AMD RyZen CPU Architecture for 2017

[Entropy]

Faster ddr4 support would be a good start already. It's quite a pity the Ryzen 3x00U chips don't improve on that. ddr4 so-dimms up to ddr4-3200 (with non-overvolted 1.2V) are nowadays easily available. Although they are stuff you'd never find in notebooks pre-installed, the budget so-dimms go up to ddr4-2666 too, that would easily get you another couple percents in graphic workloads. But as I mentioned, higher memory speeds require higher Ryzen SoC voltage at some point, and maybe 2400 is the limit for whatever voltage the 15W chips use.
IMHO there's no way for lpddr5 support for the Ryzen 7nm APUs (which will appear only next year, but that's still too early), we'll first see this memory in smartphones.

Well Synopsys has offered full 7nm DDR5 and LPDDR5 memory interface IP since october 2018, and Cadence does as well. So Joe 7nm garage SoC builder have been able to buy this off the shelf for half a year already. If AMD doesn’t support those memory types it’s by choice. Rather dubious choice IMHO given that they try to carve themselves a market niche by supplying CPUs with high core counts and APUs with theoretically high performance. They need the data flow to support that to really reap the benefits. Or rather, their customers do.

 

[xEx]

Well I think the chiplet approach would let AMD do that, decoupling the memory controller from the CPU would let AMD all the space they need for it and improve it independently from the CPU design.

 

[Entropy]

Well I think the chiplet approach would let AMD do that, decoupling the memory controller from the CPU would let AMD all the space they need for it and improve it independently from the CPU design.

We’ll see. I do believe that AMD would stand to gain a lot by being on the ball when it comes to the memory subsystem. Unfortunately, HBM hasn’t really materialised for PCs, but AMD may still be able to leverage their experience in the future.

Personally, as a user of stationary PCs, I’m not in the market for a new system until the PS5 has been introduced, and gained some multiplatform software support at its native level (i.e. not cross-gen titles). I guess that will be late 2021 at the earliest, and by then DDR5 (and possibly PCI-5) is likely to have been on the market long enough to have stabilized. It would be nice to have some headroom CPU wise in order to facilitate higher framerates than the console targets, but that may turn out to be tricky.

Then again AMD is likely to transition to TSMC 5nm around then, which seems like it could be a worthwhile lithographic step.

 

[Newguy]

Incredible necro but listening to a Jim Keller talk and he mentions this, haven't seen it mentioned before:

So I went into Wikipedia [...] and in the wikipedia page it mentioned that Intel ran a contest for the best branch predictor and the perceptron predictor won and it was a kid in Finland or something. So we called him up, paid him some money and put that in Zen

Pretty great little snippet, not bad for that kid either

 

[Kaotik]

Incredible necro but listening to a Jim Keller talk and he mentions this, haven't seen it mentioned before:

Pretty great little snippet, not bad for that kid either

So Zen is practically finnish architecture

 

[Albuquerque]

Seems legit to me! Don't underestimate the Finns!

 

[Laurent06]

I don't buy it. TAGE is the best (it won multiple times the Championship Branch Prediction) and it's French And the inventor spent some time at Intel.

 

[Laurent06]

Here is an interesting article about branch prediction (2018, perhaps a bit dated): https://hal.inria.fr/hal-01799442/document
Back then the guy who was the most vocal about "pure" Perceptron was Jimenez (IIRC from Univ of Austin), not a Finn guy.
Now it seems Perceptron and TAGE are used in conjunction to get the best results (see figure 1 of the article).