The timings don't have much affect since the bottleneck is primarily a bandwidth one rather than a latency one. I remember that it was originally planned to have GDDR5 embedded on the chip but that was canned for the first gen Kaveri -- perhaps due to budget constraints. Instead of triple channel, I don't see why they couldn't just go straight for quad channel.
Originally Posted by jakubo
Expensive memory modules, I guess.
Originally Posted by jakubo
Documents indicate that kaveri may have support for quad-channel and GDDR5 memory, but that it has been disabled. I suppose a different package would be needed for quad-channel, and GDDR5 requires very short traces, so it would likely only work right with BGA systems where everything is soldered down.
Originally Posted by mmstick
Wow, the GPU tests scale linearly with memory bandwidth available... that means the GPU is very bandwidth starved? Maybe they could have gotten much better performance by tuning the memory controller? Benchmarks indicate that the AMD memory controllers are less efficient than the Intel memory controllers.
Last edited by grigi; 01-21-2014 at 06:58 AM.
Reason: wasn't finished
You're comparing a 300US cpu to a 190US CPU... Gee Gorge can I pet the rabbits. The thing is Benchmarks can lie in Intels favor but the truth is the Kaveri (like other CPUS in it's bloodline) Is pretty memory starved and lack of a L3 hurts it's performance. Does that make it a bad dual module? No it makes it a pretty darn good one just with room for improvement...
Originally Posted by Rexilion
That being said if they make a 4 mod steamroller based FX I would buy that in a heartbeat.
The truth is, Intels are far better in fpu performance and their memory controller is more efficient. They also consume less power.
Originally Posted by nightmarex
AMD is just cheaper/price competitive on low end and with a better igpu, though it is quite bandwidth starved and you don't get to see its real performance.
Also HSA and Mantle are vaporware, especially on Linux. By the time AMD brings the software that exploits them, Carizzo or its succesor will be available...
I did a few calculations on the first test, and while it doesn't remain 1:1, but drops more and more the closer you get to 2133Mhz, the numbers seems rather predictive. I think you could make a pretty accurate mathematical projection of the performance effect of an increase in memory speed, if one was so inclined.
Originally Posted by UraniumDeer
Next test speed / current test speed * current framerate result = rough estimate of next test
actual result of next test / rough estimate of next test * 100 = actual performance compared to rough estimate of next test in percentage
You mean you kept the timing artificialy high for the lower frequencies instead of using the profiles? That makes the benchmarks (especialy the CPU ones that care more for latency than bandwidth) complete bulshit.
Originally Posted by Article
AMD's Bulldozer CPUs have half the number of FPU per core, so of course that Intel is twice better in floating point operations than AMD.
Originally Posted by TemplarGR
Carizzo will also support HSA and Mantle...
if they need short ways for the gddr memory they could place the Memory on the backside of the mainboard. Cases would have to adapt different possibilies of removing memory and/or heat dissipation. You could for instance make some jelly-like pillows on the RAM modules to push against the case without having electrical contact to use the case as a massive heatsink. there are ways. and i think that AMD should make some innovative and risky jumps in order not to lose completely to Intel.