Rumor: AMD Ryzen Threadripper PRO Lineup Leaked

[Raevenlord]

Videocardz seems to have snagged some official AMD slides detailing their upcoming Threadripper PRO lineup. AMD is tiering its Threadripper CPU offerings between the Threadripper and Threadripper PRO via added functionality that AMD considers is better suited to the prospective buyers of a PRO-branded Threadripper: professional studios, designers, engineers and data scientists. AMD's positioning for these creatives or scientists is to offer a much improved platform throughput compared to Threadripper: the PRO version supports up to 128 PCIe 4.0 lanes (64 in non-PRO); up to 2 TB of ECC memory support (either in UDIMM (Unbuffered DIMM), RDIMM (Registered DIMM), LRDIMM (Load-Reduced DIMM) and 3DS (three-dimensional stacking) RDIMM vi an 8-channel configuration (4-channel in non-PRO); as well as professional-oriented tools and features such as Pro Security, Pro manageability, and PRO business ready support.

Four different CPUs will reportedly be offered in the Threadripper PRO lineup: the 64-core 3995WX is a relatively known quantity by now; likewise, the 3975WX will mirror consumer parts core counts (32 cores), both with reduced clocks by 100-200 MHz compared to their non-PRO counterparts. AMD seems to also be launching 12 and 16-core PRO Threadrippers in the form of the 3955WX (16-core) and 3945WX (12-core), both with boost clocks being set to 4.3 GHz.



View at TechPowerUp Main Site

 

[Chaitanya]

Very odd that 24 and 48 core options are missing.

 

[SamuelL]

I'm surprised by these specs, these are essentially Epyc cpu's now. The pricing is going to have to be the significantly higher than the previous Threadrippers to keep this from hurting server sales.

 

[zlobby]

Very odd that 24 and 48 core options are missing.

Also, we are missing info if theybare going to be exclusive to OEM or one could assemble a build themselves. Actually, with the RAM and PSU out there, only the mobos are missing for a DIY build.

 

[TheLostSwede]

TRX80 chipset.

 

[ZoneDymo]

Are we not creating a bit too many categories here?

Consumer - Ryzen
Prosumer - Threadripper
Pro...prosumer? - Threadripper Pro
Professional? - Epyc

 

[ARF]

Are we not creating a bit too many categories here?

Consumer - Ryzen
Prosumer - Threadripper
Pro...prosumer? - Threadripper Pro
Professional? - Epyc

Athlon - entry
Ryzen - mainstream, mid-range
Threadripper - high-end
Threadripper Pro - professional, high-end
EPYC - server, government, serious organisations...

 

[dragontamer5788]

Are we not creating a bit too many categories here?

Consumer - Ryzen
Prosumer - Threadripper
Pro...prosumer? - Threadripper Pro
Professional? - Epyc

EPYC is more about "server" than "professional". EPYC has low clocks for more efficient compute.

Workstations need high clocks, and efficiency isn't a major concern.

 

[ARF]

EPYC is more about "server" than "professional". EPYC has low clocks for more efficient compute.

Workstations need high clocks, and efficiency isn't a major concern.

I think EPYC has lower clocks in order to be able to be cooled "passively" when installed in those rooms with cabinets and racks.

 

[dragontamer5788]

I think EPYC has lower clocks in order to be able to be cooled "passively" when installed in those rooms with cabinets and racks.

Reducing electricity usage causes a further reduction in electricity: every watt of energy saved is additional watts saved off of your air-conditioning bill.

It takes lots of air-conditioning power to move heat around. The easiest way forward is to simply generate less heat.

---------

EDIT: Data-center rooms are often power-limited as well. If you spend 15kW per rack, then 20 racks is already 300kW (220V * 1300 Amps). If each computer is using 1/2 the power or less (say, 5kW per rack), you can cram 2x or 3x the computers into the same space.

 

[thesmokingman]

2Tb memory, finally!

 

[ARF]

Reducing electricity usage causes a further reduction in electricity: every watt of energy saved is additional watts saved off of your air-conditioning bill.

It takes lots of air-conditioning power to move heat around. The easiest way forward is to simply generate less heat.

---------

Data-center rooms are often power-limited as well. If you spend 15kW per rack, then 20 racks is already 300kW (220V * 1300 Amps). If each computer is using 1/2 the power, you can cram 2x the computers into the same space.

Yes, it's interesting if there are, somewhere, data-centre rooms just outside cooled directly by the Earth's atmosphere. Say Norway or southern Chile, Argentina, or somewhere high in Switzerland.
All the electricity cost for ventilation will be virtually zero.

 

[Tomorrow]

Very odd that 24 and 48 core options are missing.

No what's odd is the inclusion of 12 and 16 core models on more expensive WRX80 platform. That was one of my initial TR3000 critisims that they only offered 24 core as bare minimum entry point. Not all TR buyers need that many cores but could sure use the PCIe lanes and quad channel memory. Plus it's a smart move to include a 12 core model to get people into the platform. Later down they line they might upgrade and AMD would have sold two CPU's instead of none (when a person saw the price of the 24 core model and decided to forego the platform alltogether due to high entry price).

Besides it makes no logical sense that a lower cost TRX40 platform lacks lower core count models and yet the higher end WRX80 has them...

 

[zlobby]

I think EPYC has lower clocks in order to be able to be cooled "passively" when installed in those rooms with cabinets and racks.

EPYC has low clocks due to the sheer number of cores it has, plus the hardships of cooling many of these running in racks.

While TR has almost the same number of cores, high clocks are possible because TR are typically placed in workstation boxes where beefy cooling is quite common.

 

[xkm1948]

TR should have come with more DRAM capacity and 8 channel support from the get go. This is creating too many tiers. With this new scheme, nobody would be buying 3990WX. I mean max 256GB DRAM for 128 threads? That is not a good balance for highly parallel workload.

 

[InVasMani]

EPYC has low clocks due to the sheer number of cores it has, plus the hardships of cooling many of these running in racks.

While TR has almost the same number of cores, high clocks are possible because TR are typically placed in workstation boxes where beefy cooling is quite common.

The motherboards aren't really as well orientated for it either and most are dual socketed doubling the cool complications as a whole. I think the main thing with Epyc however is clock are lower in part because it's more optimal for efficiency, but the other part is with so many cores all together it's harder to bin enough of them as well as they can with TR/AM4 by contrast that would drive up both the cost and cost of ownership on the electrical bill exponentially further by comparison. Binning that many high quality chips gets harder as you climb up the ladder rung further to higher core count and frequency.

TR should have come with more DRAM capacity and 8 channel support from the get go. This is creating too many tiers. With this new scheme, nobody would be buying 3990WX. I mean max 256GB DRAM for 128 threads? That is not a good balance for highly parallel workload.

If only they'd had the foresight to predict the most optimal future of their hardware in relation to customers desires. Seriously though think were a platform like this would be if AMD hadn't become competitive again with Ryzen. I mean just think about it even with Ryzen and the increased competition they've brought Intel is struggling rather heavily in these area's comparatively speaking. Intel's workstation boards were pretty laughably bad specs by contrast they stagnated innovation big time in the lull period between Bulldozer and Ryzen. Intel is playing catch up now basically for there own short comings on forward thinking innovation.

 

[Vya Domus]

I think EPYC has lower clocks in order to be able to be cooled "passively" when installed in those rooms with cabinets and racks.

They have lower clocks for power efficiency reasons, basically to minimize the cost of running them. Those "passive" heatsinks are cooled by 10 000 RPM fans, that's not the issue.

 

[wahdangun]

I'm surprised by these specs, these are essentially Epyc cpu's now. The pricing is going to have to be the significantly higher than the previous Threadrippers to keep this from hurting server sales.

but epyc doesn't boost this hight

 

[HwGeek]

Lower Clock speeds are expected when using 8Ch memory at the same TDP limits, those extra 4CH needs more juice ;-).

 

[Chrispy_]

They're seriously eating into their own 1P EPYC market here....

 

[Jism]

I think EPYC has lower clocks in order to be able to be cooled "passively" when installed in those rooms with cabinets and racks.

Eh? I guess you never seen a server from inside did you?

They are not supposed to be cooled passively. Yes they might appear with a heatsink and no fan on top of it. But servers in general have a bunch of high-performance fans that push air through the chassis. All those rooms are kept on usually airconditioned levels to maintain good performance.

Second; theres no limit on how much sound those fans are able to produce either.

To give you an impression on how much just one unit can produce in relation of audio.

 

[Tomorrow]

Plus there are even watercooled servers. So cooling is not an issue. The real reason for lower clocks on server CPU's is performance per watt, voltage efficiency curve and total cost of ownership. Obviously performance is important too but not at the cost of high power requirements.

When a single person runs a CPU that consumes 300W instead of 150W (just an example) that's not that big of a deal. But when you have 10000 of these running this extra 150W would add up very quicly and increase TCO substantially.

 

[player-x]

I'm surprised by these specs, these are essentially Epyc cpu's now. The pricing is going to have to be the significantly higher than the previous Threadrippers to keep this from hurting server sales.

No, servers and workstations are mostly very different type of products, no serious server admin would ever use a workstation board in his servers, server board are designed to run in racks, next to that the EPYC line and boards are certified for their 24/7 jobs.
Next, servers boards are more simple and robust, and have only the most basic so less can go wrong, if a workstation go's down, one user is not working, if a server go's down possible a hole company is not working.

Are the TR Pro gone be more expensive, properly yes, as it's made for a premium niche market with a lower turnover, but I think the CPUs are not gone be that more expensive, as they are just re-pined EPYC CPU's, there is even a good chance that the 12 and maybe even the 16 core is gone be sub $1000 parts, as the same EPYC parts are around $700~$1000 price point.

But the motherboards are gone be a whole other story, I think we have to be real lucky if we can find one for under $600, so I think even the cheapest TR Pro setup is gone run you +$2000 for only CPU and board and 8 sticks of memory.

ASRock, Asus, Gigabyte and MSI, all gone make 1 halo board, and maybe 1 stripped down version (without TB and/or 10Gbit), don't know what Supermicro is gone do, they could decide eider way, but there is a good chance that it will fit in their workstation line, then we will also see loose boards from them.

 

[bogami]

I wonder what the latencies will be ?

 

[Jism]

Plus there are even watercooled servers. So cooling is not an issue. The real reason for lower clocks on server CPU's is performance per watt, voltage efficiency curve and total cost of ownership. Obviously performance is important too but not at the cost of high power requirements.

When a single person runs a CPU that consumes 300W instead of 150W (just an example) that's not that big of a deal. But when you have 10000 of these running this extra 150W would add up very quicly and increase TCO substantially.

You pay for "Amperage" consumption in a rack. So lets say i can hire 1U space and my limitation is 1A for my server. It would mean that my server cannot use more then 1A on 220V or so. A CPU that has a fairly high amount of cores with a reasonable TDP gets me work done in less, sort of say.