Baseless Next Generation Rumors with no Technical Merits [post E3 2019, pre GDC 2020] [XBSX, PS5]

Status
Not open for further replies.
One clock cycle by crystal frequency = one clock cycle. The amount of work any particular chip can achieve in one clock cycle = variable. I'm old so I remember when no processor could perform a single operation in a one clock cycle unless it was NOP (No OPeration).
We are about the same age then.
But IIRC instructions now often take 5 cycles, on GCN it was 4. (Could be wrong)
I have not worked on Navi yet, so i simply don't know how compute performance compares.
From the whitepaper i guess it now takes less time until GPU is saturated after switching shaders / new dispatches, and there is the better cache and better random access. So i hope it will work a bit faster for me than GCN. But could be also the other way around.
On the other hand, for example NVs concurrent integer / floating point units - that's something you can be pretty sure you get a compute speed up from the new architecture, but on Navi there is nothing really indicating compute perf has increased.
If you look at compute benchmarks, the impressions are pretty mixed.
And looking at game benchmarks tells nothing about compute performance.
 
So we are going from :

"9.2TF???Sony is skimping! This is unnacaptable"

...to

"2.0GHz?? Are they mad? 36CU Navi at that clock will burn your console, should have gone nice wide and slow"

1. There are rules to chip design. Navi CUs are much larger then GCN, therefore they are not straight comparison vs PS4Pro
2. 2.0GHZ Navi, that was product of many years of work and big budget, will almost certainly be better chip then first Navi's that AMD released 1.5yrs before this one is released
3. I assume console manufacturers will use N7P node, as this does not require chip redesign but will give them round about ~10% better TDP at same clocks
4. Wide and slow vs narrower and high clocks is not clear cut. Bigger die sizes are getting more and more expensive. Chips with high transistor counts cost much more then the ones back in 2004, there is a reason why Nvidia and AMD are pushing frequency. Its BETTER.

Really, up until 12TF mini PC premium SKU that MS shown, everyone would have LOVED ~13TF Vega with RT and VRS in next gen console so I don't really understand what is the fuss? Only way Sony will ever match MS is if they go with 2 SKU strat or say "Fuck lower end, Lockhart can eat bottom end, we will only release premium SKU" which, IMO, would be suicidal.

So yea, bottom line. 36CU at 2.0GHz is achievable in 2020, especially on N7P node. It will still result in very advanced console and 399$ is NOT guaranteed. Fact that they went narrow and fast does not mean going wide and slow, like 56CUs at 1500MHz would result in better performance. Only if wide and slow means, 56CUs at 1700Mhz+, which I imagine was never a plan for Sony as this is very much as premium as it gets (and, will have to be unorthodox large compared to every other console released ever).

The 5700 XTX runs with 40 CUs at 1680 Mhz base clocks and 1980 Mhz at boost clocks but that's AMD highest binned Navi 10. The lowest binned Navi 10 runs at 1465 Mhz base and 1725 Mhz boost clock with 36 CUs. This configuration probably provides the best yield to performance ratio for Navi 10 and is more in line with what we would probably see from a hypothetical console (not talking PS5 or XSX) with a non binned Navi 10 based gpu at 36 CUs at N7. The Xbox X's gpu runs at 1172 Mhz with 36 CUs (cut down from 40). Compare that with AMD's lowest binned Polaris 20 which runs at 1168 Mhz at 32 CUs (cut down from 36).

You expect the move from N7 to N7P to provide a non binned 36 CU (cut down from 40) Navi (or RDNA 2)-based chip with a jump of 0.5 GHz in base clocks? That's almost a 40% jump in frequency. Not saying that can't happen but its a hell of a jump in base frequency especially considering we are talking an apu and not a discrete gpu.
 
Last edited:
The 5700 XTX runs with 40 CUs at 1680 Mhz base clocks and 1980 Mhz at boost clocks but that's AMD highest binned Navi 10. The lowest binned Navi 10 runs at 1465 Mhz base and 1725 Mhz boost clock with 36 CUs. This configuration probably provides the best yield to performance ratio for Navi 10 and is more in line with what we would probably see from a hypothetical console (not talking PS5 or XSX) with a non binned Navi 10 based gpu at 36 CUs at N7. The Xbox X's gpu runs at 1172 Mhz with 36 CUs (cut down from 40). Compare that with AMD's lowest binned Polaris 20 which runs at 1168 Mhz at 32 CUs (cut down from 36).

You expect the move from N7 to N7P to provide a non binned 36 CU (cut down from 40) Navi (or RDNA 2)-based chip with a jump of 0.5 GHz in base clocks? That's almost a 40% jump in frequency. Not saying that can't happen but its a hell of a jump in base frequency especially considering we are talking an apu and not a discrete gpu.

Xbox One X has 40 cut down from 44(believe the dev kit uses the full 44).
 
The Xbox X's gpu runs at 1172 Mhz with 36 CUs (cut down from 40). Compare that with AMD's lowest binned Polaris 20 which runs at 1168 Mhz at 32 CUs (cut down from 36).
Small correction here. Xbox One X is 40 CUs, cut down from 44.
 
Irrelevant.1 MADD == 1 MADD.
I don't say TF is a good unit to measure general compute performance, it is just the only unit we have.
Well, that's if you're simply comparing peak MADD potential in the CUs. When executing other operations, RDNA is faster. So, um...1 TF of MADD== 1 TF of MADD, yes, but 1 TF != 1TF in actual use where those FLOPS are other operations. If doing nothing but MADDs, 9.2 TF of RDNA will be exactly the same as 9.2 TF of GCN, but in real workloads RDNA can do more per trillion operations.

It's more a lousy naming convention where we count peak flops one way but talk about workloads another, using the same term. Perhaps we shouldn't count Teraflops but should count TMADDs as it's only that specific op being measured?
 
Well, that's if you're simply comparing peak MADD potential in the CUs.
Which is exactly what this number is telling.
It is NOT a measure of overall performance, so if some people use it as that, it is their mistake, not mine.
When executing other operations, RDNA is faster.
I'm listening...
So, um...1 TF of MADD== 1 TF of MADD, yes, but 1 TF != 1TF
No. TF is just about MADDs and nothing else. Correct me if i'm wrong.
If doing nothing but MADDs, 9.2 TF of RDNA will be exactly the same as 9.2 TF of GCN, but in real workloads RDNA can do more per trillion operations.
Yes, it CAN. It depends on the workloads and their implementations. Which brings us to noise an uncertainty. To prevent this, the TF number has been introduced, to have a measure that does not depend on any uncertain factors.

Again. The impression NAVI > GCN comes probably mainly from improved rasterization performance, resulting in much better game fps.
Why is a NAVI CU larger than GCN CU? More Simds? Nope. More LDS? Nope. Twice the ROPs? Yes. (IIRC)
So i'm sorry to inform you, but there is no indication about higher compute peak performance, aside the things i've said above.
If i'm wrong let me know, but you need to make a specific point.
 
The 5700 XTX runs with 40 CUs at 1680 Mhz base clocks and 1980 Mhz at boost clocks but that's AMD highest binned Navi 10. The lowest binned Navi 10 runs at 1465 Mhz base and 1725 Mhz boost clock with 36 CUs. This configuration probably provides the best yield to performance ratio for Navi 10 and is more in line with what we would probably see from a hypothetical console (not talking PS5 or XSX) with a non binned Navi 10 based gpu at 36 CUs at N7. The Xbox X's gpu runs at 1172 Mhz with 36 CUs (cut down from 40). Compare that with AMD's lowest binned Polaris 20 which runs at 1168 Mhz at 32 CUs (cut down from 36).

You expect the move from N7 to N7P to provide a non binned 36 CU (cut down from 40) Navi (or RDNA 2)-based chip with a jump of 0.5 GHz in base clocks? That's almost a 40% jump in frequency. Not saying that can't happen but its a hell of a jump in base frequency especially considering we are talking an apu and not a discrete gpu.
This was my thinking at beginning of the year as well, but hear me out.

1. Die space is premium - console manufacturers will want to push the clocks as far as they can this time around, as transistor count and density hasn't increased enough to be able to make that next gen jump without higher clocks.

2. 5700XT are Summer 2019 GPUs. They are first Navi GPUs on new node from AMD. Safe to say, GPUs coming out 1.5yr later on same (or gen2 N7P) process will make substantial improvements. I think we all can expect at least modest improvements in power efficiency with next batch of AMD GPUs next year, therefore consoles as well.

3. Console GPU released with equivalent PC GPUs on same node in a same year always had slight efficiency advantage. Why would we compare next winter's console GPUs with this summers PC parts?

Take for instance X GPU. Its 40CU RX580 equivalent with 88% of its boost clocks, but 10% more CUs. It peaks at around 180W while gaming, while RX580 card of ~TF alone pushes to almost 200W. This is just an example of generation where clocks and thermals haven't been pushed like we expect from next gen, therefore I can see 2.0GHz 36CU Navi coming out next winter in console that is pushing around 210W.
 
One thing is sure, AMD has really upped their game both in the CPU market, and they are about to do it in the GPU market too with full RDNA2, right now i'm thinking a Ryzen 3/12 core with a 16TF navi GPU might be a very nice combo, with PCI4 or 5. Seeing what they accomplish in consoles i'm not too far off.
 
The actual article is here:
https://www.eurogamer.net/articles/...laystation-5-xbox-series-x-spec-leak-analysed

Wccftech is only repeating and paraphrasing DF. They misuse the word "confirmed" too much. MS specifically refused to confirm 12TF despite saying already it's 2x the raw power of the 6TF. Not sure how math works when they confirm 2x 6TF and then refuse to answer when asked if it mean 12TF Navi. Maybe it's 11 or 13 or something... Or not yet final.
 
Last edited:
Even if the defect rates are under control, there are some implications to the TDP & clocks; between functional chips, there will be some variation in watt/frequency, which the power/circuitry or cooling/chassis designs need to consider (although I suppose the noise factor can be a "deal with it" situation).

What if retail PS5 has 40/40 CUs enabled, and clocks pushed to 2.15ghz to hit a 11TF target?

To me that seems pretty out there...
 
What if retail PS5 has 40/40 CUs enabled, and clocks pushed to 2.15ghz to hit a 11TF target?

To me that seems pretty out there...
If they wanted 40 they would have 44 with 4 disabled. There is zero chance there's no redundant CUs. It would cost more to throw away a lot of bad chips than just increasing the die size by a few percent from the start.

From past generations having different power gaps, I guess if they are within 20% of xbsx, they can probably sell at the same retail price, it won't be much noticeable and the first parties will look great. If it's 40% to 50% it should be okay at $100 less (imagine if xb1 launched at 299 against ps4 at 399, the launch sales would have been very different).

So far this seems to be 24% less (30% more) so I don't know, it's completely subjective since we don't have any visual reference of what the gap would produce. Unless other aspects like memory, or RT, or storage speed, compensate to create a different balance but equal perceived value.
 
Last edited:
RX 5700 XT has 20 / 20 WGP enabled, it sells for $330. The yield should be even better by late 2020.

I think MS was exploring this option for the XB1.
 
Status
Not open for further replies.
Back
Top