AMD 3000 TR-32core Spotted

[AlexOak]

Quote

The latest Ryzen Threadripper 32 core entry is an ES1 sample with the ‘2D2832E6UIVG5_42/36_N‘ which means that it is an ES1 sample which means it’s still in the early state but the clocks have been listed at 3.6 GHz base and 4.2 GHz boost. The benchmark reports that the chip maintained an average 3.75 GHz clock through the tests. The clocks are a definite improvement from the 2990WX which had a base clock of 3.00 GHz and a boost clock of 4.20 GHz.-wccftech

Opinion:

So TR isnt dead yet and yet 3.6 ~ 3.75 clock.

 

 

Source:

https://www.userbenchmark.com/UserRun/19698768 (Benchmark)

https://www.forbes.com/sites/antonyleather/2019/08/31/amd-threadripper-3000-32-core-benchmark-leak-reveals-30-performance-boost/#67d974b22125

https://wccftech.com/amd-ryzen-threadripper-3000-32-core-64-thread-cpu-userbenchmark-30-percent-faster-2990wx/

 

Credit to @VegetableStu for other sources

Quote

geekbench entries for reference:

https://browser.geekbench.com/v4/cpu/14281648 (linux)

https://browser.geekbench.com/v4/cpu/14448604 (windows 10)

https://browser.geekbench.com/v4/cpu/compare/14448604?baseline=14281648 (comparison)
 

 

[TheKDub]

128 core 256 thread when?

[Guest]

oh wow must be slow

[Antistatic12]

31 minutes ago, AlexOak said:

Opinion:

So TR isnt dead yet and yet 3.6 ~ 3.75 clock.

Source:

https://www.userbenchmark.com/UserRun/19698768 (Benchmark)

https://www.forbes.com/sites/antonyleather/2019/08/31/amd-threadripper-3000-32-core-benchmark-leak-reveals-30-performance-boost/#67d974b22125

https://wccftech.com/amd-ryzen-threadripper-3000-32-core-64-thread-cpu-userbenchmark-30-percent-faster-2990wx/

I never thought TR was dead personally. High End hardware often goes by its own schedule.

 

If AMD really do bring out a 64c128t beast, it really shows that there has been such a change in the industry in the last 5-10 years. For the better I might add.

[Dabombinable]

5 minutes ago, Antistatic12 said:

I never thought TR was dead personally. High End hardware often goes by its own schedule.

 

If AMD really do bring out a 64c128t beast, it really shows that there has been such a change in the industry in the last 5-10 years. For the better I might add.

Especially price wise, since AMD didn't go with the far more expensive monolithic route. TBH it's good to see the days of multi die packages come around once more.

[Jurrunio]

25 minutes ago, Antistatic12 said:

If AMD really do bring out a 64c128t beast, it really shows that there has been such a change in the industry in the last 5-10 years. For the better I might add.

I'm just hope they maintain compatibility with older boards (I dont care if there's no more PCIe 4) and use an integrated voltage regulator if they do release a 64 core. The reason for the latter being the absurd current the boards need to push without it.

[leadeater]

8 hours ago, Jurrunio said:

The reason for the latter being the absurd current the boards need to push without it.

Do note Zen has always had on die voltage regulation but it's doing a bit of a different function than what you are wanting/talking about. Honestly you don't actually want much power regulation going on in the die, the added heat is only going to reduce clocks and stability. AMD went for a middle ground approach. Pushing 300W-400W in to a CPU isn't actually that much of a problem, just need to take a bit of care in the matter and not just recycle existing electrical and cooling designs. The lesson has been learned so it shouldn't be an ongoing problem.

[Doobeedoo]

This series will be quite a good upgrade. 

[GOTSpectrum]

People upgrade desktops more often, I wouldn't be surprised to see up to 48 cores this generation and 64 next myself. But then and is really in the bury intel while they can to prove a point mode of 64 core TR is also highly possible

[MeatFeastMan]

Still waiting for the 64 core beast. Where the hell is it? It's September now and we haven't had a single sighting of the 64 core part....

[cj09beira]

5 minutes ago, MeatFeastMan said:

Still waiting for the 64 core beast. Where the hell is it? It's September now and we haven't had a single sighting of the 64 core part....

we might not get it, they might top out at 48, or they might go for 64 but on a new chipset as there was a leak of 3 new chipsets trx40 trx80 and wrx80, and tr4 gets up to 32, after that memory would become a large bottleneck

[Jurrunio]

3 hours ago, leadeater said:

 Pushing 300W-400W in to a CPU

If it's a consumer platform, it will be overclocked. If it requires the boards to push that current at core voltage, all of the existing boards are screwed. Even the flagships.

[descendency]

12 hours ago, Antistatic12 said:

I never thought TR was dead personally. High End hardware often goes by its own schedule.

 

If AMD really do bring out a 64c128t beast, it really shows that there has been such a change in the industry in the last 5-10 years. For the better I might add.

I'm trying to imagine a non-server workload that would benefit from a 2Ghz 64 core machine...

 

I really doubt AMD brings that out for the reason that the market just has no reason to buy a $3000-4500 CPU that has such a niche use case.

[Humbug]

8 minutes ago, VegetableStu said:

quick update: probably temper expectations on TR3K's singlecore boost ._. 4.5 (and lower) seems to be the most reasonable target.

Yep, people buy HEDT for multi-threaded performance anyway...

[Kisai]

5 hours ago, descendency said:

I'm trying to imagine a non-server workload that would benefit from a 2Ghz 64 core machine...

 

I really doubt AMD brings that out for the reason that the market just has no reason to buy a $3000-4500 CPU that has such a niche use case.

Realtime things are probably where you'd go with that. At 64 cores you're hitting the entry level GPU parallelization that could be achieved with openCL already. Intel's Phi chip had 72 cores. Intel's Larabee had 10 cores but they wanted to get as many as 64 cores. Imagine doing realtime ray-tracing on the CPU itself. Or realtime AI stuff.

 

So the -end game- here might be to merge the CPU and GPU core capability. Where cores can be used to do CPU or GPU type of operations dynamically. 

 

Most end-users do not benefit from high-core counts because most developers of business software and games are still building software like it was 1979. One thread to rule them all.

[Unixsystem]

 

 

8 hours ago, GOTSpectrum said:

People upgrade desktops more often, I wouldn't be surprised to see up to 48 cores this generation and 64 next myself. But then and is really in the bury intel while they can to prove a point mode of 64 core TR is also highly possible

This has been my assumption. While I don't doubt that they could conceivably push a 64 core this year/early 2020, I don't see the point. 48 cores with Zen 2's IPC bump and the removal of the older 32 core's weird memory distribution is already a massive upgrade over any HEDT system out there from AMD or Intel. 

 

Then they can just wait for Intel to respond (assuming they can get anything remotely close to competitive to those core counts within the next few years) and drop the 64, possibly with Zen 3 improvements to boot. Not unlike when Intel unveiled their chilled 28 core and AMD announced the 32 core the next day. 

 

While mainstream is much more neck and neck between the two, AMD has such a headstart in HEDT over Intel that I don't see any financial reason for them to overplay their hand this early. 

 

Spoiler

And yes, I know that Intel still has the clocks and AVX-512 to their benefit, but with 48 cores at equal/better IPC there are increasing few workloads that make sense, especially on an HEDT platform where n-core multithreaded software is more likely to be used. Until Intel gets their node shrinks in order, I don't see any way that they could compete period, let alone when considering price/performance. 

 

[leadeater]

6 hours ago, Jurrunio said:

If it's a consumer platform, it will be overclocked. If it requires the boards to push that current at core voltage, all of the existing boards are screwed. Even the flagships.

Not all of them and it's not even the VRMs that are the problem but the cooling on them. Even if you had full voltage regulation on the die and provided 2V from the VRM to the CPU the reduction in current wouldn't actually be enough to reduce the heat enough to matter. Any more than 2V output from the VRM and the CPU would have to have significantly large VRMs in the die producing a ton of heat. You're wasting die space and killing clocks for a motherboard cooling issue, not that good of an idea.

 

The current boards that really failed for the 2990WX did so in overclocking and mostly only when pushing 1.4V.

[Ithanul]

8 hours ago, VegetableStu said:

quick update: probably temper expectations on TR3K's singlecore boost ._. 4.5 (and lower) seems to be the most reasonable target.

 

EDIT: i mean 4.2 sounds reasonable already, so... o_o

I be happy enough if I could get solid 4.2GHz across all cores.  Though....I fear the wattage draw a tad.  1950Xs when OCed can eat a good bit already.  Though, I would love improved IPC and stable 4.2GHz (plus, better RAM and latency too, I really hope the next TRs improve in that area).

 

Also, I'm hoping some better motherboards drop when the next threadrippers drop.

[leadeater]

13 minutes ago, Ithanul said:

I be happy enough if I could get solid 4.2GHz across all cores.  Though....I fear the wattage draw a tad.  1950Xs when OCed can eat a good bit already.  Though, I would love improved IPC and stable 4.2GHz (plus, better RAM and latency too, I really hope the next TRs improve in that area).

 

Also, I'm hoping some better motherboards drop when the next threadrippers drop.

Currently it looks like Zen2 cores use about 9.3W to do 4.1Ghz so some crude math and assumptions get total power for 64 cores around 595W or for 32 cores 297W. It looks like Zen2 cores use about 15.8W at 4.4Ghz, I'd say overclocking a 64 core TR if one were to exist is likely not a good idea, or an amazing idea XOC LN2 

[Ithanul]

19 minutes ago, leadeater said:

Currently it looks like Zen2 cores use about 9.3W to do 4.1Ghz so some crude math and assumptions get total power for 64 cores around 595W or for 32 cores 297W. It looks like Zen2 cores use about 15.8W at 4.4Ghz, I'd say overclocking a 64 core TR if one were to exist is likely not a good idea, or an amazing idea XOC LN2 

I sure hope the wattage usage goes down.  Be nice if a 32 core part eats around or under 300W.  Personally, I probably stop at a 32 core/ 64 thread.  The cost of RAM starts getting nuts to feed more cores, especially when running BOINC WUs that use VMs or running several VMs as well.

 

On cooling, yeah, a 64 core would probably be pushing serious amount of heat if pushing some high clocks (be wild though to see folks push those). 

[bcredeur97]

23 hours ago, Jurrunio said:

I'm just hope they maintain compatibility with older boards (I dont care if there's no more PCIe 4) and use an integrated voltage regulator if they do release a 64 core. The reason for the latter being the absurd current the boards need to push without it.

wont it be identical to whatever is the epyc 64 core chips now, just higher clockspeeds? Do the epyc (rome) chips have integrated voltage regulation?

[williamcll]

Can you overclock an EPYC processor?

[Jurrunio]

4 hours ago, leadeater said:

Not all of them and it's not even the VRMs that are the problem but the cooling on them. Even if you had full voltage regulation on the die and provided 2V from the VRM to the CPU the reduction in current wouldn't actually be enough to reduce the heat enough to matter. Any more than 2V output from the VRM and the CPU would have to have significantly large VRMs in the die producing a ton of heat. You're wasting die space and killing clocks for a motherboard cooling issue, not that good of an idea.

 

The current boards that really failed for the 2990WX did so in overclocking and mostly only when pushing 1.4V.

Problem with cooling an overclocked a 64 core Zen 2 TR wont be the total heat output from the CPU, but the high heat density caused by pushing high frequency in 7nm process (i.e. heat not leaving the CCDs to the IHS, already seen on even a lowly 3600 in single thread workloads). I really doubt adding an integrated (I still expect SOC to be powered directly regardless, it draws so little power to be worth making an extra voltage regulator in the CPU for it) voltage regulator for Vcore will give out enough heat to hurt overclocking under liquid.

 

Of course it will only be a problem when overclocked, the thing for "running stock" means the worst boards can run it without blowing up. that's how the 2990WX end up running about 3GHz stock even though the silicon easily reaches 4GHz.

 

And from how you put it, I hope you dont think "this board has 480A worth of powerstages" = "I can push 480A through it into the CPU and it's fine". 2700X in a more reasonable 4GHz 1.3V overclock can already sustain 150w of power draw from the EPS 8pin in say, Blender rendering. 4 times that is 600W. Let's say the SOC pulls 100w max (way too high if you ask me, 4 times the PCIe lane count and double the memory channel count shouldnt lead to 4x the power draw), that's still 500w (btw I havent compensated for the VRM efficiency in 2700X power draw, so I'm estimating 2990WX to do 500W at 90% VRM efficiency which I assure you wont be the case for most boards) at 1.3V to the CPU cores. Let's say the Asus Zenith Extreme board with 8 IR3555 60A powerstages. Expect at least 7.5w per powerstage (60w total) to come out from its VRM (actually the heat loss according to the datasheet at 25C temperature which I assure you it's not maintaining). You can also find the same VRM on Asrock boards that's not the Phantom 6 and Asus X399-A, X399-E. What they all share in common is how the heatsink has no hope cooling a GTX 750ti worth of heat, not even if you blast it with air. MSI's 10 IR3555 setup gives them a better chance, but at 38.5A per powerstage it will still kick out 4.6W per powerstage. 46W in total. This you might be able to cool by blasting a case fan right on top of the heatsink (more exposed fins here too), it will depend on the actual environment. Oh and Gigabyte? the 8x 50A powerstage boards are just screwed, completely. the 10x IR3578 50A doesnt have public datasheet but I dont see any indication that it will do better than the MSI 10x IR3555 setup.

 

Basically, you're wrong. Most boards cannot afford to overclock the 2990WX with any overvolting. They have to keep voltages the same as stock unless you strap a waterblock to them.

 

If the new 64 core behaved like 8 3700x (100w stock down the EPS 8pin, I assume you dont overclock 3rd gen from the core) glued together and I assume you just aim to match the 8 core, nothing more, then that's 800W. Let's say the core only contribute to 700W and the voltage is estimated to be at 1.2V (that's how much the 3700x uses at stock, full load), that's 583A of current draw. The good thing of this number is that we do already have boards that can handle this, the Asus Zenith Extreme Alpha and MSI Creation (with 16x IR3555 60A and 16x TDA21472 70A respectively). The bad thing is... LN2 overclocking could be a problem. Tho AMD's releasing new chipsets so LN2 battle should move over to the new platform anyways.

[leadeater]

1 hour ago, Jurrunio said:

Problem with cooling an overclocked a 64 core Zen 2 TR wont be the total heat output from the CPU, but the high heat density caused by pushing high frequency in 7nm process (i.e. heat not leaving the CCDs to the IHS, already seen on even a lowly 3600 in single thread workloads). I really doubt adding an integrated (I still expect SOC to be powered directly regardless, it draws so little power to be worth making an extra voltage regulator in the CPU for it) voltage regulator for Vcore will give out enough heat to hurt overclocking under liquid.

Getting the heat out of the CCD and previous Zen CPU dies I suspect has something to do with the already included voltage regulation that exists in Zen. I don't think as it is it is adding heat but the overall heat is impacting this section of the die. I honestly have no more insight than anyone else does but since it covers 45% of each CPU core and slightly the L3 cache I'd be willing to bet it does play a part. There is something common to the architecture and silicon design itself that gives Ryzen/TR CPUs the same clock, voltage and power behavior even across different nodes and foundries.

 

Quote

AMD uses a Metal-Insulator-Metal Capacitor (MIMCap) layer between the two upper level metal layers for fast current injection in order to mitigate voltage droop. AMD stated that it covers roughly 45% of the core and a slightly smaller coverage of the L3. In addition to the LDO circuit integrated for each core is a low-latency power supply droop detector that can trigger the digital LDOs to turn on more drivers to counter droops.

 

Quote

In Zen, SoC voltage coming from the Voltage Regulator Module (VRM) is fed to the RVDD, a package metal plane that distributes the highest VID request from all cores. In Zen, each core has a digital LDO regulator (low-dropout) and a digital frequency synthesizer (DFS) to vary frequency and voltage across power states on individual core basis. The LDO regulates RVDD for each power domain and create an optimal VDD per core using a system of sensors they've embedded across the entire chip; this is in addition to other properties such as countermeasures against droop. This is in contrast to some alternative solutions by Intel which attempted to integrated the voltage regulator (FIVR) on die in Haswell (and consequently removing it in Skylake due to a number of thermal restrictions it created). Zen's new voltage control is an attempt at a much finer power tuning on a per core level based on a collection of information it has on that core and overall chip.

*my own bolding

 

 

Quote

AMD uses a Metal-Insulator-Metal Capacitor (MIMCap) layer between the two upper level metal layers for fast current injection in order to mitigate voltage droop. AMD stated that it covers roughly 45% of the core and a slightly smaller coverage of the L3. In addition to the LDO circuit integrated for each core is a low-latency power supply droop detector that can trigger the digital LDOs to turn on more drivers to counter droops.

 

Quote

A larger number of sensors across the entire die are used to measure many of the CPU states including frequency, voltage, power, and temperature. The data is in turn used for workload characterization, adaptive voltage, frequency tuning, and dynamic clocking. Adaptive voltage and frequency scaling (AVFS), an on-die closed-loop system that adjusts the voltage in real time following real-time measurements based on sensory data collected. This is part of AMD's "Precision Boost" technology offering high granularity of 25 MHz clock increments.

https://en.wikichip.org/wiki/amd/microarchitectures/zen

 

Whether you think the above is enough on die regulation or not adding more of it isn't going to change the current going through the board VRMs enough to matter, not for 64 cores and workstation clocks, let alone overclocking it.

 

1 hour ago, Jurrunio said:

Basically, you're wrong. Most boards cannot afford to overclock the 2990WX with any overvolting. They have to keep voltages the same as stock unless you strap a waterblock to them.

 

If you have no airflow a fair amount will over heat at 1.25V and at 1.4V many will. If you have some case airflow in the VRM area then the situation improves a lot. Sure better VRM designs will do a lot more but the current problem on existing boards that's really killing them is the junk slab of metal heatsinks and not actually using proper ones. Like X299 you can't solve everything with a VRM redesign, you actually have to put a proper heatsink on. I'm sure you are going to point to the Auros Extreme as that does have a proper heatsink but also has a particularly bad VRM, there's no single fix all solution. If MSI had paired the Creation with a heatsink like that on the Auros Extreme that would be a board today that could do 64 cores and OC, but they didn't so it can't. A water block for the VRMs isn't required, just a decent heatsink for the non garbage VRM boards.

 

Existing X399 boards have been used to OC current TR CPUs up to 500W and they didn't melt in to the ground, not saying most can but boards exist today that will do ok for Zen2 TR, up to 32 cores. I also don't think 64 core TR is coming out soon anyway.

 

1 hour ago, Jurrunio said:

If the new 64 core behaved like 8 3700x (100w stock down the EPS 8pin, I assume you dont overclock 3rd gen from the core) glued together and I assume you just aim to match the 8 core, nothing more, then that's 800W. Let's say the core only contribute to 700W and the voltage is estimated to be at 1.2V (that's how much the 3700x uses at stock, full load), that's 583A of current draw.

We have a slightly more accurate way of knowing how much power Zen2 TR might use, Anandtech has already supplied the watts required at the core level across given clocks. Look for my earlier post about it. Zen2 TR isn't going to come in clocks as high as Ryzen, kind of sadly.

https://www.anandtech.com/show/14605/the-and-ryzen-3700x-3900x-review-raising-the-bar/19

 

The reason why I'm not worried about current boards is I'm very doubtful 64 core TR is coming any time soon and by the time it does a motherboard upgrade isn't too much to ask for anyone that went in on TR first or second generation. Anyone that expected a motherboard to realistically handle an increase from an original 16 core design parameter all the way up to 64 cores is a tad unreasonable.

 

With what I expect a 32 core Zen2 TR to use, 300W, you've got 200W room to OC and admittedly that isn't a lot. I don't see the point though other than sticking all cores to a more fixed near stock multiplier, Zen scales so badly as it is getting note worthy clock increases requires sub ambient cooling.

 

TL;DR I was purposely discounting 64 core TR becuase I'm highly doubtful it's coming in a time frame that matters. Edit: I know the conversation was around the 64 cores so my bad for not stating this in the first place.

[Jurrunio]

55 minutes ago, leadeater said:

Quote

AMD uses a Metal-Insulator-Metal Capacitor (MIMCap) layer between the two upper level metal layers for fast current injection in order to mitigate voltage droop. AMD stated that it covers roughly 45% of the core and a slightly smaller coverage of the L3. In addition to the LDO circuit integrated for each core is a low-latency power supply droop detector that can trigger the digital LDOs to turn on more drivers to counter droops.

 

I still see people use loadline calibration on Zen 2 tho (by manually setting the multiplier, not that I recommend doing so)... not sure how effective this is or how much worse it will get if these things don't work anymore.

 

59 minutes ago, leadeater said:

Quote

In Zen, SoC voltage coming from the Voltage Regulator Module (VRM) is fed to the RVDD, a package metal plane that distributes the highest VID request from all cores. In Zen, each core has a digital LDO regulator (low-dropout) and a digital frequency synthesizer (DFS) to vary frequency and voltage across power states on individual core basis. The LDO regulates RVDD for each power domain and create an optimal VDD per core using a system of sensors they've embedded across the entire chip; this is in addition to other properties such as countermeasures against droop. This is in contrast to some alternative solutions by Intel which attempted to integrated the voltage regulator (FIVR) on die in Haswell (and consequently removing it in Skylake due to a number of thermal restrictions it created). Zen's new voltage control is an attempt at a much finer power tuning on a per core level based on a collection of information it has on that core and overall chip.

*my own bolding

I guess Haswell got a FIVR simply because Intel developed it for Haswell-E. Haswell never drew enough current to really need one. If it is indeed a thermal problem, we won't see it on Skylake-X whose thermal problems are much, much worse than Haswell (even early Haswell like the 4770k).

 

1 hour ago, leadeater said:

Whether you think the above is enough on die regulation or not adding more of it isn't going to change the current going through the board VRMs enough to matter, not for 64 cores and workstation clocks, let alone overclocking it.

Aye, it might bear the name "voltage regulator", I'm only focused on the part on allowing the VRM to pump in higher voltage and less current for the same power input to the CPU and reduce stress on the VRM. We can talk about stuff like ripple suppression and transient response later, after the weaker boards stop melting down their capacitors.

 

1 hour ago, leadeater said:

not for 64 cores and workstation clocks, let alone overclocking it.

btw might want to swap these two's position. OC is more stressful so it should go in the front.

 

1 hour ago, leadeater said:

 

I strongly suspect the Zenith Extreme's result to be wrong. It contradicted Kitguru's test https://www.kitguru.net/components/leo-waldock/gigabyte-x399-aorus-xtreme-review-w-amd-2990wx/6/ which also includes the Gigabyte Aorus Extreme (which didnt hold up well either). HUB did a video about the Zenith Extreme Alpha later, and the gains are surprisingly minimal considering it carried twice the VRM, twice the heatsink and twice the fans as the older Zenith.

 

1 hour ago, leadeater said:

If you have no airflow a fair amount will over heat at 1.25V and at 1.4V many will. If you have some case airflow in the VRM area then the situation improves a lot. Sure better VRM designs will do a lot more but the current problem on existing boards that's really killing them is the junk slab of metal heatsinks and not actually using proper ones. Like X299 you can't solve everything with a VRM redesign, you actually have to put a proper heatsink on. I'm sure you are going to point to the Auros Extreme as that does have a proper heatsink but also has a particularly bad VRM, there's no single fix all solution. If MSI had paired the Creation with a heatsink like that on the Auros Extreme that would be a board today that could do 64 cores and OC, but they didn't so it can't. A water block for the VRMs isn't required, just a decent heatsink for the non garbage VRM boards.

You are correct, but dont forget a good heatsink wont make up for running powerstages near their max ratings. There's only so much heat a heatsink that doesnt get in the way of top case fans or air coolers can get rid of from powerstages packed into an area this small.

 

1 hour ago, leadeater said:

Existing X399 boards have been used to OC current TR CPUs up to 500W and they didn't melt in to the ground,

I mean, you can buy the Creation and the Alpha already so I can't talk back to that...

 

1 hour ago, leadeater said:

Anyone that expected a motherboard to realistically handle an increase from an original 16 core design parameter all the way up to 64 cores is a tad unreasonable.

I'm not giving up on anything that can make it possible  I understand why AMD doesnt do it (cost, Intel could have a lot of patent on FIVR's design that AMD cant work around), but I want them to do it regardless.

 

1 hour ago, leadeater said:

With what I expect a 32 core Zen2 TR to use, 300W,

I think the 32 core will stick to the current 32 core's 250w, if not lower. Zen 2 has already shown about 30% lower power draw and more performance on AM4 so no reason to go up, especially when the prime enemy, Intel bakery's W-3175X sits at 255w TDP (forget whether it's true or not for now, we're talking marketing).