AMD EPYC Server Processor Thread - EPYC 7000 series specs and performance leaked

[csbin]

https://videocardz.com/70266/amd-epyc-7000-series-specs-and-performance-leaked

AMD EPYC 7000 launches June 20th
The NDA on this material ends on June 20th. Here are some buzzwords from the slides:

• up to 32 High-Performance “Zen” Cores
• 8 DDR4 Channels per CPU
• Up to 2TB Memory per CPU
• 128 PCIe Lanes
• Dedicated Security Subsystem
• Integrated Chipset
• Socket-Compatible with Next Gen EPYC Processors

No compromise 1-Socket

• Right-size underutilized servers
• Optimize storage and heterogeneous compute
• Low power consumption
• Up to 20% lower cap-ex

• 
  

 

[Rngwn]

That 8-core part... That one is likely be 1 core active per CCX. I wonder how much yield of those dies with that bad of a defects.

I expect that AMD will not make this too appealing price-wise to the mass, or they will have to cripple the better dies for it.

 

[cbn]

That 8-core part... That one is likely be 1 core active per CCX. I wonder how much yield of those dies with that bad of a defects.

I expect that AMD will not make this too appealing price-wise to the mass, or they will have to cripple the better dies for it.

Maybe those are sourced from early production wafers.....so by this time AMD has accumulated enough "mostly broken" dies to make a reasonable amount of the 8C/16T octachannel processor?

 

[SarahKerrigan]

That 47% number for the 7601 would appear to put it around 1300-1400 SPECint_rate - above the current SPECint_rate record holder (SPARC M7, 1120 base) but likely below Power9 and potentially Skylake-SP. However, it has more memory bandwidth (8ch vs 6ch) and considerably more PCIe than SKL-SP, and matching memory bandwidth and more PCIe than Power9 (48 lanes PCIe 4 vs 128 lanes PCIe 3.)

I think Epyc is looking very well-positioned, unless we end up with surprises from the fact it has eight different locality domains per socket.

 

[krumme]

Boot up Johan De Gelas !
Time to go aces hardware beserk

 

[raghu78]

AMD is going to be extremely disruptive at these prices. AMD is going to attack Xeon Platinum SKUs with Epyc 24C/48T and 32C/64T but with prices below Xeon Gold range.

 

[stockolicious]

admittedly im not very technically savy but the base clocks seem a little weird on there for EPYC - if a threadripper 16core chip has a 3.4base how can the EPYC 16 core server variant be at 2.4gh? What am i missing?

 

[csbin]

admittedly im not very technically savy but the base clocks seem a little weird on there for EPYC - if a threadripper 16core chip has a 3.4base how can the EPYC 16 core server variant be at 2.4gh? What am i missing?

EPYC: 8 DDR4 Channels，128 PCIe Lanes
Threadripper: 4 DDR4 Channels，64 PCIe Lanes

 

[Atari2600]

Why start the naming series at 7000? If the EPYC name is supposed to replace Opteron, surely you give it as much growth room as possible.... i.e start with EPYC 1000 series.

Maybe expecting some logic in naming conventions is a bit too much?

Or is it a BS leak...

 

[Topweasel]

admittedly im not very technically savy but the base clocks seem a little weird on there for EPYC - if a threadripper 16core chip has a 3.4base how can the EPYC 16 core server variant be at 2.4gh? What am i missing?

4 Ryzen Dies instead of 2. They set the power window within the same as ThreadRipper. This isn't like Intel where two Cores here or there just means 2 cores. All of EPYC is going to be 4 distinct dies with 4 distinct power usages. So while power usage will be down for unused cores overall it will still require more power for the 16 enabled cores (and the 64 MB of L3 vs. 32MB) then Threadripper. That and clocks aren't everything on the servers especially in markets they will be targeting with EPYC.

 

[dooon]

Ctrl + U
https://www.supermicro.com.tw/aplus/

 

[Topweasel]

Why start the naming series at 7000? If the EPYC name is supposed to replace Opteron, surely you give it as much growth room as possible.... i.e start with EPYC 1000 series.

Maybe expecting some logic in naming conventions is a bit too much?

Or is it a BS leak...

Last leak I thought was BS turned out to be Threadripper in the end. That said BS link. The 8c16t part makes no sense. 1 Core per CCX. Using 4 different 2 core full sized dies? Not buying it.

 

[noneis]

admittedly im not very technically savy but the base clocks seem a little weird on there for EPYC - if a threadripper 16core chip has a 3.4base how can the EPYC 16 core server variant be at 2.4gh? What am i missing?

Ryzen uncore cosumes 15-17W with those die-to-die communication links disabled. 8 Cores without uncore at 2,1Ghz consumes ~17W in Prime Small FTT without voltage stability margins and with stability margins roughly 20-22W. According to Stilt, there are voltage regulators. They are in bypass mode on AM4 platform. 4*([15-17]+[20+22]) / 0.9 = 156-173 W + Die-to-Die communication. These clocks and TDPs look believable.

That 8C SKU doesn't look like high margin product. Even R5 1400 would have better margin.

 

[maddie]

Ryzen uncore cosumes 15-17W with those die-to-die communication links disabled. 8 Cores without uncore at 2,1Ghz consumes ~17W in Prime Small FTT without voltage stability margins and with stability margins roughly 20-22W. According to Stilt, there are voltage regulators. They are in bypass mode on AM4 platform. 4*([15-17]+[20+22]) / 0.9 = 156-173 W + Die-to-Die communication. These clocks and TDPs look believable.

That 8C SKU doesn't look like high margin product. Even R5 1400 would have better margin.

The issue is that you can't get a R5 1400 with 1 core/CCX working. These would have to be the most defective die being used.

If info is true, then pretty much every die will be sold.

 

[dnavas]

That 8C SKU doesn't look like high margin product. Even R5 1400 would have better margin.

The pricing on the 16-core variants seems to imply something like $650 / $850 / $1150. I assume TR pricing will be similar. As with the 8-core, the lowest priced 16-core seems really cheap.

 

[Atari2600]

This doesn't make sense to me...

The numbering is all over the place.

Leaving that aside for the moment - why have the 7251 at 120W rather than clock it up to say 3.0 GHz // 3.5 GHz at 170W? (or offer a 7252 that allows a trade of power for clock).

But back to the numbering - what on earth does the 2nd numeral represent?

 

[dnavas]

But back to the numbering - what on earth does the 2nd numeral represent?

I assume the combination of numerals 2&3 are used somewhat arbitrarily to indicate performance. Which is sort of sad, because apart from 7281 and the 7601, those numbers could have indicated core counts.
Now, what is the second number for TDP? How can a chip have two TDPs?

 

[maddie]

This doesn't make sense to me...

The numbering is all over the place.

Leaving that aside for the moment - why have the 7251 at 120W rather than clock it up to say 3.0 GHz // 3.5 GHz at 170W? (or offer a 7252 that allows a trade of power for clock).

But back to the numbering - what on earth does the 2nd numeral represent?

Is it? For the most part, 75xx = 32C, 74xx = 24C and 73xx = 16C. The 7601 and 7281 are the only odd ones.

edit: Seems dnavas has similar thoughts.

 

[nix_zero]

This doesn't make sense to me...

The numbering is all over the place.

Leaving that aside for the moment - why have the 7251 at 120W rather than clock it up to say 3.0 GHz // 3.5 GHz at 170W? (or offer a 7252 that allows a trade of power for clock).

But back to the numbering - what on earth does the 2nd numeral represent?

that 8 core is for datacenters that need lots of pcie/disks with little processing power so an higher tdp wouldnt make sense

the two number probably are configurable tdp (if you read the stilt thread he noted that ryzen has an "hidden" feature to configure max tdp) like in apus.

 

[.vodka]

Found a close up.

 

[krumme]

Found a close up.

Looks like smart business to me

 

[senseamp]

$4000 for AMD CPU? Keep smoking that crack.

 

[LaikaSpaceCat]

$4000 for AMD CPU? Keep smoking that crack.

Because apparently a:
32 core
64 threaded
8 channel
128 PCI-E lane

Server CPU that is more or less on par with Intel on most metrics per core is somehow not worth 4000$.

 

[Space Tyrant]

$4000 for AMD CPU? Keep smoking that crack.

Do you care to explain why that doesn't make sense?

 

[senseamp]

Because apparently a:
32 core
64 threaded
8 channel
128 PCI-E lane

Server CPU that is more or less on par with Intel on most metrics per core is somehow not worth 4000$.

They are pricing it against Intel's monopoly pricing, but they are in a duopoly now. Prices should come down by at least half.