Tuesday, December 17th 2019
AMD "Renoir" APU iGPU Configuration and Platform Spread Detailed
AMD's upcoming "Renoir" silicon will be the company's most important, as it will sit at the heart of not just desktops, but also notebooks and ultraportables. A brilliant report by _rogame on Reddit compiles the chip's many iGPU variants along with iGPU device-IDs, and slots them in various platform variants. Renoir will target four key market segments characterized by TDP: 15 W ultraportables, 45 W mainstream notebooks, 65 W mainstream desktops, and 35 W low-power desktops.
As for the iGPU itself, "Renoir" was last reported as being a processor that combines "Zen 2" CPU cores with an iGPU that has SIMD machinery from the "Vega" architecture, but with updated display- and multimedia-engines from "Navi." According to _rogame, Renoir's iGPU will have up to 13 NGCUs, which work out to 832 stream processors. AMD internally marks the iGPU as RV B##, where RV refers to "Radeon Vega," and B## referring to the iGPU variant. The commercial name of the iGPU will be different. B12 is the highest variant, with 12-13 CUs, B10 has 10-11 CUs, B8 has 8-9 CUs, B6 has 6 CUs, and B4 has 3-4 CUs. The B12 configuration will be exclusive to the mobile parts. The desktop parts cap out at B10. Renoir is expected to dominate AMD's processor launch cycle through the first half of 2020.
Source:
_rogame (Reddit)
As for the iGPU itself, "Renoir" was last reported as being a processor that combines "Zen 2" CPU cores with an iGPU that has SIMD machinery from the "Vega" architecture, but with updated display- and multimedia-engines from "Navi." According to _rogame, Renoir's iGPU will have up to 13 NGCUs, which work out to 832 stream processors. AMD internally marks the iGPU as RV B##, where RV refers to "Radeon Vega," and B## referring to the iGPU variant. The commercial name of the iGPU will be different. B12 is the highest variant, with 12-13 CUs, B10 has 10-11 CUs, B8 has 8-9 CUs, B6 has 6 CUs, and B4 has 3-4 CUs. The B12 configuration will be exclusive to the mobile parts. The desktop parts cap out at B10. Renoir is expected to dominate AMD's processor launch cycle through the first half of 2020.
63 Comments on AMD "Renoir" APU iGPU Configuration and Platform Spread Detailed
I guess I'll have to wait and see but bleh.
This is more for notebooks..
browser.geekbench.com/v5/cpu/compare/816466?baseline=758268
I swear AMD is the fucking worst when it comes to naming their products. They ditched Radeon HD 8000, 9000 and went straight to Rn 200. Then after RX 500, they leaped back to RX 5000. Jesus fuck! And just forget about whatever they named their mobile and desktop APUs in the meantime. People make fun of Intel's 10k series naming scheme but at least they and Nvidia to some extent follow the numbering scheme.
Intel and Nvidia have the advantage of time. Core and GTX have been around for a long time with well understood naming schemes (if you don't account for rebrandings). Its only recently that these things have derailed too with Intel's absurd i9-10900KFS or whatever and Nvidias RTX confusion along with endless GTX 1660 variations.
In the end - no company is perfect in their naming because often it's a marketing decision that can change on a whim.
Everything so far points at this being Vega, so not that big of an improvement.
At least hopefully higher clock speeds due to 7nm.Getting to number 10 or the next generation after 10 have always been problematic and marketing decisions at that point have been strange.
10900KF sounds stupid but actually follows their established naming model. 10 - 10th gen, 900 - higher end, K - unlocked multiplier, F - without/disabled iGPU.
spot on. At least they got Zen / Ryzen / Threadripper right in naming. I wonder if they can keep that current convention up or feel obliged to follow Intel/Nv again and pick at their modelnumbers.
But yeah... for GPU its pretty ridiculous. We went from HD to R---(X) to RX and got none the wiser, and Vega is all over the place. Then again, Nvidia is also going borderline these days with 20xx and 16xx side by side... never mind the jump from 10xx to 20xx. WTF was that anyway.
I think the theme here is: slow GPU progress in absolute performance per gen = more naming confusion to hide it.
Look at RX5500 review on TPU:
www.techpowerup.com/review/amd-radeon-rx-5500/31.html
Vega idle power draw is very high, but multi-monitor is just +2-3W.
Radeon VII is exactly the same.
However, RX5500 gets +14W (4->18).
Even if we assume there's a 3W measurement error, Navi figures are absurd (even if it becomes something like 2W->9W in a smaller mobile SoC).
So AMD has to stay with Vega for most important segment: mobile <=25W SoCs (and likely the desktop 35W as well).
I guess dividing the lineup and making larger SoCs on Navi didn't make much sense.
Multi-monitor power draw is a tricky thing. 2 vs 3 monitors are different, high refresh rate monitors are different plus combinations. This seems to heavily depend on power management implementation in drivers. 2 monitors are generally figured out but if one of them is a high refresh rate one, both vendors' GPUs tend to run higher than expected. Once you add a third display you usually get a very noticeable increase in clocks and power consumption. Add to this that monitors can be hotplugged and that will throw another wrench into it.
I have mostly had Nvidia cards and both Maxwell and Pascal cards went a little back and forth on how they handled multiple monitors, mostly with driver versions. Somewhere right before Turing was released Nvidia managed to semi-reliably fix clocks/power for 2 monitors with one high-refresh rate one. Then it was broken for several months again. Right now, everything works again for my Turing GPU, including lowered clocks and power for all 3 screens connected at once (1440p 165Hz, 1440p 60Hz, 2160p 60Hz). I have no faith in this being permanent :D
AMD had a pretty good handle on single and dual monitor power with Vega. Single-monitor idle was and is downright impressive, dual-monitor is OK but at least once the early issues were resolved it worked reliably. Triple-monitor seems to be hit and miss. Mostly miss with load clocks applied. I have much better impression from Vega with this but I did not use the card for nearly as long periods of time. I have the same impression from Vega 11 in my 2400G but this one only rarely has more than one monitor connected.
Resolutions in this scenario are 1920x1080 and 1280x1024, so likely lower than we would expect in 2019, so power consumption could be even worse.Basically, it was very bad with Polaris, but they somehow fixed it in Vega (and in Polaris-based Radeon Pro for Apple).
Now it's (kind of) a new architecture and power issues came back.
There's no way Apple could accept this, so they'll help them with optimization (I guess: again).
Tigerlake will add, according to rumours a lot of GPU power to the already good Ice Lake GPU, so we will see how this ends up.
As for desktop, yes, these APUs won't have any competition.
Vega on 7nm will still lead Intel's TGL in most cases IMO, unless of course Intel does something ridiculous & just dedicates like 50~75% die to the IGP & doesn't go beyond 4c/8t for ULV versions. Remember those AVX512 instructions, they take a helluva lot of Si real estate!
That said, this is the moment where chiplet design gets in the way a bit.
Assuming no changes to the general SoC layout, it'll be 80mm2 max for the GPU chip.
I bet AMD would love to use a larger one if possible.
Ice Lake 1065G7 is relatively tiny for what it offers. The GPU part is around 40mm2.
Imagine what will happen when they start making larger SoCs for desktop...
Even so, Zen2 and 7nm should help with power efficiency. The current APUs are hampered by a slightly disappointing idle power that affects on-paper battery runtimes that a lot of people use to make their purchasing decisions, and additionally, the older manufacturing process meant that the 15W 3700U throttled under combined CPU+GPU load unless given an extra 7-8W. 2500U/2700U/3500U/3700U run games almost 100% faster if you can give them 50% more TDP - beyond that you're well past the efficiency sweet spot and get almost nothing more out of the IGP.
I'm hoping that we can see ultraportables in 1H2020 with 15W Vega12 and an efficiency sweet spot for CPU+GPU combined loads that more closely match the sustained 15W TDP.
AMD refuses to employ a better memory controller to take advantage of higher memory speeds that could feed a newer, better iGPU. AMD's refual to embrace higher memory speeds, LPDDR4X, refual to use the latest transistor processes, refual to use the newerst CPU arch, and refusal to use the latest GPU arch merely shows AMD doesnt take mobile seriously, or doesnt care about mobile users. They are content to simply absorb those $450 walmart specials and leave the high margin machines to intel. Again. The 3000 series APUs should have had 7nm zen 2 cores and 7nm rDNA cores for this upcoming holiday season, instead of launching in the spring with last gen hardware.
I know some will tout the argument that AMD just doesnt have the resources to do all this at once. Fine, just like ryzen APUs launching without AMD supplied drivers, if AMD cannot give their all in this market, they shouldnt bother. They are just wasting money and time when they could be putting those resources into the server market and its rediculous margins, or inject some life into their comotose GPU driver department, or invest even more into zen 4 and 5.
www.techpowerup.com/258859/amd-renoir-apu-to-support-lpddr4x-memory-and-new-display-engine