News Intel GPUs - Battlemage IGP benchmarks are here

[gaav87]

Good. I thought gimping the A750 with 8GB instead of 12GB minimum and having an 8GB A770 were poor decisions. Good that we see progress here.

24xe2 core 192EU 12GB 19GB/s 456GB/s 192-bit ~15TFLOPs of single precision. 1.8Ghz (current base clock) Gaming performance ~A750+50%. (Slower then rx6800 and rtx3070ti)
32xe2 nowhere to be seen. (est. perf between 4070 and 4070s ~6800xt)

 

[DAPUNISHER]

Hi guys enjoy:

https://intel-gfx-ci.01.org/tree/intel-xe/xe-pw-136821v1/shard-bmg-3/boot0.txt

BMG 12GB 19GB/s 456GB/s bandwidth 192-bit bus width
Actual physical size: 0x0000000300000000 =(12 GB)
[drm:xe2_hpd_get_bw_info.isra.0.constprop.0 [xe]] QGV 1: deratedbw=53000 peakbw: 456000
<7>[ 5.433883] xe 0000:03:00.0: [drm:icl_get_qgv_points.constprop.0 [xe]] QGV 1: DCLK=19000 tRP=0 tRDPRE=0 tRAS=0 tRCD=0 tRC=0

LMAO WCCF added you to the AT forums 15 mintutes of fame list - https://wccftech.com/intel-arc-batt...2-gpu-cores-12-gb-192-bit-bus-19-gbps-memory/

 

[gaav87]

LMAO WCCF added you to the AT forums 15 mintutes of fame list - https://wccftech.com/intel-arc-batt...2-gpu-cores-12-gb-192-bit-bus-19-gbps-memory/

Yes i know but its not 14xe2 cores.
Shard-bmg-3 is 20xe2 (G21) BMG-X2. It has 5xe2 cores "active" rest are "fused" off.
BMG G21 X2 (5x4x8) 20xe2 cores (160EU) ~270mm² on N4 / 19GB/s / 456GB/s / 192-bit
BMG X3 was (7x4x8) 28xe2 cores (224EU) 192bit / 12GB /19Gbps - possibly scrapped
BMG X4 was (4x10x8) 320eu /40core - scrapped
BMG G31 (not called X3) 32xe2 cores (256EU) ~400mm2 (<- source bionic)
My guess 256bit / 16GB / 19Gbps / 608GB/s ~22FP32

There is also bmg-4 with potential 16GB vram (ggtt mappable size: 16384 MiB).
The mappable size shows how much memory can be addressed through the GGTT it doesn’t directly reflect the amount of dedicated VRAM on the GPU as it can include system memory.

 

[jlnsrk]

Yes i know but its not 14xe2 cores.
Shard-bmg-3 is 20xe2 (G21) BMG-X2. It has 5xe2 cores "active" rest are "fused" off.
BMG G21 X2 (5x4x8) 20xe2 cores (160EU) ~270mm² on N4 / 19GB/s / 456GB/s / 192-bit
BMG X3 was (7x4x8) 28xe2 cores (224EU) 192bit / 12GB /19Gbps - possibly scrapped
BMG X4 was (4x10x8) 320eu /40core - scrapped
BMG G31 (not called X3) 32xe2 cores (256EU) ~400mm2 (<- source bionic)
My guess 256bit / 16GB / 19Gbps / 608GB/s ~22FP32

There is also bmg-4 with potential 16GB vram (ggtt mappable size: 16384 MiB).
The mappable size shows how much memory can be addressed through the GGTT it doesn’t directly reflect the amount of dedicated VRAM on the GPU as it can include system memory.

View attachment 105014

Do you have any information on power consumption you can share? Will it support the vvc codec?

 

[gaav87]

Do you have any information on power consumption you can share? Will it support the vvc codec?

H.266 (VCC) - yes (source ASUS event on Weibo)
base 1800mhz boost ~2500mhz (march/april sisofsandra data)
Do not have any info on power.

 

[511]

So LNL GPU is on par eith Strix Point but the 45W TDP kind of doesn't make sense as the Pl2 is 30W only i wonder if they had shown 30W what would have happened 🤣 also we don't know what was actual consumption

GPD Unveils Pocket 4, The First AMD Ryzen AI 300 "Strix" Handheld: Shares Radeon 890M "RDNA 3.5" & Arc 140V "Xe2" iGPU Benchmarks

GPD has unveiled the first AMD Ryzen AI 300 "Strix" handheld & also shared benchmarks of Radeon 890M against Intel Arc 140V "Xe2" iGPU.

wccftech.com

Found the source for the score apparently 17W TDP = 45W TDP GPD Maffs

 

[poke01]

So LNL GPU is on par eith Strix Point but the 45W TDP kind of doesn't make sense as the Pl2 is 30W only i wonder if they had shown 30W what would have happened 🤣 also we don't know what was actual consumption

GPD Unveils Pocket 4, The First AMD Ryzen AI 300 "Strix" Handheld: Shares Radeon 890M "RDNA 3.5" & Arc 140V "Xe2" iGPU Benchmarks

GPD has unveiled the first AMD Ryzen AI 300 "Strix" handheld & also shared benchmarks of Radeon 890M against Intel Arc 140V "Xe2" iGPU.

wccftech.com

Found the source for the score apparently 17W TDP = 45W TDP GPD Maffs
View attachment 105164

I wouldn’t trust information from GPD. They have to sell their silly handhelds so they do whatever to prop up the hx 370.

 

[gdansk]

And even if they weren't making up numbers TimeSpy performance doesn't seem to translate into great game performance.

 

[511]

And even if they weren't making up numbers TimeSpy performance doesn't seem to translate into great game performance.

That is especially true for Intel it just shows how good they can be if done right way larger delta with optimization vs non optimization scenarios

 

[mikk]

Found the source for the score apparently 17W TDP = 45W TDP GPD Maffs

It's more than 45W=17W.

They also lied with 268V because he tested an unspecified Pre-QS with 16GB whereas 268V is a 32 GB model. 32GB with Dual rank can be up to 10% faster. Also they lied with the timespy score of Strix Point prior to launch. They claimed it can do 4221 points at 54W.

 

[511]

It's more than 45W=17W.

They also lied with 268V because he tested an unspecified Pre-QS with 16GB whereas 268V is a 32 GB model. 32GB with Dual rank can be up to 10% faster. Also they lied with the timespy score of Strix Point prior to launch. They claimed it can do 4221 points at 54W.

Yeah we will se BMG in LNL on 3rd September

 

[Hans Gruber]

The other major consideration people are forgetting about with Intel discrete GPU's. The Intel playbook was always to move from TSMC silicon to either 20A or 18A down the road. The power efficiency of Battlemage (N4P) will be a big efficiency upgrade over N6. If Intel fixes their hardware mistakes with ARC and can deliver 4070 performance with the same power efficiency with either 12GB or 16GB of VRAM. They would have a winner if priced correctly.

Add in AMD RDNA 4 GPU's and Battlemage vs. Intel 40 series GPU's that were already on 4N eg. N4 but Nvidia specialty silicon. Keep an eye on power efficiency this generation. We will see if it's chip design or silicon that provided Nvidia amazing power efficiency. The 4060 was not well received but the power efficiency was amazing. I personally think it's the silicon that provided the power efficiency for Nvidia.

The top Battlemage GPU should see real world power use of 170w-190w under full load. That is right about where the 4070 power usage is.

 

[511]

The other major consideration people are forgetting about with Intel discrete GPU's. The Intel playbook was always to move from TSMC silicon to either 20A or 18A down the road. The power efficiency of Battlemage (N4P) will be a big efficiency upgrade over N6. If Intel fixes their hardware mistakes with ARC and can deliver 4070 performance with the same power efficiency with either 12GB or 16GB of VRAM. They would have a winner if priced correctly.

Add in AMD RDNA 4 GPU's and Battlemage vs. Intel 40 series GPU's that were already on 4N eg. N4 but Nvidia specialty silicon. Keep an eye on power efficiency this generation. We will see if it's chip design or silicon that provided Nvidia amazing power efficiency. The 4060 was not well received but the power efficiency was amazing. I personally think it's the silicon that provided the power efficiency for Nvidia.

The top Battlemage GPU should see real world power use of 170w-190w under full load. That is right about where the 4070 power usage is.

Yeah but Nvidia Jumped from Samsung 8N to Custom 4N which is like 1.5 real node jump unlike the meager one we will see from now on

 

[Hans Gruber]

Yeah but Nvidia Jumped from Samsung 8N to Custom 4N which is like 1.5 real node jump unlike the meager one we will see from now on

I think N4P will be as efficient or within 1-2% of 4N. The 50 series is on some kind of 4NX type stuff that has added silicon density. Not sure if it's much more efficient than 4N was on the 40 series. Look at the nice efficiency boost for Zen 5 using N4P silicon. That's going from N5 to N4P. Battlemage will be going from ARC N6 (7nm) all the way to N4P. That is also a generation and a half silicon jump.

Because the N4P stuff is quite good. I do not see Intel in a rush to move to 20A or 18A for the discrete GPU's. Maybe Celestial will use Intel silicon.

 

[511]

I think N4P will be as efficient or within 1-2% of 4N. The 50 series is on some kind of 4NX type stuff that has added silicon density. Not sure if it's much more efficient than 4N was on the 40 series. Look at the nice efficiency boost for Zen 5 using N4P silicon. That's going from N5 to N4P. Battlemage will be going from ARC N6 (7nm) all the way to N4P. That is also a generation and a half silicon jump.

Because the N4P stuff is quite good. I do not see Intel in a rush to move to 20A or 18A for the discrete GPU's. Maybe Celestial will use Intel silicon.

Celestial has some Intel 3 + N3E Sillicon

 

[gaav87]

Lnl with 7 cores (active subslices) and 8 l3 cache banks

BMG with 20 cores and 24 l3 cache banks




Anyone ? bmg G21 is 20 cores or 24 cores (with 4 cores disabled for tests) or am i just Ret? Why would it have more l3 banks then cores ?

 

[DavidC1]

Anyone ? bmg G21 is 20 cores or 24 cores (with 4 cores disabled for tests) or am i just Ret? Why would it have more l3 banks then cores ?

I think BMG went through multiple revisions hence the delay into 2025 for the most part. They were talking about using insane 100MB+ L3 cache for Battlemage, and even before that they were talking about quad tile graphics. I wouldn't be too surprised if G31 is basically BTS 2025 now. Won't change the game significantly.

I gave another thought about Pat's comment regarding ARC drivers. He said they thought reusing the iGPU drivers would have been enough. That means him and entire high level management division had zero clue about what they needed. You know, despite the fact until 2011 or so their vision on graphics was essentially HD Audio but for graphics? Takes near zero silicon and no one cares to buy it?

Recently he said Larrabbee cancellation was a mistake and they could have been a leader. Larrabbee was in development before 2010. If they had no idea in 2022, how would they have got it 12 years prior? I have to wonder he now has a software mindset after being out of Intel, or that his skill limit floor is a top level engineer, not a CEO.

I hope they stick with GPUs, and actually soldier on for once.

 

[ToTTenTranz]

Wait, so Battlemage is confirmed to be delayed into 2025?

 

[igor_kavinski]

No, at least one SKU (maybe the A770 counterpart) is supposed to be released before year end. 2025 for the rest of the line-up including their flagship BM part.

 

[gaav87]

Wait, so Battlemage is confirmed to be delayed into 2025?

No they are already on A0 QS samples for at least 4months G21 launch is imminent linux drivers are almost ready.

 

[ToTTenTranz]

No, at least one SKU (maybe the A770 counterpart) is supposed to be released before year end. 2025 for the rest of the line-up including their flagship BM part.

But isn't the A770 counterpart the one with RTX 4070 Ti performance? Or is that the 2025 flagship?

 

[igor_kavinski]

But isn't the A770 counterpart the one with RTX 4070 Ti performance? Or is that the 2025 flagship?

2025 flagship.

 

[Aapje]

A bit chaotic, but some info about Battlemage, which is apparently going to be sold as Xe2.

 

[gaav87]

But isn't the A770 counterpart the one with RTX 4070 Ti performance? Or is that the 2025 flagship?

FWIK There is no 4070 ti performance card anymore. Perf est. on 32xe2-core variant is only ~4070/6800xt/7800xt [A770+50%]. But there is 0 real life perf leaks for this card.

 

[Tup3x]

A bit chaotic, but some info about Battlemage, which is apparently going to be sold as Xe2.

Deeper version.