• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

If the Instinct MI100 has 120 CUs, couldn't RDNA 2 have 96-120 CUs?

g67575

g67575

g67575

Soldato
Joined
30 Jun 2019
Posts
8,159
The Arcturus / CDNA 1.0 based Instinct MI100 has 120 Compute Units, running at a TDP of 200w, according to a leaked BIOS, in Feb 2020, link here:

https://www.techpowerup.com/263743/...i100-arcturus-hits-the-radar-we-have-its-bios

AMD seem to have achieved a very low TDP of 200w by clocking the GPU at just 1334 Mhz, leading to a GPU capable of 20.49 TFlops. Previous Vega based Instinct GPUs like the MI60 ran at 300w, and was capable of 14.75 TFlops.

Apparently, the MI100 has no 3D engine, which could explain how they were able to increase the CU count to 120 (due to reduced die size).

All this makes me wonder, for the top RDNA 2 GPU, couldnt AMD reduce the CU count to ~96 (cheaper to produce), then increase the GPU clock upto Navi10 levels? That should keep tdp at 300w or less. AMD wouldn't need to use the HBM2 VRAM or 4096 bit memory bus, so power savings could be made there.

I guess this comes from the idea that the CDNA and RDNA 2 architectures might be very similar, in some ways. The CDNA architecture is very power efficient it seems, it would be a waste not to carry these improvements across to next gen GPUs.
 
Last edited:
Unlikely, a huge core has much higher risk of being defective so low yields, more cost hence Nvidia using GA102 not GA100 for GPU.
 
Yeah, but even Nvidia's RTX 3070 has 5888 shader units :D

I think AMD will need a GPU with at least 92-96 CUs to compete with it.
 
It's not just down to CUs, and remember those cards are designed for compute, which utilises all the cores much easier, so you can actually run them full tilt. For graphics it doesn't really work that way, and we saw that with GCN, culminating with Vega which was an arch that had a LOT of power but in games the cards never got utilised better than at 60% (best case). And indeed you can even see it with Ampere now, it has a lot of cores on paper but in reality the performance gains don't translate 1:1.

Really the most important thing is proper utilisation of available resources, and they achieved that much better with RDNA 1.
 
I wouldve thought 78 CUs for the top RDNA 2 would be the minimum, since the 200w TDP series X GPU has 52 CUs.

I think the most difficult thing to predict is how much will difference improvements to performance per GPU clock will make. For the series X GPU microsoft said it is a 25% improvement. If thats true, that would make it around 55% faster than the 5700 XT (which has 12 less CUs).
 
Last edited:
I'm a bit surprised there hasn't been more discussion about it on the interwebs, do people think the CDNA architecture isn't relevent to future consumer graphics cards, because it isn't designed for gaming?

I'm optimistic because I had no idea AMD could produce a GPU with 120 Compute Units, they've never build one with more than 60 in the past (64 if you count the Instinct MI60).

what do you chaps think? Will it influence the design of RDNA 2 and future AMD gaming graphics cards?
 
Last edited:
g67575 said:
Yeah, but even Nvidia's RTX 3070 has 5888 shader units :D

I think AMD will need a GPU with at least 92-96 CUs to compete with it.
Click to expand...

Why would they need 92-96 when the 40CU RDNA1 5700XT is around 35-40% off 2080Ti territory already? I don't see why a 64CU RDNA1 card without any IPC improvements wouldn't be similar let alone the upgrades almost certainly coming with RDNA2.
 
64 CU might be alright if there are some large improvements in IPC too (or whatever the GPU equivalent is).

I hate not knowing if RDNA 2 will be a worthwhile upgrade. I've been waiting 4-5 years to upgrade, don't know why another month or so would seem like such a long time :D
 
Last edited:
I suppose another possibility is that the flagship RDNA 2 GPU will have the same clock rate as the PS5 GPU (2233 MHz), but with twice the Compute Units (72). This happens to be 12 more CUs than the Radeon VII.

That would mean the GPU would be capable of 20.5 TFlops, slightly more than the RTX 3070.

If the cooling was sufficient, that would probably be cheaper than a graphics card with 80 or more Compute Units. I think this spec makes more sense than any other estimates I've seen so far.

I'm not sure I believe the PS5 GPU has a TDP of 180, its probably closer to 150w or less. The estimate on techpowerup appears to be based on the 180w of the RX 5700 (Navi 10).

We know there have been optimizations to power usage and efficiency with RDNA 2, this might just translate into higher GPU max clocks, at lower TDP than RDNA v1.
 
Last edited:
You must log in or register to reply here.
Back
Top Bottom