/O, higher mem BW, higher mem capacity, lower TDP, I'd still call that a massacre.

 

[lobz]

Not anymore with the "Price Cut". The 24/28C Gold Refresh prices are much closer to the 32 core Epycs.

Servethehome showing off the new Intel value king: https://www.servethehome.com/intel-xeon-gold-6248r-benchmarks-and-review/2/
I will still call that a massacre. The only thing that happened is Intel going from cringe level milking to just offer far less value for the same money than the competition does.