Broadwell E5 Xeon prediction thread

[cbn]

This may be a little premature considering we won't know specs till later in the year (I am thinking around the time of IDF 2015?), but if you have a prediction on core count and die size configurations please post it in this thread.

The current Haswell E5s come in three die configuartions: 8C, 12C, 18C. Ivy Bridge came in 6C, 8C, 12C. I believe Sandy Bridge E5 only came in two die configurations: 4C and 8C

If you want to make a prediction on the maximum delta between base clock and 1C turbo please do so as well (Currently the largest delta between base clock and 1C turbo is 1.3 Ghz found on the eighteen core E5-2699 v3). Previous to that Ivy bridge had a maximum 1C turbo to base clock delta of 800 Mhz (on several SKUs)

 

[ArizonaSteve]

Wouldn't be surprised if they didn't show up at all, or had a lower maximum clock speed than the Haswell Xeons. Seems that all the 14nm process is good for is mobile junk.

Has Intel even got a quad core working on this process yet?

 

[cbn]

Has Intel even got a quad core working on this process yet?

Broadwell K (for a LGA 1150) is a quad core:

http://blogs.intel.com/technology/2015/03/gdc-2015/

 

[Enigmoid]

Wouldn't be surprised if they didn't show up at all, or had a lower maximum clock speed than the Haswell Xeons. Seems that all the 14nm process is good for is mobile junk.

Has Intel even got a quad core working on this process yet?

Xeon D

Though that is low power.

 

[gbeirn]

Yeah between Xeon D and Core M it seems like broad well is really only targeting lower power devices, I wouldn't be surprised it it skips the regular Xeon line altogether and we see proper Xeons with Intels next generation 14nm chips.

 

[cbn]

According to following article from May 2014 there was talk of unlocked Haswell Xeons, so Intel may have a back up just in case 14nm yields are not good enough for a large die Broadwell E5 Xeon:

http://vr-zone.com/articles/computex-will-show-desktop-alive-well/77282.html

Add to this the rumoured confirmation that, unlike their predecessors, Haswell-EP Xeons, including likely thae 14-core and 18-core flavours, will have several top bin un-locked and even liquid-cooling optimised variants meant for HPC, workstations and high frequency trading, and you can guess the implications: the Haswell-E and Haswell-EP platforms will again be the overclocker’s heaven.

In general, these are very good news as, with Haswell and Broadwell next-gen high end platforms, we will get the unlocking and speeding-up capabilities we saw in the high end desktops of the past, this time spread across both single socket and dual socket desktops and workstations, not to mention HPC supercomputing platforms.

However unlocked or high power optimized chips (on 22nm) would only address part of the market for these chips.

 

[IntelUser2000]

Wouldn't be surprised if they didn't show up at all, or had a lower maximum clock speed than the Haswell Xeons. Seems that all the 14nm process is good for is mobile junk.

Has Intel even got a quad core working on this process yet?

Actually, mobile optimization is perfect for the Xeons. Lower power cores = More cores.

The low power focus is good everywhere except on Desktops.

Broadwell-EP is already planned anyway: http://vr-zone.com/articles/broadwe...s-reliability-per-core-performance/66845.html

I assume though since the top core count for Haswell EP is 18, even though pre-release set it up as 14 cores, probably just to throw off competitors, I am expecting Broadwell-EP to have 24 cores. They've done the "trick" several times already.

 

[ShintaiDK]

According to following article from May 2014 there was talk of unlocked Haswell Xeons, so Intel may have a back up just in case 14nm yields are not good enough for a large die Broadwell E5 Xeon:

http://vr-zone.com/articles/computex-will-show-desktop-alive-well/77282.html





However unlocked or high power optimized chips (on 22nm) would only address part of the market for these chips.

More like vr-zone dumping its realibility and wanting part of the gossip revenue.

 

[NTMBK]

Nice low power Broadwell cores should be great for servers. Lower power consumption = moar cores.

My only concern is whether the bidirectional ring will scale to >20 cores. The 18 core Haswell-EX already uses three separate bidirectional rings, with some pretty crazy topology. My guess is that BDW-EP will be the last huzzah of the ring, and we will see a mesh based topology in SKX.

 

[Idontcare]

Nice low power Broadwell cores should be great for servers. Lower power consumption = moar cores.

My only concern is whether the bidirectional ring will scale to >20 cores. The 18 core Haswell-EX already uses three separate bidirectional rings, with some pretty crazy topology. My guess is that BDW-EP will be the last huzzah of the ring, and we will see a mesh based topology in SKX.

Mesh based? I expect a lot more from Intel than that.

Quantum entanglement lends itself directly to the application of cache coherency in geographically disparate replication domains.

No kidding, but SKX maybe a gen or two too soon for that. Entropy, err, I mean time will tell

 

[NTMBK]

Mesh based? I expect a lot more from Intel than that.

Quantum entanglement lends itself directly to the application of cache coherency in geographically disparate replication domains.

No kidding, but SKX maybe a gen or two too soon for that. Entropy, err, I mean time will tell

Ah, they're saving that for Icelake :biggrin:

(I guess mesh, because that's what they're using on Knight's Landing- clusters of 2 cores + L2, connected by a 2D mesh. Replace a two core cluster with a single Cannonlake core, and you've got a nice design for high core count 10nm Xeons.)

 

[cbn]

I am expecting Broadwell-EP to have 24 cores.

That is what I am thinking the largest die will be. This followed by 16 core and 10 core dies (with Broadwell i7s for X99 remaining 6 cores and 8 cores). With that mentioned, I do wonder if there will be a delay in the release of the chips? Maybe sometime in Q1 2016 if there is a delay?

 

[cbn]

Has Intel even got a quad core working on this process yet?

Broadwell K (for a LGA 1150) is a quad core:

http://blogs.intel.com/technology/2015/03/gdc-2015/

Some additional info from VR Zone on Broadwell Desktop Core i5 and i7 desktop SKUs:

https://translate.google.com/transl...75c-and-i5-5675c-03232015/&edit-text=&act=url

Two SKUs for LGA 1150: Core i7-5775c and Core i5-5675c

Both come with 65 watt TDP, Iris Pro 6200 graphics and DDR3L 1600 memory, but the i7 has a 4C/8T core configuration and 6MB cache with 3.3 Base clock and 3.7 Ghz turbo. The Core i5 has 4C/4T with 4MB cache and 3.1 Ghz base/3.6 Ghz turbo.

Those clocks aren't that high, but at the same time the TDP is only 65 watts.

 

[cbn]

Regarding Broadwell E5 Xeon at 24 cores, if that happens base clocks would remain at 2.3 GHz (same as E5-2699 v3 on 22nm) provided we see a 33% reduction in power at 14nm to match the 33% increase in cores.

If the 14nm node cannot deliver a 33% reduction in power then I'd imagine we will see base clocks on the top SKU drop to some level below 2.3 Ghz.

P.S. Looking ahead to Skylake E5 (also on 14nm) I think base clocks will drop (in the same way we saw base clocks drop moving from the 22nm 12C Ivy E5-2697 v2 to 22nm 18C E5 2699 v3) if Intel continues to significantly increase core count on the top E5 SKU.

 

[imported_ats]

Regarding Broadwell E5 Xeon at 24 cores, if that happens base clocks would remain at 2.3 GHz (same as E5-2699 v3 on 22nm) provided we see a 33% reduction in power at 14nm to match the 33% increase in cores.

If the 14nm node cannot deliver a 33% reduction in power then I'd imagine we will see base clocks on the top SKU drop to some level below 2.3 Ghz.

P.S. Looking ahead to Skylake E5 (also on 14nm) I think base clocks will drop (in the same way we saw base clocks drop moving from the 22nm 12C Ivy E5-2697 v2 to 22nm 18C E5 2699 v3) if Intel continues to significantly increase core count on the top E5 SKU.

We can do some estimation based off over Xeon-D. Dual 10 GbE probably costs in the range of 5-8W of power. Chipset complex probably another 8-12W of power. Memory controllers probably another 5-8W of power. So we're looking at 18-28W of power for the uncore. That leaves 37-47W of power for 8C.

For the 130W TDP processors we have 2 MCs, @ 10-16W total power. Probably roughly another 10-15W of power for the rest of the uncore leaving roughly 99-110W for the cores.

3x37W = ~110W for 24C@2GHz. Though we don't yet know how close to the TDP limits Xeon-Ds actually are, though we do know that normal Xeons are almost always fully thermally constrained.

 

[cbn]

Another thing to think about: What Broadwell E5 will be used in the FB's Open Compute Platform?

http://www.anandtech.com/show/9138/open-compute-hardware-tried-and-tested

Currently they are using the E5-2670 v3 in 2P configuration:

(E5-2670 v3 is a 120 watt 12C/24T with 2.3 Ghz base clock and 3.1 Ghz turbo.)

 

[IntelUser2000]

Regarding Broadwell E5 Xeon at 24 cores, if that happens base clocks would remain at 2.3 GHz (same as E5-2699 v3 on 22nm) provided we see a 33% reduction in power at 14nm to match the 33% increase in cores.

24 cores are very likely considering recent information.

I think we'll see higher clocks along with 24 cores. With Ivy Bridge EP we saw 50% core gains along with 10% clock speed increase on the high-end parts.

With Haswell EP, it bucked the trend of 33% or 50% core increase and went to an 18 core, rather than 16*. But it lost clocks by about 10%, making it effectively like 16 cores while keeping clocks same. Broadwell E5 Xeon should effectively be equal to 50% increase in cores, and because 24 is only 33% increase over 18, I'd expect clock increase of 10-15% to compensate for it which would put it back to the range of Ivy Bridge EP clocks.

Tock=33% because its on the same process but with better IPC
Tick=50% because its on the new process and small IPC gains

*This I assume because Haswell project was having issues, and they failed(though minor) to meet their goals. Remember how throughout the entire Haswell family it lost clocks or kept the same? They knew that early enough that in order for the Xeon E5 v3s not to be a disappointment, more cores were needed.

 

[cbn]

Regarding Broadwell E5 Xeon at 24 cores, if that happens base clocks would remain at 2.3 GHz (same as E5-2699 v3 on 22nm) provided we see a 33% reduction in power at 14nm to match the 33% increase in cores.

^^^^ Regarding the power consumption reduction on the advanced node to match clocks (for 33% increase in cores), it should actually be 25% reduction in power (re: 1 x xtor * 100% power consumption = 1.33 xtor * 75% consumption).

 

[cbn]

I think we'll see higher clocks along with 24 cores.

I think this will happen too. I also expect the die to be rather small as well (re: 14nm is a 1.5 node jump from 22nm while 16nm would have been the typical one node jump)

P.S. For the FB Open Compute Broadwell E5 with my guess is that it will have 16 cores @ 120W TDP with 2.5 Ghz base clock and 3.3 Ghz turbo.

 

[cbn]

Predictions on what the E5 v4 Workstation SKU will have for specs?

Current and past E5 workstation SKU specs:

E5-2687W v3 (160 watt, 10C/20C, 3.1 Ghz base w/ 3.5 Ghz turbo)

E5-2687W v2 (150 watt, 8C/16T, 3.4 Ghz base w/ 4.0 Ghz turbo)

E5-2687W (150 watt, 8C/16T, 3.1 Ghz base w/ 3.8 Ghz turbo)

 

[cbn]

For my Broadwell Workstation SKU prediction (see post above):

E5-2687W v4 (160 watt, 10C/20T, 3.4 Ghz base with 4.0 GHz turbo)