Discussion Intel current and future Lakes & Rapids thread

[mikk]

Yes, but it matters a lot on mobile. Trifecta of battery life, graphics, and CPU wins coming with Remembrandt. Alderlake may be a big advancement, but will face stiff competition at the very least.

Intel is also slow on driver updates even with Xe.

Can we have more consistent execution please? It's impressive how AMD is doing that on both areas. It feels like AMD executing well is better looking at the big picture.

My point is that the performance advantage over Vega 7/8 currently is a bit masked because of the bandwidth limits. Vega 7/8 does not gain as much from 4266 over 3200. And yes they are slow with drivers but they have added speed improvements for Xe already:

 

[exquisitechar]

Yes, but it matters a lot on mobile. Trifecta of battery life, graphics, and CPU wins coming with Remembrandt. Alderlake may be a big advancement, but will face stiff competition at the very least.

Intel is also slow on driver updates even with Xe.

Can we have more consistent execution please? It's impressive how AMD is doing that on both areas. It feels like AMD executing well is better looking at the big picture.

I think Intel’s execution has been pretty decent, they delivered/will deliver with ICL/TGL/ADL. ADL will be better than Rembrandt in terms of CPU performance. If anything, it’s AMD that needs to pick up the pace with Zen4 being so late, especially in laptops. Intel’s graphics still need more work for sure, though. Like uzzi38 said, the 7nm delays didn’t help there.

 

[uzzi38]

My point is that the performance advantage over Vega 7/8 currently is a bit masked because of the bandwidth limits. Vega 7/8 does not gain as much from 4266 over 3200. And yes they are slow with drivers but they have added speed improvements for Xe already:

It won't be enough. With the move to RDNA2 iGPUs (combined with LPDDR5) with Rembrandt, we're talking about a potential 80%+ improvement to performance. RMB comes with 12CUs vs the 8 on Cezanne/Renoir, with significant per CU performance improvements (the 40CU 5700XT performed the same as the 60CU Radeon VII despite both clocking the same as one another out of the box - around 1750MHz) whilst consuming less power per-CU.

80% is the baseline for my expectations assuming there is no Infinity Cache on die assuming they gain as much performance as they gain in additional bandwidth and how well they used that bandwidth using the 5700XT and Vega64 as comparison points (25% higher perf with 10% lower bandwidth). Having any level of Infinity Cache on die - even just 16MB - will be enough to propel performance past the 100% gain gen on gen (with the help of additional clocks, something that isn't an issue with RDNA2 as the dGPUs have shown).

 

[mikk]

It won't be enough. With the move to RDNA2 iGPUs (combined with LPDDR5) with Rembrandt, we're talking about a potential 80%+ improvement to performance. RMB comes with 12CUs vs the 8 on Cezanne/Renoir, with significant per CU performance improvements (the 40CU 5700XT performed the same as the 60CU Radeon VII despite both clocking the same as one another out of the box - around 1750MHz) whilst consuming less power per-CU.

80% over what? Over the current Vega? This might be not enough to beat ADL-P DDR5/LPDDR5.

 

[uzzi38]

80% over what? Over the current Vega? This might be not enough to beat ADL-P DDR5/LPDDR5.

Yes. And of course that would be enough to win in pure performance, Tiger Lake doesn't hold that much of a lead over Cezanne/Renoir as is (usually about 25-30% over Renoir, but the extra clocks on Cezanne helps claw back ~5% depending on the title). It wouldn't be a massive lead, sure, more like a reversal of the current situation, but it'll be plenty to take back the crown.

 

[lobz]

I think Intel’s execution has been pretty decent, they delivered/will deliver with ICL/TGL/ADL. ADL will be better than Rembrandt in terms of CPU performance. If anything, it’s AMD that needs to pick up the pace with Zen4 being so late, especially in laptops. Intel’s graphics still need more work for sure, though. Like uzzi38 said, the 7nm delays didn’t help there.

Uhmm, decent how? They've delayed evety single one of their products multiple times ever since they failed to set up realistic design goals for 10nm. Their execution as of late has been the exact opposite of decent. Moreover, you speak of ADL vs Rembrandt as if you _knew_ how they'd fare against each other. How could you possibly _know_ that? From subjective and cumulatively indirect comparisons of 1st party 'up to*' figures?

I say that because none of us knows exactly, what and how AMD plans with Rembrandt/Zen 3+/6nm or 7nm etc. You just can't KNOW. You could guess pretty well in general, against Cézanne though, but that's it.

There is the possibility, of course, that you're just joking, in which case I apologize 🙂

 

[mikk]

Yes. And of course that would be enough to win in pure performance, Tiger Lake doesn't hold that much of a lead over Cezanne/Renoir as is (usually about 25-30% over Renoir, but the extra clocks on Cezanne helps claw back ~5% depending on the title). It wouldn't be a massive lead, sure, more like a reversal of the current situation, but it'll be plenty to take back the crown.

Tigerlake has a 20-30% speed advantage over Vega and I believe an 50% uplift is doable with a combination of GPU clock speed increases, DDR5/LPDDR5, driver optimizations. Keep in mind Iris Xe is more bandwidth starved than Vega 7 or 8, the current advantage of 20-30% is a bit masked. 80% over the current Vega probably won't be enough against ADL-P.

 

[uzzi38]

Tigerlake has a 20-30% speed advantage over Vega and I believe an 50% uplift is doable with a combination of GPU clock speed increases, DDR5/LPDDR5, driver optimizations. Keep in mind Iris Xe is more bandwidth starved than Vega 7 or 8, the current advantage of 20-30% is a bit masked. 80% over the current Vega probably won't be enough against ADL-P.

Were bandwidth an issue for Tiger Lake as it's shipping, Intel would reel back clocks to save power as they already have a clear lead in graphics performance. There's no reason to clock the iGPU higher. It's safe to assume that LPDDR4X-4266 is enough to feed 96EUs @1.3GHz.

Anyways, you do realise you're quoting almost identical numbers to my expectations as I've stated them, but stating that Intel are keeping up, right? 1.8x uplift over Vega is exactly 20% over the 1.5x you're predicting for ADL-P, which is as I mentionned earlier, effectively a role reversal from the current generation.

 

[lobz]

Silly double post

 

[jpiniero]

Considering Gen13 is only rumoured to appear alongside Meteor Lake, coasting is probably the right word. Progress is just too slow. I know that's partially down to 7nm delays, but still.

ADL-P is getting the updated display engine.

 

[mikk]

Were bandwidth an issue for Tiger Lake as it's shipping, Intel would reel back clocks to save power as they already have a clear lead in graphics performance. There's no reason to clock the iGPU higher. It's safe to assume that LPDDR4X-4266 is enough to feed 96EUs @1.3GHz.

It is an "issue", it has been confirmed in many tests. Iris Xe gains a crazy 10% just from switching single rank 3200 to dual rank 3200, it speeds up by another 20% from LP 4266 over dual rank 3200 or 30% over single rank 3200. Without this limitation they might have clocked Iris Xe higher which they will surely do on ADL-P, we have seen ES samples with 1.5 Ghz already. Most DDR4 dualchannel devices are single rank, this is quite a bottleneck at the moment. That's why many DDR4 Iris Xe devices are rather low scoring in 3dmark. We may see Tigerlake refresh with LPDDR5 5400 support, this will be interesting.

 

[TESKATLIPOKA]

I think Intel’s execution has been pretty decent, they delivered/will deliver with ICL/TGL/ADL. ADL will be better than Rembrandt in terms of CPU performance. If anything, it’s AMD that needs to pick up the pace with Zen4 being so late, especially in laptops. Intel’s graphics still need more work for sure, though. Like uzzi38 said, the 7nm delays didn’t help there.

ADL vs Rembrandt
In single threaded performance ADL should be faster, but It's questionable by how much considering Intel mentioned Up to 20% performance uplift, which is most likely the best case scenario in an unknown test, and we don't know against what(Tiger Lake or Rocket Lake?).
The problem is multi thread in my opinion at least for 15W.
For 15W TDP Intel is planning only 2+8 which won't be enough against an 8C Rembrandt.
For 28-45W TDP Intel has 6+8 which should be enough to be more powerful than an 8C from AMD, and It doesn't look like AMD plans to release anything with more cores for mobile until Zen4.
For 55W TDP there is 8+8, this will be the new king for mobile multi threaded performance.
This all depends on If ADL's HW scheduler will be effective in using both big and small cores for best performance and If there will be a need for software support for ADL, like It was needed for Lakefield.

I personally find this approach of using big+small cores interesting. I understand Intel chose this approach, because of 10nm process. We will see If It was a good move or not.

P.S. IGP could have been 128EU.

 

[andermans]

I think a lot of the extra performance of the Xe IGP on mobile Tiger Lake compared to the Vega of Renoir/Cezanne is that they just went for +50% ALUs. I expect Rembrandt to catch up to that as the current leaks put it at 8->12 CUs.

That said, that for the U line (15W) that shouldn't result in +50% perf as things are mostly power/thermals limited. I also doubt that even Intel has much room to improve there just by reclocking without significant efficiency work. With the current clocks you're mostly already taking 30W+ to reach the top clocks.

Above 30W+ clocking really turns interesting, but a lot of laptops with that also have discrete graphics. Before that it is really about perf/watt at various power budgets.

 

[IntelUser2000]

Tigerlake has a 20-30% speed advantage over Vega and I believe an 50% uplift is doable with a combination of GPU clock speed increases, DDR5/LPDDR5, driver optimizations.

It might be possible, but I feel that Alderlake slide would have pointed out the GPU if it resulted in 50% faster performance.

Also Remembrandt is a 5nm part. Vega to RDNA is 50% perf/watt gain, and RDNA2 is another 50% gain, but that includes the jump from 14nm to 7nm, while 7nm to 5nm will be a bit less.

If they really achieve the 1.5x per generation like they did with the dGPUs, it'll be faster than Alderlake by quite a bit, even with the 50% improvement. Not to mention if they include 50% extra CUs on top of that.

I think a lot of the extra performance of the Xe IGP on mobile Tiger Lake compared to the Vega of Renoir/Cezanne is that they just went for +50% ALUs. I expect Rembrandt to catch up to that as the current leaks put it at 8->12 CUs.

Yes but Renoir also clocks substantially higher. It almost makes up for the ALU differences.

 

[mikk]

I think a lot of the extra performance of the Xe IGP on mobile Tiger Lake compared to the Vega of Renoir/Cezanne is that they just went for +50% ALUs. I expect Rembrandt to catch up to that as the current leaks put it at 8->12 CUs.

That said, that for the U line (15W) that shouldn't result in +50% perf as things are mostly power/thermals limited. I also doubt that even Intel has much room to improve there just by reclocking without significant efficiency work. With the current clocks you're mostly already taking 30W+ to reach the top clocks.

Above 30W+ clocking really turns interesting, but a lot of laptops with that also have discrete graphics. Before that it is really about perf/watt at various power budgets.

Vega 8 seems to deal better with the limited bandwidth, they may have the better compression and keep in mind Vega 8 iGPU clocks up to 2000 Mhz, means Iris Xe is low clocked in comparison. They should get more headroom with enhanced SuperFin and the GPU doesn't need 30W+ to reach the top clocks, Intels CPU Turbo is too aggressive in 3d workloads. Actually I get a slightly better iGPU performance with disabled CPU turbo, also the device runs much cooler. It's GPU limited in a typical iGPU gaming scenario unless it's Counterstrike with 100+ fps, CPU turbo kicks in for no real benefit.

 

[jpiniero]

Rembrandt is N6.

 

[IntelUser2000]

Vega 8 seems to deal better with the limited bandwidth, they may have the better compression and keep in mind Vega 8 iGPU clocks up to 2000 Mhz, means Iris Xe is low clocked in comparison.

Actually Hardware Unboxed tests show while LPDDR4x benefits the Iris Xe G7 a lot at 15W, the differences aren't as profound under 28W:

Rembrandt is N6.

Ok thanks. I'm still on the side Remembrandt will kick ass on the iGPU side as there is nothing exciting on the horizon post-Xe.

For 15W TDP Intel is planning only 2+8 which won't be enough against an 8C Rembrandt.

In an ideal scenario, 2+8 will be far more competitive than Tigerlake 4C is against Cezanne today.

Right now the top Renoir part is 30-50% faster than the fastest Tigerlake configuration.

 

[eek2121]

Nothing spectacular, but I noticed the tiger lake NUC made an appearance: https://browser.geekbench.com/v5/cpu/7162029

 

[andermans]

Actually Hardware Unboxed tests show while LPDDR4x benefits the Iris Xe G7 a lot at 15W, the differences aren't as profound under 28W:

I wonder if the LPDDR4X advantage under 15W is not due to higher bandwidth, but due to lower power usage & more power saving features of LPDDR4X compared to DDR4, giving a higher budget to the rest of the SoC.

 

[IntelUser2000]

I wonder if the LPDDR4X advantage under 15W is not due to higher bandwidth, but due to lower power usage & more power saving features of LPDDR4X compared to DDR4, giving a higher budget to the rest of the SoC.

I thought of that too, but I'm not sure. The mobile DDR4 uses at max 2W, and typically at the low 1W level. LPDDR saves ~15% power when active which is nothing compared to 15W for the CPU.

 

[andermans]

I thought of that too, but I'm not sure. The mobile DDR4 uses at max 2W, and typically at the low 1W level. LPDDR saves ~15% power when active which is nothing compared to 15W for the CPU.

Yeah, not sure about that, but at 15W you're clearly not going to be more bandwidth starved than 28W with the same memory so either there is some memory(/subsystem) reclocking being suboptimal (power management is hard) or the power budget changes are significant. Not sure what else one could attribute it to.

 

[IntelUser2000]

Yeah, not sure about that, but at 15W you're clearly not going to be more bandwidth starved than 28W with the same memory so either there is some memory(/subsystem) reclocking being suboptimal (power management is hard) or the power budget changes are significant. Not sure what else one could attribute it to.

I agree. Based on the numbers I cannot conclude Iris Xe is very memory bandwidth bound. They were at one point, with the 40EU Iris versions using Gen 8 graphics without eDRAM. But ever since Gen 9 it wasn't really sensitive.

The 28W DDR4 significantly outperforms the 15W LPDDR4X, yet the 28W LPDDR4X isn't much faster.

Also the consensus is that the Iris Xe fares significantly better against Vega 8 at 28W then it does at 15W. At 15W it just falls behind, while at 28W it can be noticeably faster.

Possibilities:
-Vega 8 is too small so it has little room to stretch with extra TDP
-Intel CPU cores are power hungry and takes away budget from the GPU
-Tigerlake's design point is at 25W+
-Xe scales more linearly with extra resources

Alderlake can improve on #2, #1 means 12 CU or more on Remembrandt will be awesome, #4 means dGPU may not be terrible.

 

[andermans]

I think a big factor in #2 is the power management firmware giving more power to the CPU side than needed for optimal gaming performance. I think for both AMD and Intel systems you can improve gaming performance by biasing power to the GPU (by limiting CPU frequency or trying to force GPU clocks).

It should be mostly software so I believe it is fixable, but the question is whether it will be improved.

 

[IntelUser2000]

It should be mostly software so I believe it is fixable, but the question is whether it will be improved.

Renoir cores are impressively efficient. There's a video on twitter with Renoir getting 600 Cinebench R15 MT score without fan or heatsink.

At all clock levels Tigerlake CPU uses more power.

The fact that AMD gains the advantage at low power levels may also be an indicator that Intel's 10nm SF is still behind TSMC 7nm in terms of perf/watt.

 

[tamz_msc]

I thought that Intel Xe scored higher in synthetic tests like 3DMark than in actual games. Intel has still got a long way to go with their drivers though. There are frame pacing issues, visual glitches and straight up unplayable games when it comes to Iris Xe.