[Manuel Calavera]

Evo dio jednog teksta koji ima teoriju da će 8 core Zen (barem na onim prezentiranim clockovima) biti nešto iznad i5 6600K i ispod i7 6700K, nekakav cca 300$ cpu. Ovo je tip utemeljio na onome amd "orochi" slajdu

My attention, in trying to solve this mystery, turned to AMD's latest slide. The slide doesn't actually say that Summit Ridge doubled the FX-8350 performance. Instead, it says Summit Ridge doubled "Orochi's" performance.

As it turns out, Orochi and the FX-8350 aren't the same thing. You see, Orochi referred to the first Bulldozer module family of CPU's, dubbed "Zambezi"/Orochi. The FX-8350, however, was based on subsequent Piledriver modules. As Anantech said when the FX-8350 was launched (emphasis is mine):

Of course, Vishera CPUs were more performant than Orochi CPUs. But the problem doesn't end there. You see, AMD has been very keen on comparing same-frequency CPUs, and there wasn't one Orochi 8-core CPU, there were several. And what was closest to what AMD has been dubbing "Summit Ridge"? It was the FX-8100:

- It had the same base frequency (2.8 GHz) as Summit Ridge.

- It had a similar turbo frequency (3.1 Ghz full-load vs 3.2 Ghz for Summit Ridge). However, here there's doubt on whether Summit Ridge refers to a full or half-load turbo frequency. The FX-8100 half-load turbo frequency was 3.7 Ghz.

- It even had the same TDP (95W).

Thus, it's very likely that AMD's chart referred to the FX-8100 and certainly not the FX-8350. Of course, you already wonder why AMD isn't clearer about these comparisons. That's never a good sign.

This, in turn, changes the "double performance" comparison quite a lot. You see, the FX-8350 has a Cinebench R15 multi-thread score of 640. Doubling that gives us 1280, which would be close to Intel's i7 5960X's 1337. But the thing is, the FX-8100 is necessarily lower.

A Problem

There's something of a problem here. Anandtech doesn't actually carry a Cinebench R15 score for the FX-8100. It does have a score for the FX-8150, but the base clock on the FX-8150 is 3.6 Ghz (so 28.5% higher than the FX-8100), the full-load turbo frequency is 3.9 Ghz (25.8% higher than the FX-8100) and the half-load turbo frequency is 4.2 Ghz (13.5% higher than the FX-8100). The FX-8150 is a significantly faster CPU than the FX-8100, and the FX-8350 is faster still.

Comparing the FX-8350 to the FX-8150, we notice that the FX-8150 is 13.75% slower, whereas the base frequency difference alone would imply just a 10% difference. This is understandable, as the FX-8350 gained not just in frequency but also in architecture improvements.

Anyway, assuming the FX-8150 scaled linearly with frequency from the FX-8100, then the FX-8100 would have a Cinebench R15 score of around 429. That's a far cry from the 640 attained by the FX-8350 and then doubled by the media. Doubling 429 would put Summit Ridge at ~858. Now here's what this implies:

- 858 largely exceeds the 654 posted by the i5 6600K, which in my previous article based on the Ashes of the Singularity benchmark, showed up as being 33% faster. Not anymore. Part of this is expectable, the Ashes of the Singularity doesn't scale well above 8 threads, whereas Cinebench will make use of all 16 threads Summit Ridge can provide. However, it should also be noticed that on a multi-thread and Cinebench R15 basis, the FX-8350 was already competitive with the Intel i5 6600 (611 score, vs 640 for the AMD).

- The Cinebench multi-thread improvement over the FX-8350 looks to be around 34%, which is more or less in line with what the Ashes of the Singularity benchmark predicted (38%).

- This puts Summit Ridge below the Intel i7 6700K (a $350 CPU, though, vs the $177 FX-8350) on multi-thread performance. Also, it's likely that this will put Summit Ridge around the bottom of all 7th generation Kaby Lake i7s. And of course, it puts Summit Ridge significantly away from the i7 5960X. This benchmark confirms the Ashes of the Singularity benchmark, except in as much as this benchmark is better able to use all available threads. Such comparisons, however, were already valid with the FX-8350.

Conclusion

AMD's Cinebench R15 slide was misleading, and media seems to have taken the bait, leading it to believe Summit Ridge would be competitive with a high-end Intel CPU (the i7 5960X). A closer analysis shows that such isn't so.

As it stands, the Cinebench R15 benchmark remains consistent with the Ashes of the Singularity benchmark I used in my previous article, save for the fact that Cinebench better uses all available thread-processing capabilities. This, however, already happened with the FX-8350 in the past. So in Cinebench R15 the FX-8350 was already competitive with the Intel i5, yet in the marketplace it wasn't (it struggled to be competitive with the Intel i3).

The reason why it wasn't competitive with the i5 in the marketplace is obvious. While the FX-8350 was competitive in the Cinebench R15 multi-thread benchmark, it struggled horribly in the single-thread benchmark (i5 6600 at 169 versus FX-8350 at 96). This discrepancy was based on IPC (76% difference) since frequencies were similar and a 40% improvement in IPC isn't going to cure that. Moreover, a drop in frequency forced by the 8-core nature of the chip will degrade IPC gains (which are then only partially compensated by better multi-thread performance on account of SMT).

Unfortunately, everything indicates that even if Summit Ridge allowed for some narrowing of the IPC disadvantage, the reality of the marketplace won't change significantly:

- On situations needing single-thread performance, even an Intel i5 will remain more competitive than a Summit Ridge chip. The improvement in IPC does allow Summit Ridge (and presumably, the 4-core Zen chips as well) to be competitive on single-thread performance with the Intel i3.

- On situations needing multi-thread performance, an Intel i5 won't necessarily be better - but then again, it already wasn't. Yet, an Intel i7 will remain competitive versus the Summit Ridge, and an 8-core Intel chip will still trounce it (though at a significant price difference).

Being optimistic, the Zen improvement might allow for a bit better ASPs for AMD in as much as Summit Ridge will go from competing with low end i5s to competing with low end i7s in multi-threaded applications, and from not competing with i3s to competing with them in single-threaded applications.

This, however, won't change the competitive nature of the market. Going from the data we have, AMD will still be forced to use 8-core chips to compete with Intel 4-core chips. The benchmarks we have, flawed as they might be, indicate that this is the most likely outcome. The same benchmarks indicate that the multi-thread gains versus the FX-8350 should be slightly less than 40%, because the gain on IPC and SMT are then partially lost on lower frequencies.

We now have 2 different benchmarks confirming these conclusions: The Ashes of the Singularity benchmark, and the Cinebench R15 benchmark, though the Cinebench R15 reality was rather hidden.