5900X GB5 = 1605 according to GB5 Browser result. 3900X GB5 = 1280 per GB5 benchmark chart. (Presumed we can get 100+ more with VAES/other updates but let's just use these numbers.)
Part 1
Let's assume that SPECint2006 1T scales along with GB5 single-core. Known GB5/SPECint2006 results are in
bold, projected results in
italics.
A14:.......
GB5 1586,
SPECint2006 63.13 @ 3.0 GHz, SPEC/GHz = 21.04
A13:.......
GB5 1327, SPECint2006 52.82 @ 2.66 GHz, SPEC/GHz = 19.86
5900X:...
GB5 1605,
SPECint2006 62.72 @ 4.95 GHz, SPEC/GHz = 12.67
1065G7:.GB5 ?????,
SPECint2006 47.40 @ 3.9 GHz, SPEC/GHz = 12.15
3900X:...
GB5 1280, SPECint2006 50.02 @ 4.6 GHz, SPEC/GHz = 10.87
9900K:...
GB5 1334, SPECint2006 54.28 @ 5.0 GHz, SPEC/GHz = 10.86
10900K:.
GB5 1412,
SPECint2006 57.45 @ 5.3 GHz, SPEC/GHz = 10.84
* I scaled by architecture - A14 inferred SPEC score = (A13 SPEC score * A14 GB5 score) / A13 GB5 score. The 9900K/10900K also grouped, and 3900X/5900X grouped in a similar way.
First, perusing the top CPU charts for raw single-core GB5 results, you see the 3800XT at 1354 (4.7 GHz), 3800X at 1291 (4.5 GHz). This correlates to a miniscule 0.4% GB5 per GHz gain with the XT chips over X chips, with my inference being that this reflects binning rather than any actual process performance/efficiency gains. This is important because Zen3 uses the same process, hence we can infer that any GB5 per GHz gain from Zen2/2+ to Zen3 would be from core/CCX redesign rather than process improvements. In single-core GB5, the 5900X shows a 25.4% gain in raw GB5 score, and a 16.5% gain in GB5 score per GHz. This is lower than the claimed 19% IPC and likely reflects the algorithm and benchmark specific issues as discussed by Hans de Vries and others above.
What's remarkable here is not just the 25.4% raw gain and 16.5% IPC gain from Zen2 to Zen3. What's remarkable is that as generationally amazing as Apple's chips have been, A13 to A14 gains only 19.5% raw and 6.0% IPC between generations.
What AMD have done here is exceeding even Apple's advancement. And remember - some of Apple's IPC gain is coming from process change. AMD has NO process change. So AMD is all design, Apple is part design part process.
Part 2
Since scaling with GB5 is inference, why not just use manufacturer claims? Even if they might be slightly (or wildly) different than our results...
Now, if we simply assume (and give lots of leeway to AMD and Apple on their claims) that:
- Zen3 will gain 19% in IPC over Zen2
- A14 will gain 17% in raw performance over A13*
* Apple said A14 is +40% of A12, and that A13 was +20% on A12, so we can assume A14 is +17% A13 from their marketing
Then the chart looks like this (A13, A14, 3900X, and 5900X only):
A14:.......
SPECint2006 61.80 @ 3.0 GHz, SPEC/GHz = 20.60
A13:.......
SPECint2006 52.82 @ 2.66 GHz, SPEC/GHz = 19.86
5900X:...
SPECint2006 64.05 @ 4.95 GHz, SPEC/GHz = 12.94
3900X:...
SPECint2006 50.02 @ 4.6 GHz, SPEC/GHz = 10.87
SPEC/GHz (IPC) generational gains:
A13 -> A14 = +3.7%
Zen2 -> Zen3 = +19%
I'd like to point out that Apple's actual performance gains have fit - with clock speed increase and process improvements, they've been able to squeeze a 19.8% raw performance increase between A12 -> A13 and it appears a 19.5% raw performance increase between A13 -> A14, for a total actual raw performance increase of 43.1%. This matches or exceeds their performance claims.
Part 3
What Apple has done with performance efficiency is remarkable, but it appears AMD is making even better generational leaps than they have, if it plays out this way...
From A12 -> A14, half of Apple's 43.1% performance gains have come solely from clock speed increases (20.5% increase in clocks from 2.49 GHz A12 -> 3.0 GHz A14). That means that about 18.8% came from design changes or process improvements. N7 (A12) -> N5 (A14) is +15% speed at isopower, hence we can infer that possibly as little as 3.8% of the raw performance increase came from actual core / uarch changes over the last two years. Well, that's a low-end estimate. I'm being lazy by not taking into account possible reduction in performance gains in order to reduce power draw since they also increased clocks, but I don't feel qualified to closely examine these things, nor do I know how to scale performance vs power on node changes to figure out exactly how much performance and how much power savings from the N7 -> N5 change Apple decided to use -- particularly since we have no actual chip review out yet for A14 regarding details. However, we do have a history with A12 -> A13 that Apple increased power usage on the Lightning core over Vortex core by 28.4% and increased clock speeds 6.8% for a 16.5% SPECint gain. Since N7 -> N7P provides 7% isopower performance gains, we can infer that Apple didn't use the efficiency gains of the process change, but rather the performance gains, pushing power consumption up by 28.4% for a 16.5% performance gain. Clocks and process improvements account for ~14% of a 16.5% SPECint2006 performance gain, leaving little room in the performance gains for actual uarch/core improvements; I wouldn't expect things to be much different on A13 -> A14.
Meanwhile, in a single generation from Zen2 -> Zen3, AMD performance gains in single-core GB5 (totaling 25.4%) are 7.6% from clockspeed increases (4.6 GHz -> 4.95 GHz) and approximately 16.5% from design changes or process improvements. Since the process no Zen2 and Zen3 is the same, that means 16.5% of the performance improvement is from core / uarch changes.
Food for thought, huh?
Facebook
Twitter