Comment by dragontamer | Hacker News Reader

Comment by dragontamer | Hacker News Reader

Parent Full thread

> What is anandtech doing different?

AMD's officially supported memory is 2933 MT/s. Intel's officially supported memory is 2600 MT/s.

Other review sites overclock memory and run both systems at 3200 MT/s RAM (closer to reality: you may have to spend $20 on RAM but the faster speeds seem to be worth it these days). It seems like Anandtech prefers to run the system on default settings however.

Its definitely a discrepancy. But I don't necessarily think Anandtech is in the wrong. It just goes to show how much these little decisions matter in benchmarking, and why its good to have multiple benchmarks from different sites.

So, do you go for the "obvious overclock" with 3200 MT/s RAM, or do you actually follow the official recommendations from the manufacturers? Regardless, its good to see that different websites have tested the two different cases and leaves it up to the reader.

See this line: https://www.anandtech.com/show/12625/amd-second-generation-r...

> As per our processor testing policy, we take a premium category motherboard suitable for the socket, and equip the system with a suitable amount of memory running at the manufacturer's maximum supported frequency

Personally speaking, I think running RAM at an overclock at 3200 MT/s is more fair to both systems. Both systems achieve 3200 MT/s easily, with Intel overclockers often hitting 4000MT/s or more (so really, Intel has the better memory controller). So its just... odd... that AMD would be equipped with better RAM.

But Anandtech's approach is reasonable (although they probably should list the timings of their RAM in their setup page). After all, most computer builders probably don't know the finer overclocking details between the systems, and may instead just go for the officially supported numbers.

I like that they went stock.

I build fast machines but I don't think I've ever overclocked one, I certainly can't remember.

I value stability and longevity, my 2500K lasted til end of last year (it still works fine, it's just my laptop is faster).

A lot of modern RAM can hit 3200 MT/s within the specified voltage settings. RAM literally has gotten better over the past few years, but "official" JEDEC timings lag behind.

Its why Intel has XMP: so that your sticks of RAM will talk to the motherborad and automatically overclock themselves to factory-tested values these days.

These "easy overclocks" I'm talking about are very simple! The factory gives you an XMP profile of the RAM under specified conditions, and then the motherboard enables a checkbox for XMP-profiles in the BIOS.

No need to tweak timings or anything. These days, overclocking RAM is a question of "Do you want to run your RAM at the speed tested at the factory" ?? Or would you rather run RAM at speeds specified by JEDEC from 2014?

Every RAM manufacturer worth a darn has an XMP setting, where they've thoroughly tested the RAM at these higher speeds. You might as well spend a minute, click the button in the BIOS/UEFI prompt and accept the free performance benefit.

Thanks for the recap.

> Personally speaking, I think running RAM at an overclock at 3200 MT/s is more fair to both systems

at that point you're comparing a system that AMD doesn't sell to a system Intel doesn't sell

FWIW, SDRAM transfer rate isn't really a big performance driver. The bus is idle almost all the time on modern parts, which are dominated by the precharge/read-write cycle and not transfer. When the 512 bits of data actually arrives, it gets blasted out very fast. Shrinking that transfer time has a measurable effect, but it's not remotely linear in the clock frequency.

Its certainly non-linear, and you're right that its dominated by precharge / read-write cycles.

But DDR4 RAM has support for 16 banks per stick! A CPU with two memory controllers can have 64 outstanding requests to DDR4 RAM simultaneously (16 per stick, x2 from the chip select, x2 for the 2nd memory controller).

And since modern cores are highly multithreaded, and also have prefetchers and whatnot, I'm sure the modern DDR4 bus is quite active.

Perhaps it doesn't matter too much for single-threaded situations (single-threads only benefit from prefetch and out-of-order executions). But a Ryzen 2700x is a 8-core / 16-threaded beast, which will surely eat up all of those outstanding DDR4 requests. Intel's 8700k is 6-core/12-threads as well, so they'll rack up those memory requests and keep the bus busy.

And if each of your Assembly commands are AVX (256-bit requests per instruction)... well... that's going to be a busy bus.

RYZEN seems to like fast memory more than Itel one review I just looked at said that the new RYZENS are targeting 3400

This is because Ryzen's "infinity fabric" or internal interconnect between the cores is tied to the RAM speed.

This is kind of like the old days when CPUs used a Front-side Bus (FSB). If you increased the FSB frequency it would increase the CPU frequency, the RAM frequency, and even PCIe depending on how it was wired up. It made things simpler to have the FSB be the sole clock-generator.

So despite the increased bandwidth moving from 2600 to 3200mhz RAM not having much impact on performance (in most applications), the increased frequency of the bus between the cores allows for greater performance beyond that.

IIRC Intel uses a ring-style bus with it's own clock, hence why it this effect only applies to AMD.

Yeah, up until 3200 MT/s the IF has a lot of impact.

> the old days when CPUs used a Front-side Bus (FSB)

They still kinda do. It's not the bus, but there is one clock called BCLK (base clock) now and almost everything is derived from it. It's not adjustable on cheaper boards, but with a good board you can absolutely do the same adjustment :)

Yeah I have an older box running an i7-920 which is overclocked from 2.6 to 3.6ghz with the base clock.

Sandy Bridge and newer get extremely unstable when you toy with the base clock more than a few MHz though. You're basically stuck shelling out an extra few bucks for a "K" chip if you want to overclock on Intel now.

Skylake and newer have fixed this, they decoupled PCIe and stuff from BCLK. On Skylake you can actually do non-K (I ran a 6400 at up to 4.7GHz):

https://overclocking.guide/category/intel-oc-guides/skylake-...

Of course on Kaby and later they closed the loophole...

Reply on news.ycombinator.com