juanrga (noemail.delete@this.juanrga.com) on November 22, 2020 6:09 am wrote:
> Chester (lamchester.delete@this.gmail.com) on November 21, 2020 2:49 pm wrote:
> > > They are testing some overclocked chips, not getting IPC on stock settings,
> >
> > You can see perf at stock setting in regular reviews.
>
> Most regular reviews tested Zen on non-stock settings. But what is relevant
> to this thread is the above IPC comparison isn't measuring stock IPCs.

So there are two interesting data points: IPC at stock clocks, and IPC at identical clocks.

IPC at stock clocks will naturally favor lower clocked CPUs, because memory latency will be lower in terms of clocks. Because of that, I think measuring IPC at identical clocks is better if you're trying to directly compare architecture.

IPC scaling with clock speed should be tested separately, but they'd need performance counters to account for DVFS and sadly most review sites don't do that.

> > > and
> > > the choice of applications is too narrow: rendering, rendering, rendering,...
> >
> > Try scrolling down a bit. The little bar on the right of your browser window can do that - try dragging
> > it down or pushing the down arrow at the bottom, and you'll see a whole lot more than rendering.
> >
> > There's compression, gaming, video editing, image editing, as well as a couple
> > of synthetics.
>
> The author of the review differentiates between "games" and "applications"; it would
> be a bit weird if I add gaming when discussing his applications results. Yes, there
> is compression, this was included in the ellipsis that I wrote above.

Nice, you've picked up compression and rendering. Now try scrolling down more :) If you keep hitting the down arrow, the page will keep scrolling and you'll see more results like video/image editing. Keep going until the little bar on the right reaches the bottom.

> > Not the most complete coverage, but miles better than SPEC.
>
> Whatever you say, but the review is rather irrelevant to get the average IPC,
> because it is comparing stock vs non-stock, using a very narrow choice of application
> workloads

Number of useful SPEC subtests: 1.5 (counting GCC as 1, blender as 0.5 because you don't like rendering)

Number of useful subtests in that review:
Cinebench R20 (counting as 0.5 because it's not as long as a real high quality render - it just runs a short one)
7z compression
7z decompression
Blender, V-Ray, Corona (sure, count as 0.5 if you don't like rendering)
DaVinci Resolve Studio
Adobe Premiere Pro
Adobe Photoshop
Adobe After Effects
And a set of games people actually play

Of course they could always expand coverage. Compilation could be added for example. But if you don't complain about spec, you should definitely not complain about that review. It's meant to be a quick comparison at matched clocks, and its relevant coverage is already better than spec's.

> and not providing full CPU tests in the gaming measurements.

What's not 'full' about CPU tests in the gaming measurements?