AMD Streamroller with 16 cores and PCIE v3 will be realized

So, week ago we seen Kaveri and this is exactly four times more cores on the chip:


Source: http://support.amd.com/TechDocs/47414_15h_sw_opt_guide.pdf (Figure 14, page 197)

Newer models of Family 15h processors offer five links for connections to I/O and other processors. Of the five links, one link suports PCIe 3.0, two support coherent Hypertransport, and two are capable of either coherent Hypertransport or PCIe 3.0. These processors have 8 compute units (16 cores).

What does that mean? Well, probably some server based CPUs or CPUs for people that need lots of processes? Could be (with the Mantle scaling almost perfectly with more cores) also interesting for gamers, as the power will be significant.

Never the less, the mystery about four memory controller channels, with only two being active on Kaveri, is starting to make more sense. No-one in the right mind would expect that 16 cores could get along with memory access with only two channels. Especially considering that the space for L3 cache is limited - cache eat most of the die size, so roughly when we know, that four Streamroller cores with L2 cache eat abou 62mm square of die size, then 16 cores are somewhere around 245-260mm die size. Of course we have to add memory controler, northbridge, and PCIE lines… so we are getting something around 300mm die size, witch is not uncommon to be realized under 125W TDP.
However with 16MB L3 we getting to about 360mm and with 32MB we getting 430mm die size need, witch is somewhat extreme and over the line.

On the other hand, 16MB L3 is not that much for 16 cores. So it all depends on how adventurous AMD will be. Sometimes they cross the limits (eg. the 5GHz CPU with overhelming TDP even for most watercooling solutions, lol).

Looks interesting!  Though it all depends on how they perform as Kav….is less than impressive.

I'm dying for some CPU power from AMD, they used to be my go to team.

A 16 core with 32 Thread paths would be divine plus some real world power and no longer shared cache.

DDR4 Quad channel would be great as well. NO NO NO onboard graphics just pure CPU power.

Going to get a little FM2+ rig just to support AMD and to support my steambox needs on my tv 🙂

This will be good, only if the scheduler can use them all, and games can be optimized for all these workers. Otherwise its just a number, right now.
Would be interesting to see how single threaded performance improves, as this is a point AMD is sorely lacking in,

-Shiver

Well, I don't think Kaveri can be compared with hi or even higher end line Intel CPU + dedicated graphic. Then it perform relatively mediocre. But the point of APU is lower to middle end machines. For me the Kaveri seems to be good, because I cannot even run BF4, not to mention getting 60fps on high details (tought in simple scene).
IMHO the Kaveri is a good performer in it's class and it definitively save money (and even power and cooling) for dedicated graphic card. Tough we all know, that dedicated card will be always (at least current cards) much, much faster (just because of the dedicated superfast memory), then the point of APU is to save price. You need to cool the CPU - but with quiet requirments at hand, you need also to get 3rd company cooler for GPU. Therefore for low end gaming machine it, IMHO will suffice.

Now we can look at some upcoming Opterons - the Abu Dhabi generation based on Piledriver cored get updated to the Warsaw cores. Sadly, the manufacturing process will stay the same 32nm one, but still they promising lower TDP. 99W 🙂 So more effective cores?

Opteron 6370: 16 cores, 99W TDP, 16MB L3 ($664)
Opteron 6338: 12 cores, 99W TDP, 16MB L3 ($421)

No clocks are revealed yet, but some expecting lower clocks. I did not think that this could be true, but we see. Current Piledriver cores run at 115W TDP and 16 cores go from 2.3-2.5GHz to 3.2-3.4GHz with turbo. 16 core ones got faster base clock of 2.6-2.8GHz that go up to the same 3.2-3.4GHz with turbo.

The future models will be introduced around next half of 2014 with the Streamroller cores and 28nm process.

I have no idea about performance, but I would consider that in HWbot Cinebench benchmark these Opterons win all the top places:
http://hwbot.org/benchmark/cinebench_r11.5/

Also brute-force cracking RAR archive bench seems love Opterons:
http://hwbot.org/benchmark/ucbench_2011/
(tough some guy that go with nick Dead Things are pushing there with the Xeon 8870 to highly respectable 5th place ;D  Don't know that one? 😉 )

  • 5
    Posts
  • 1366
    Views