One of the important press releases that came out as a result of the AMD Financial Analyst Day has been some insights into how AMD is approaching the Carrizo and Carrizo-L platform. Have a read of Ryan’s round up of the Financial Analyst Day, which included some broad details about Zen and the big x86 cores, but Carrizo and Carrizo-L focus on AMD’s mobile strategy as well as describing the next iterations of the Bulldozer architecture (Excavator) and the Cat family of low power SoCs (Puma+). We covered some of AMD’s releases on Carrizo back in February, but despite the similar name Carrizo-L functions for a slightly different market by virtue of the different architecture.

Carrizo-L features ‘Puma+’, which by virtue of the naming scheme suggests an updated version of Puma which was seen in Beema. What the ‘plus’ part of the name means has not been disclosed, as both Puma and Puma+ are reported to be 28nm, but chances are that the design has attacked the low hanging fruit in the processor design, rather than purely just a frequency bump. Carrizo-L will be advertised under the new ‘AMD 7000 Series’ APUs, featuring up to four low power separate cores up to 2.5GHz, up to 25W and up to DDR3-1866 support. These are aimed square at the Atom ecosystem within a similar power budget.

AMD Carrizo-L
  A8-7410 A6-7310 A4-7210 E2-7110 E1-7010
Cores / Threads 4 / 4 4 / 4 4 / 4 4 / 4 2 / 2
CPU Frequency Up to 2.5 GHz Up to 2.4 GHz Up to 2.2 GHz Up to 1.8 GHz Up to 1.5 GHz
TDP 12-25W 12-25W 12-25W 12-15W 10W
L2 Cache 2MB 2MB 2MB 2MB 1MB
DRAM Frequency DDR3L-1866 DDR3L-1600 DDR3L-1600 DDR3L-1600 DDR3L-1333
Radeon Graphics R5 R4 R3 'Radeon' 'Radeon'
Streaming Processors 128 ? 128 ? 128 ? 128 ? 128 ?
GPU Frequency Unknown Unknown Unknown Unknown Unknown

AMD is stating that these APUs are currently available in Greater China already with a global rollout commencing in due course. All APUs are listed with AMD Radeon graphics, although the Rx number has no indication as to the streaming processors in the graphics part – a similar situation happened with Beema, and all those parts came with 128 SPs, differing only in frequency which is likely the case here. The SoC design also ensures all the IO is onboard, including an AMD Secure Processor, which for Puma was a Cortex-A5 supporting ARM TrustZone. It is likely that Carrizo-L also uses only a single memory channel, similar to Beema.

One of the more interesting elements is that Carrizo and Carrizo-L will share a socket, known as FP4. This means the processors are pin compatible, and what we know about Carrizo at this point suggests that both segments will play within the same sort of power budget (10-25W vs 15-35W). This allows OEMs to build two designs with almost identical hardware under the hood except for the SoC – would you prefer a single/dual Excavator design, or a faster frequency quad-core Puma+ design? There also leaves scope for differential integrated graphics performance, as mobile Kaveri up to 25W had up to 384 SPs or 3x what we are expecting with Carrizo-L. A lot of the performance metrics in this part will be down to binning the various designs, which adjusts the cost.

At some point we will source a Carrizo-L low-power notebook in order to test the hardware – it would be an interesting data point to get a corresponding Carrizo design as well.

Source: AMD

POST A COMMENT

50 Comments

View All Comments

  • BlueBlazer - Wednesday, May 13, 2015 - link

    From http://www.anandtech.com/show/7314/intel-baytrail-... mentions "The power consumption, at least on the CPU side, also looks very good. From our SoC measurements it looks like Bay Trail’s power consumption under heavy CPU load ranges from 1W - 2.5W, putting it on par with other mobile SoCs that we’ve done power measurements on.". Thus that is definitely around 2W as mentioned earlier. AMD's Carrizo-L has much higher wattage of around 10W to 25W thus AMD's Beema TDP range (although Mullins and Beema uses the same chip). There are no AMD Mullins APUs with 2W SDP. The lowest is 2.8W SDP as shown here http://www.anandtech.com/show/7974/amd-beema-mulli... for both the quad core and dual core versions. Despite that "2.8W SDP", so far all the tablets using AMD's Mullins APU are at least 10 inches and bigger (including the AMD Discovery http://www.anandtech.com/show/7974/amd-beema-mulli... at 11.6 inch, and Bungbungame Photon 2 http://www.bungbungame.com/EN/products/products.as... at 10.1 inch). Reply
  • ToTTenTranz - Tuesday, May 12, 2015 - link

    2 GCN CUs seems really anemic, even if it was for a smartphone SoC.
    10W puts it into large tablet territory, but again with 2CUs, they're either clocked at >1GHz or the thing may not even compete with Cherry Trail in GPU power.

    If neither Kabini in 2013 and Beema in 2014 couldn't get any traction. why bother launching what it seem to be the same chip with a couple of non-substantial tweaks?
    So what's the point? Is this just to keep investors "happy" claiming they're launchign "new chips" every year?
    Reply
  • testbug00 - Tuesday, May 12, 2015 - link

    2 GCN CUs is WAY more graphics power than Atom has. Atom's have 4 EU iirc. 8 GCN units beats 20 GCN... 2 then should beat 5... So, this is 25%+ faster than the 4 EU Atoms. Reply
  • mczak - Tuesday, May 12, 2015 - link

    BayTrail Atoms had 4 EUs. And this would be way more than 25% faster than those Atoms. However, CherryTrail Atoms have 16 EUs (albeit at a lower clock), which might pose quite a challenge (in practice the CherryTrail gpu seems to be easily twice as fast as the one from BayTrail). I suspect they could be close in the end, though probably only because they won't quite compete in the same TDP class (BayTrail atoms even in desktop format didn't go past 10W, whereas this is 12-25W). Reply
  • MrSpadge - Tuesday, May 12, 2015 - link

    Cherry Trail also has the advantage of dual channel memory. It could be that Carrizo-L gets it as well (due to sharing the socket with Carrizo), but considering AMDs recent track record of (almost) not changing anything I wouldn't get my hopes up. Reply
  • testbug00 - Tuesday, May 12, 2015 - link

    Ah, though they stayed at 4! Oops. Thanks.
    Yeah, Cherry should be faster, but, at the power it would be faster, well... There are better options for iGPU from either vendor at that point.
    Reply
  • azazel1024 - Tuesday, May 12, 2015 - link

    But...but...these aren't in Atom's class! Even if the performance is similar or even a little better, these are 10w on the lowest end and 25w on the high end. Atom is a 2w SDP processor, maybe clocking 4w TDP at most. Even the notebook/desktop chips are only about 11w TDP and that is being generous. You are talking double to quadruple the TDP.

    The 10w dual core Puma+ might be honest, but at best the top end quad core Intel Atom chip is maybe hitting the 11w TDP, the dual core Celeron Atom parts are almost certainly hitting a more realistic 6-7w at most.

    These are not targetting Atom at all. They are maybe targetting Broadwell ULT, but I suspect not really on performance.
    Reply
  • azazel1024 - Tuesday, May 12, 2015 - link

    To further my thought, last I checked, the Puma design is not significantly better IPC than Silvermont cores...and the dual core Celeron desktop/notebook designs are hitting 40+% higher clock speeds than the Puma+ dual core part...at a realistic 30-40% less power consumption. Reply
  • testbug00 - Tuesday, May 12, 2015 - link

    Hard to judge AMD's products until we can get power consumption numbers. At least, for the cat cores. Brazos at "18W" ran rings around a "9W" Atoms and used less power.

    Of course, that was the old atom and the bobcat core, but, the point is, need to see some systems with these in them first!

    I expect Intel to have lower power, higher pref/watt and AMD have higher absolute performance. What that will lead to in realistic scenarios where you are racing to sleep depends on the near-idle power usage.
    Reply
  • movax2 - Tuesday, May 12, 2015 - link

    Agree. AMD makes the best from 28 nm process. Intel has an advantage on 'TDP vs Performance' here with its 14-22 nm nodes. Reply

Log in

Don't have an account? Sign up now