AMD - Zen chitchat

[badass]

The series is quite new so I wouldn't be surprised if they didn't switch to old Intel chips now given AMD didn't really upgrade FX past the 8370, just putting later cores into low cache dual module APUs so AMD are done at this point. Unless they do a cookery episode with the 9000 series

I would be more interested in the 3570K as that was the CPU I nearly got when I went for the FX 8350, though from memory it was something like £190 for the faster Intel chip vs £140 for the FX and I do like cheap The gap seemed to close between the two chips over the years to my amusement, but who knows what it is like now.

I got one of those too. Whilst I wish it had a bit more oomph for those "all 4 cores maxed out" moments it doesn't affect the very little gaming that I do. My monitor only does 60 FPS so I couldn't give two hoots about any more than that. It plays Eliteangerous just fine and with an RX 580 it seems to stay at the 60 FPS cap the whole time.

Highly detailed benchmark over

[kompukare]

CB were given some Renoir to review:
https://www.computerbase.de/2020-08/...-4750g-test/2/
More of a min review though. Guess they were given these to review from an OEM so no overclocking.
Also no dGPU - at least yet.

The idle system power figures are actually a lot better than I would have expected from a B550 board at around 25W.

Pity that Thunderbolt is so rare on AMD platforms as those tiny 1.9 litre ASRock DeskMini (A300 / X300) would be nice to pair with a eGPU.

[Xlucine]

Pity that Thunderbolt is so rare on AMD platforms as those tiny 1.9 litre ASRock DeskMini (A300 / X300) would be nice to pair with a eGPU.

As a luggable sort-of-laptop equivalent?

[watercooled]

Very interesting, the 2C 3000G is actually a 2C die, not the same 4C with 2 cores disabled: https://twitter.com/FritzchensFritz/...80052620378117

[CAT-THE-FIFTH]

Very interesting, the 2C 3000G is actually a 2C die, not the same 4C with 2 cores disabled: https://twitter.com/FritzchensFritz/...80052620378117

That might explain why some of the 300U laptops have been priced so competitively in the past. I wonder how it compares in size to the Atom SOCs??

[Xlucine]

The long heralded 2C/3CU die! It exists! This thing has been rumoured since before raven ridge launched.

[watercooled]

I wonder if anyone's measured it? I might do some pixel-counting if not.

[watercooled]

Curiosity got the better of me.

It looks to be roughly 12.6 x 11.52mm so ~145mm2. Compared with ~210mm2 for the full die: https://en.wikichip.org/wiki/amd/ryzen_7/2700u

[CAT-THE-FIFTH]

The package PCB according to Reddit is 40MM x 40MM. I got 12.51MM x 11.74MM which comes to around 146MM2. I tried the same PCB measurement,with this Raven Ridge die shot:
https://www.hwcooling.net/wp-content...-The-Stilt.jpg

I got around 207MM2,so around the 150MM2 mark seems accurate. The Intel Core i7 6700K had a die size of 122MM2,but it lacks many of the SOC functions Raven Ridge has.

[CAT-THE-FIFTH]

Cheaper B550 motherboards are tested by HUB. It appears the MSI and ASRock have the best VRM cooling,and the Gigabyte and Asus ones are bottom of the pile.

Edit!!

I can see why the B550 Mortar and PRO VDH do reasonably well:
https://www.scan.co.uk/products/msi-...lusc-micro-atx
https://www.scan.co.uk/products/msi-...1-aplusc-micro

The VRM heatsinks,do appear to actually have some fins. Even a BIOS reflash button on the backplate!

[CAT-THE-FIFTH]

Videocardz and Toms Hardware have leaked slides from the MS presentation at Hot Chips:
https://videocardz.com/90788/microso...x-architecture
https://www.tomshardware.com/news/mi...ture-deep-dive

Looking at Beyond3D it appears RDNA1 can do 5 instructions per clock,and RDNA2 7 instructions per clock. Apparently part of the RT pipeline is shared with the texture units:
https://forum.beyond3d.com/threads/a...61042/page-127

You can see it in one of the slides,4 texture or RT ops at the same time. So some of the peak RT figures might in reality be lower,once you start switching between texture and RT operations. However,this has apparently been done to reduce the number of transistors,and hence the die area. So I suspect Nvidia will be more brute force and AMD is more about efficiency??

Edit!!

RDNA2 appears to use the hybrid RT method AMD patented last year:
http://www.freepatentsonline.com/20190197761.pdf

[ik9000]

Videocardz and Toms Hardware have leaked slides from the MS presentation at Hot Chips:
https://videocardz.com/90788/microso...x-architecture
https://www.tomshardware.com/news/mi...ture-deep-dive

Looking at Beyond3D it appears RDNA1 can do 5 instructions per clock,and RDNA2 7 instructions per clock. Apparently part of the RT pipeline is shared with the texture units:
https://forum.beyond3d.com/threads/a...61042/page-127

You can see it in one of the slides,4 texture or RT ops at the same time. So some of the peak RT figures might in reality be lower,once you start switching between texture and RT operations. However,this has apparently been done to reduce the number of transistors,and hence the die area. So I suspect Nvidia will be more brute force and AMD is more about efficiency??

or more likely NVidia is about performance and AMD will be propping up the rear as an also-ran as so often the case in recent times. For all their faults the green team have at least kept pushing on with performance. It's not quite the same as an intel resting on its laurels.

[CAT-THE-FIFTH]

or more likely NVidia is about performance and AMD will be propping up the rear as an also-ran as so often the case in recent times. For all their faults the green team have at least kept pushing on with performance. It's not quite the same as an intel resting on its laurels.

Nvidia's method probably requires more die space,ie,it needs bigger dies and greater cost. Looking through the slides(you should look at them),the GPU part of the XBox SOC is not only under 200MM2,it appears to consume as much power as the XBox One X GPU(Polaris based). So it appears not only is it efficient in terms of die area,but power effiency is improved. It makes me wonder whether we potentially could be seeing another GTX200 vs HD4000 series,or Fermi vs HD5000 series scenario. Nvidia winning through just larger GPUs??

The RDNA2 changes don't look that minor at all.

Edit!!

If AMD can produce a more area efficient RT solution,it will mean you will see it in IGPs. So eventually AMD APUs,once we moved to DDR5 and maybe even smartphones.Samsung is an RDNA licensee.

Second Edit!!









Looking at the comments,it looks like the RT and ML parts add a relatively small area cost to the GPU.

[Terbinator]

I suppose it is a H/W presentation and not software, but was keen to learn more on DirectML implementation given MS already had it up and running on Turing.

Seems that MS are using the Renoir parts for CPU too, as it is monolithic 8 cores. 76MB cache, though, which should help massively for RT. Allegedly the PS5 SoC isn't as small as it should be given the difference in CUs.

[kalniel]

Finally some info on AMDs RT solution, thanks cat. That RT performance seems.. flipping fast. Hard to compare like for like, but peak theoretical 8-9x faster than Turing? Could end up faster than Ampere in RT, depending what else the GPU is doing at the same time.

[CAT-THE-FIFTH]

Well looking at how AMD is doing RT,it seems to be more general purpose,unless I am reading it wrong. So if it is using the texture units for part of the RT ops,I expect realworld those figures will be lower,BUT it also mean it will be easier to scale up,and rasterised performance will take less of a hit in non-RT games??

Ideally we would want a standardised benchmark to compare both. However,I do think Nvidia will have better RT performance(especially if they go for larger GPUs),and AMD is going for more of a middle path here. Either way,I only really care about the market under £500,especially around the £250 mark,so if AMD has something which is competitive in performance and power consumption that is fine with me!