Results: SiSoft Sandra 2014 SP2a
The latest version of Sandra recognizes and utilizes most of the cutting edge features of the latest processors. AVX and FMA are integrated into these tests, as well as a handy OpenCL test. While Sandra is not a productivity app, it does have a nice grasp on theoretical performance.
I was more surprised by the Dhrystone results than anything else. The latest Kaveri processors seem to do well, with a pretty large increase in performance over the Richland parts, and a small lead over the competing Intel part. Whetstone performance for Kaveri just is not that hot. The Intel part easily surpasses all of the other products here. In multimedia performance we see that Intel again has a sizeable lead with its AVX2 unit.
OpenCL results are fairly interesting here. I am not entirely sure what Intel’s OpenCL driver is doing, but the Double Float and Quad Float shaders have identical results as the Single Float shaders. I am removing those results for Double and Quad on the Intel part, as they do not look valid. Single float results do look inline with expectations.
Memory bandwidth was supposed to be improved on Kaveri, but with this particular case we are not seeing the improvements we were hoping for. That is not to say that AMD has not done a lot of work on the memory controller, as it does feed the graphics portion with as much memory bandwidth as possible.
How about 24 compute cores,
How about 24 compute cores, at least 10 of those “compute cores” CPU cores, yes based or your new x86 microarchitecture, the one your top CPU boffins are cooking up, and don’t skimp on the stacked on die RAM, stack um up high like the mondo-pancake stack at Ned’s house of sweet butter & lard! Those CPU cores need to be Full, no shared execution units, or instruction decoders, with lots of IPCs.
Yup, just pop in that new x86
Yup, just pop in that new x86 architecture. (That’s not expected to be ready for another couple years.)
Let’s add more than double the CPU compute cores. (Despite the transistor count already being nearly 2.5 billion, more than doubling bringing it close to 5+ billion, which would dramatically reduce yield and in turn dramatically increase manufacturing costs.) Their new CPU uArch will be twice as space efficient, after all. (Even though they can only do 4 jaguar cores in the same space of a single Steamroller module, SURELY their new big-core x86 chips will be able to double the core count without dramatically increasing transistor count.)
To seal the deal, add in ‘on-die stacked RAM’ (which is apparently very expensive, potentially causing even more yield issues, and therefore driving costs even higher.)
Chip design is just so easy!
Just go package on package,
Just go package on package, or put the GPU die, and the CPU die on a Mezzanine module, those CPUs, and GPUs were designed with scalability in mind years ago, just bump up the core count an enable some of the resources that were fused off. Yes a big fat module, with a 1024 bit bus, to some on mezzanine module stacked Ram, All on one big fat coherent interconnect fabric. There, if the CPU is bad, well it does not have to be, throw the GPU out with the bath water, same for the GPU. Hay it cold be 10 ARMv8 ISA based custom cores, as long as the SIMD instructions were there, hell some ARMv8 custom cores like the Apple A7/A8, or the Nvidia Denver cores, wide order superscalar designs, for rendering workloads(ray Tracing) to go along with the AMD graphics, and HSA goodness. It does not have to be x86, I’m Sure AMD has some good custom ARM designs ready to go to compete with Apple, and Nvidia. Or some Power8 based product, just license a reference design from openpower/IBM and integrate it with AMD graphics, no big x86 redesign worries there, Just some on die/on module interconnect work, AMD has some IP of its own there, or just get CAPI, tweak it and rebrand it, like Nvidia did for its NvLink, lots of IP out there besides just CPU IP/ISA that can be licensed to get to market quicker. The HPC world is going to be a big user of GPU accelerators on Mezzanine Modules, to go along with the CPUs, for speeds/bandwidth that even PCI can not handle. That HPC goodness will work its way down to the consumer market, But I need more cores, at least 10(CPU cores) to go along with the GPU, that is unless hardware ray tracing circuitry comes the market in a big way, Though I have not heard any new news concerning the PowerVR wizard GPU with the hardware ray tracing, but them Rays Need tracing, and tracing Needs CPU cores, lots of cores/threads. Chip design worries, Just look at the latest Intel E chips, Just some server SKUs that did not make the top bin, just fused of and gimped down, under a layer of really sticky thermal epoxy, just fuse and use. More cores, more cores, CPU cores that is, AMD!
This is what opteron boards /
This is what opteron boards / chips are for. You can get all this stuff today, in fact, you could get a dual socket AMD board for about the same price as a mid-range i7.
I’ll take a custom ARMv8 ISA
I’ll take a custom ARMv8 ISA chip with 24 cores, and some AMD graphics, in a portable workstation form factor, 7 or more IPC custom wide order superscalar cores, and plenty of SIMD/Neon goodness, to go along with the Firepro! Now AMD get some Dedicated Ray Tracing going on your Pro Graphics SKUs, and all those cores will not be necessary! Do they have Opteron based portable workstation SKUs, at rock bottom prices, hell AMD I am waiting for your Custom ARMv8 designs, Apple has got its Cyclone, soon to be replaced with a newer chip, and Nvidia has its Denver custom cores. I do not need any reference design ARM 3 IPC cores, give me 7 IPCs and some SIMD extended instructions. Hell make a Xeon Pi competitor(pricewise at least) with 96 custom ARM cores, and wider SIMD, and other Ops for Tracing them Rays, and keep your eyes open to what Imagination is doing with the PowerVR wizard, and get some Hardware Ray Tracing going on your Pro GPUs. And for sure look into getting your GPUs intimate with some Power8, on a Mezzanine module with stacked Ram for pillows, and an 1024 bit bus tying it all together, don’t let the green team get ahead, with a Power8 hammer drop, its only a license away, just like ARM stuff.
P.S. AMD get your gaming console APUs with some beefed up GPU cores on a discrete PCI card, and start the whole home gaming/computing cluster market going, and Fill those slots with complete gaming platform/s for some real low latency gaming, and tell the motherboard CPU to kiss off and take care of some house keeping chores, screw the consoles get the Clusters! Now that’s some HSA goodness on a PCI card, and Nvidia is heading that direction too!
I was expecting to see a
I was expecting to see a Pentium or a Celeron to compare with 7400K. I think this processor is in fact the most interesting between these two. 7800 we already know more or less what it can or it can’t do.
Thanks for the review.
On its own, the current and
On its own, the current and previous gen AMD APUs are ok. Unless these perform to at least the level of a GTX650, HD 7770/90, or its R7s, you’re still better off getting a discrete graphics card for a proper, budget gaming PC.
The 7600 seems like an
The 7600 seems like an absolute, super cheap gaming powerhouse for people who play on their TV`s @ 720p or dont bother with new 3xA games @ 1080p.
Happy with the Core i7 4790K
Happy with the Core i7 4790K drawing 88 watts.
PhoneyVirus
https://twitter.com/PhoneyVirus
https://phoneyvirus.wordpress.com/
Looks weak…my 4 year old
Looks weak…my 4 year old GPU is still better.