AMD: Navi Speculation, Rumours and Discussion [2019-2020]

Status
Not open for further replies.
Code:
+#if defined(CONFIG_DRM_AMD_DC_DCN2_0)
+
+#define FAMILY_NV 143 /* DCN 2*/
+
+enum {
+    NV_NAVI10_P_A0      = 1,
+    NV_NAVI12_P_A0      = 10,
+    NV_NAVI14_M_A0      = 20,
+    NV_NAVI21_P_A0      = 40,
+    NV_NAVI10_LITE_P_A0 = 0x80,
+    NV_NAVI10_LITE_P_B0 = 0x81,
+    NV_NAVI12_LITE_P_A0 = 0x82,
+    NV_NAVI21_LITE_P_A0 = 0x90,
+    NV_UNKNOWN          = 0xFF
+};
+


New set of GFX10 ASIC revisions (NV10, NV12, NV14, NV21)
https://cgit.freedesktop.org/~agd5f...0&id=019d9cdfd8a698762897a856f4c37f6a0b149d00
 
I thought it was supposed to come by the end of 2020, along with PS5 and Xbox "Scarlett" APU parts - that would be "next gen RDNA - hardware raytracing" on the slides.
Roadmaps change, sometimes for non-obvious reasons. By no means am I saying that RDNA 2 is absolutely a 2021 product, merely that AMD did intentionally change things. Meanwhile, hardware vendors often take the more liberal interpretation of their roadmaps.
 
Code:
+#if defined(CONFIG_DRM_AMD_DC_DCN2_0)
+
+#define FAMILY_NV 143 /* DCN 2*/
+
+enum {
+    NV_NAVI10_P_A0      = 1,
+    NV_NAVI12_P_A0      = 10,
+    NV_NAVI14_M_A0      = 20,
+    NV_NAVI21_P_A0      = 40,
+    NV_NAVI10_LITE_P_A0 = 0x80,
+    NV_NAVI10_LITE_P_B0 = 0x81,
+    NV_NAVI12_LITE_P_A0 = 0x82,
+    NV_NAVI21_LITE_P_A0 = 0x90,
+    NV_UNKNOWN          = 0xFF
+};
+


New set of GFX10 ASIC revisions (NV10, NV12, NV14, NV21)
https://cgit.freedesktop.org/~agd5f...0&id=019d9cdfd8a698762897a856f4c37f6a0b149d00

so it looks like replacement for current line:

Polaris10 = Navi 10 256bit memory bus
Polaris 11 = Navi 12 128 bit memory bus
Polaris 12 = Navi 14 128/64 bitmemory bus

Navi 21 = Vega 20 (?) 384bit memory bus or HBM3

What then "Lite" moniker means ? It looks that Lite moniker is applied to all chips not just one as I initially thought.
 
Unless otherwise noted, you should consider dates in roadmap slides to be inclusive. So AMD is saying that RDNA 2 will come out by the end of 2021.

That slide interpretation would also mean that Vega 10 came out early 2017, which it didn't.

Unless there's a RDNA1 solution coming up with DXR hardware acceleration in the meantime, I find it a bit hard to believe AMD will take another 2 years to release cards with that feature. That would also mean AMD launching cards with DXR almost a year after both consoles released with RTRT hardware.

Not having add-on graphics cards for the PC that compete with their direct competitor's and the consoles' IQ features for almost a year seems like a terrible idea. I don't know how RTG could ever plan for a roadmap that could ultimately drive them out of business..
 
so it looks like replacement for current line:

Polaris10 = Navi 10 256bit memory bus
Polaris 11 = Navi 12 128 bit memory bus
Polaris 12 = Navi 14 128/64 bitmemory bus

Navi 21 = Vega 20 (?) 384bit memory bus or HBM3

What then "Lite" moniker means ? It looks that Lite moniker is applied to all chips not just one as I initially thought.

Lite are consoles GPU it seems.
 
That slide interpretation would also mean that Vega 10 came out early 2017, which it didn't.

Unless there's a RDNA1 solution coming up with DXR hardware acceleration in the meantime, I find it a bit hard to believe AMD will take another 2 years to release cards with that feature. That would also mean AMD launching cards with DXR almost a year after both consoles released with RTRT hardware.

Not having add-on graphics cards for the PC that compete with their direct competitor's and the consoles' IQ features for almost a year seems like a terrible idea. I don't know how RTG could ever plan for a roadmap that could ultimately drive them out of business..

IMO, despite what MS and Sony said, we don't know if accelerated RT will be like the nVidia solution. Maybe it will be compute based, and they could do that with navi already... I don't know. For the out of business thing, yeah... Navi doesn't seem to catch on nvidia size/perfs/consumption wise... But they seems fine with the IP side of things. They could take the Videologic/IMG route...
 
AMD just publicly stated they are going back to a one year cadence of product releases... and people have already forgotten and are misreading roadmaps and whining about delays.
 
Roadmaps change, sometimes for non-obvious reasons. By no means am I saying that RDNA 2 is absolutely a 2021 product, merely that AMD did intentionally change things. Meanwhile, hardware vendors often take the more liberal interpretation of their roadmaps.

Sure, but on 1Q 2019 Earning Call, AMD presented roadmap where Next Gen GPU still slated for year 2020 and they seem very confident about that.

https://www.anandtech.com/show/14286/amd-7nm-navi-gpu-and-rome-cpu-to-launch-in-q3
 
Ran across this story about AMD's adding a broad swath of changes for Navi's Linux drivers.
https://www.phoronix.com/scan.php?page=news_item&px=AMD-Navi-10-Linux-Kernel-Patch

A few things I noted from skimming is that there are two graphics rings, with the additional ring able to be set to different priorities. I think this might not mean two command processors, as how the hierarchy of microcode engine to pipe to queue maps to hardware is a bit nebulous to me and has shifted at times in other contexts. There's a block that would be handling this, although whether there's an internal set of controllers or dual-threaded support to give the two pipes may be up to AMD's discretion.

Also, there's a new IP block for scheduling the graphics and SDMA pipes, which may be part of what enables the priority tunneling mentioned Navi. A microcode engine scheduler manages the queuing and prioritization of the formerly independent rings--the lack of this prior might explain why GCN couldn't fully stop existing wavefront launches if the graphics pipe had enough control to always give itself some of the resources.
 
About the production yields of those APUs: is there a link between final clocks and production yields ? will they produce less APUs if the target frequency is higher ? If yes how much less ? Is there a linear relationship ?

I remember Sony saying they had phenomenal production yields just before PS4 launch. Was it because of the conservative clocks of the APU ?
 
Indirectly, yes. The clocks are a function of stability in the fabrication. Higher binned parts simply function at the target clocks at a reasonable TDP. It just so happens that one way to ensure function of transistors is to boost voltage (to a certain point), but that increases power consumption non-linearly, and there would be increased demands on power delivery and the amount of heat you need to dissipate. Of course, that can also work against operation if the designed system cannot deal with the heat adequately.

I don't know if QA rules have changed for consoles, but they're probably still expecting continuous operation at load for a significant number of hours as well.

---
I imagine once they get a slew of functioning chips, they can gather some data about achievable clocks and voltages, then see how much they want to spend on cooling & circuitry :?:
 
About the production yields of those APUs: is there a link between final clocks and production yields ? will they produce less APUs if the target frequency is higher ? If yes how much less ? Is there a linear relationship ?

I remember Sony saying they had phenomenal production yields just before PS4 launch. Was it because of the conservative clocks of the APU ?

Transistor characteristics will follow a bell/normal distribution, so the relationship will not be linear.
 
About the production yields of those APUs: is there a link between final clocks and production yields ? will they produce less APUs if the target frequency is higher ? If yes how much less ? Is there a linear relationship ?

I remember Sony saying they had phenomenal production yields just before PS4 launch. Was it because of the conservative clocks of the APU ?

There is a link in that the amount of chips you can manufacture at x MHz depends on x. I think it should be something like 1 - the cumulative distribution function(x) of the probability distribution. The distribution of max clock speeds would probably be roughly normal (someone please correct me if I'm wrong here), in which case the relationship would look like this, i.e., not really linear, except over the middle section (roughly): https://en.wikipedia.org/wiki/Normal_distribution#Cumulative_distribution_function

If the distribution is not normal, then it would look (potentially very) different.

And I wouldn't read too much into Sony's statements, as they were referring to production yields for the entire console, not just the APUs, and it's not clear whether the statements were about yields as a ratio of functional_consoles/consoles_built, or just the raw number of functional consoles built.
 
Yeah, we know from the past, that other IC components like RAM can be a bottleneck in terms of "yields."
 
Seems a fair point in the context of PS4's launch - the decision to double the RAM was quite late.
 
I thought it was supposed to come by the end of 2020, along with PS5 and Xbox "Scarlett" APU parts - that would be "next gen RDNA - hardware raytracing" on the slides.

Only Sony and Microsoft know exactly what arch the PS5/Scarlett will be on, and even they are probably still weighing tradeoffs and exact specs. Is it going to be easier to go with 7nm+ next year, or 7nm? What's the node upgrade path for TSMC like, so they can bring costs down, is 6nm which is cell compatible with 7nm better than 7nm+? There's no indication 7nm+ has slipped, in fact it's probably in oversupply. It needs EUV scanners, which TSMC bought almost all of this year, seemingly in the anticipation of Apple paying up for the newest node as always. But (rumored) Apple balked, so TSMC might have a ton of extremely expensive EUV scanners sitting around doing nothing. Will this lower the price and timeline for RDNA2? I dunno.

Even the very terms "RDNA 2" and "Next Gen" seem intentionally vague. Are they still considered "Navi" for the purposes of "PS5/Scarlett have Navi GPUs"? AMD also has reason to be vague, they're answerable to investors. If they can't guarantee RDNA2 will launch next year changing a chart to vaguely show it might somehow launch in 2021 is just covering their ass legally, as well as in PR terms. The only conclusions that can solidly be drawn are "Navi or something very close, with raytracing, exists somewhere, on some node, and will launch in consoles next year, and for consumers by 2021 or earlier".
 
Last edited:
Status
Not open for further replies.
Back
Top