Hacker News Reader

Comment by bdz | Hacker News Reader

Looking at those gaming benchmarks. Turns out the single core performance is still more important than having more cores.

https://www.purepc.pl/procesory/premiera_i_test_procesora_am...

View on HN

what's up with some of the tests? For example this one:

https://www.purepc.pl/procesory/premiera_i_test_procesora_am...

A Broadwell with 3300 Mhz (3700 Mhz boost) beats a Kaby Lake (Skylake...) with 4200 Mhz (4500 Mhz boost)? Can the Broadwell use the 128MB EDRRAM from the iGPU as a L4 cache if the iGPU is not used? Would that actually make such a difference?

> Can the Broadwell use the 128MB EDRRAM from the iGPU as a L4 cache if the iGPU is not used? Would that actually make such a difference?

That's exactly what it does, and it gives you a "free"* ~20% increase in memory operations performance (which tends to be a ~20% increase in general performance).

* not actually free because you paid for those 128MB EDRRAM...

Nice, now I'm wondering why Intel doesn't slap 128MB EDRAM on an Kaby Lake i7 4 core 8 thread and then release it for the enthusiast platform x99. Maybe even select chips to make it possible to release them at 5ghz. Of course with a large premium (expensive EDRAM etc.).

Because they had no competition, so no desire for higher performance. They can easily remove the GPU from Kaby Lake, use that transistor/thermal budget for 4 additional cores and add 128MB of L4 cache for a monster CPU.

The Iris chip (the extra EDRAM for GPU/L4) is seen by Intel as a performance improvement for mobile solutions (ultrabooks, MS Surface Pro, maybe Macbooks some day?) and not a desktop solution.

But from a consumer point of view, it increases performance (either for graphics or for the whole system if I have a discrete GPU), and it can also help with GPU numerical processing, because you get "free" transfers from memory (L4) to graphics memory by flipping a bit (or a few bits) in hardware. Don't know why this wasn't expanded further, appart from the cost of the embedded RAM.

This is what I got frustrated with and decided to stop waiting & buy a Ryzen, in fact. I've been looking forward to an enthusiast segment desktop chip with eDRAM and Iris/Iris Pro (or better!) from Intel, for a few years now, but it hasn't surfaced.

It's hard to put any faith in the large spread of numbers presented by this article regarding gaming performance. Anandtech just recently did a review of the I7 7700K and pretty much all the cpus over $200 posted the same numbers since most games are gpu bound.

http://www.anandtech.com/show/10968/the-intel-core-i7-7700k-...

main thread has to go somewhere. :\

Maybe knowing consumers will probably have more than 2 cores will make games optimise better for parellelism, it's not like the engine programmers _cant_ do it, they do it all the time with GPUs.

GPU parallelism is very different from main-core parallelism. The GPU is an entirely different architecture than the CPU. The stuff that's still on the CPU are things that don't trivially parallelize to begin with.

I can give you an example in my actual work;

When making AI bots it's quite common to have all of the work of processing the AI it's own thread, so each rendered frame will do all calculations for all bots. That in of itself would be cheap, except it has to share it's core with a bunch of other crap.

With 2 cores you can have a main thread for updating the screen and then an "everything else" thread. If you can't be sure people will be playing with >2 cores then you have to keep your budgets down to 2 cores when developing and that's quite limiting.

it's not that it's "hard" to parralellise(?), it's that you run out of resources super quick if you make assumptions about CPU topology.

Reply on news.ycombinator.com