David Kanter is very against HBM/HBM2 for CPUs. I wonder if it is due to the fact that HBM/HBM2 deliver very high bandwidth (great for gfx/HPC) but maybe not so great latency?
HBM2 is probably way too expensive to fully replace the DRAM any time soon. Also to justify it´s cost AMD really needs to step-up in EVERY area. If Zen cores and therefore Raven Ridge will perform as I expect them to perform, AMD cannot ask top price for the product regardless if it has huge pile of HBM2 on it or not. I´m not sure if the latency of HBM2 is anywhere near sane levels either.
David Kanter is very against HBM/HBM2 for CPUs. I wonder if it is due to the fact that HBM/HBM2 deliver very high bandwidth (great for gfx/HPC) but maybe not so great latency?
David Kanter is very against HBM/HBM2 for CPUs. I wonder if it is due to the fact that HBM/HBM2 deliver very high bandwidth (great for gfx/HPC) but maybe not so great latency?
The latency is outright horrible. So that's most certainly why. But GFX doesn't care for that, its all about the bandwidth.
But can they fix that fairly easily if they wanted to?
Example: Micron lists a GDDR5 (MT51J256M32HF-60 (GDDR5 which has a 8Gb density)) with a programmable CAS latency of 7 to 27 (click on data sheet when in link below):
https://www.micron.com/products/dram/gddr5/gddr5-part-catalog#/
(I would assume the CL of 7 would be if the GDDR5 were tuned for much lower bandwidth than we normally see it used at)
Carrizo-E/Bristol Ridge-E/Summit Ridge-E will be using a single stack of 256MBytes of eDRAM. While, Raven Ridge-E will be using 2GB?(2-stack 3D)/4GB?(4-stack 2.5D) HBM2.
With HBM? No not really.
CL7 at GDDR5 would mean a very low frequency btw. CL27 is most likely what you see at 1.75Ghz base clock.
The latency is outright horrible. So that's most certainly why. But GFX doesn't care for that, its all about the bandwidth.
What're you basing that on? Joe Macri indicated that HBM1 latency is lower than that of GDDR5:
http://www.legitreviews.com/amd-high-bandwidth-memory-detailed-with-joe-macri_163855
HBM2 should be lower.
HBM also has a lot more channels, good for parallel handling of multiple simultaneous accesses.
FX-8800P is a 15W part with cTDP limit of 35W, where as the FX-9830P is a 35W part with cTDP limit of 45W :sneaky:
133% more power budget is expected to bring some clock improvement
Keep in mind that in case of Carrizo / Bristol Ridge there is no such thing as the "base clock". The clocks can drop below the "base clock" level when and if the power limit is exceeded.
A Quad Core 14nm ZEN with 12 CUs and 8-16GB of HBM2 memory for both system and graphics could be a premium Laptop APU for AMD in 2017 with high margins that could completely transform the laptop landscape in 2017.
At least the good news is that it would be difficult for OEMs to kill performance of FX-9830P in similar fashion as already happened with FX-8800P.
In majority of laptops it was locked at 15W TDP without opportunity to use 25/35W mode.
Not to mention a typically single stick of memory (usually DDR3L-1600).
Btw, don't you think that support of only DDR4-2400 is somehow disappointing?
At this moment 8GB kit of DDR4-3200 cost only 20 dollars more than DDR3L-2133/DDR4-2400.
HBM is DDR based, meaning the clock runs max at 250Mhz.
Btw, don't you think that support of only DDR4-2400 is somehow disappointing?
Never heard of 250 MHz being a max clockspeed for HBM or HBM2. Source? That aside, DDR1 could get you some nice, low latency at clockspeeds of 200 MHz years ago . . . HBM/HBM2 implementations have the benefit of extremely short trace lengths (among other things, cited in the link I pasted above). High latency certainly does not appear to be inevitable for HBM or HBM2.
Clearly you expect Zen 8 core to faster and cheaper than Skylake or Kabylake.Since you don't know what my expectations for Zen are, how do you it will not happen?
Clearly you expect Zen 8 core to faster and cheaper than Skylake or Kabylake.It doesn't matter if the performance comes from higher clocks or IPC. So if it performs as 8 core IB/Haswell-E, then would you be interested in buying it assuming the price is right?
I'm impressed. Then you more know about what I expect than I do myself, because I'm not entirely sure what to expect. I only have a rough guess.
I think it's your memory that is failing you. Or you're misinterpreting conclusions about a potential outcome based on assumptions on this forum with my own expectations.
If you really are interested in my expectations, I'd say something along the lines mentioned above: 8 core IB/HW-E performance. However I'm not certain at all so it's a very rough guess, since so much is unknown. But if that's where it ends up, it should be much faster than 4 core KabyLake in MT performance, and come close in ST performance.
At stock speeds, due to a combo of lower clocks and likely lower IPC, 8 core Zen will be at least 25% off 4 core Skylake/Kabylake ST performance.
I expect Zen to have IPC roughly equivalent to Haswell; lower than SKL/KBL. I also do not expect it to come anywhere close to the 4GHz+ stock speeds that SKL does.
3GHz for 8 cores would be a solid achievement, IMO.
It is far from clear if AMD will even be offering an 8 core Zen for desktop users in 2017.
I expect Zen to have IPC roughly equivalent to Haswell; lower than SKL/KBL. I also do not expect it to come anywhere close to the 4GHz+ stock speeds that SKL does.
3GHz for 8 cores would be a solid achievement, IMO.