Alchemist needs to get a proper launch first, imo.

IF this happens, and IF Intel can improve the drivers further, then this will eat nGreedias breakfast in the low to mid-range.

Could 2024 be the year competition in the GPU market may finally bear fruit for the consumer? Because it's been a bad joke so far.

I hope the A.I. bubble bursts soon, as all the big tech companies are on record saying that they hate kowtowing to nGreedia for A.I. Maybe then that will finally bring consumer GPUs back to being "all about the gamers", and the prices back to sanity.

AusWolfAlchemist needs to get a proper launch first, imo.

Disagree. Alchemist isn't terrible, but it's just barely competitive. Efficiency is it's Achilles' heel: Alchemist cards require an outsize amount of silicon and power to reach performance parity with the other manufacturers. The sooner they can launch something that closes that gap, the better, not least for Intel. Alchemist margins have got to be tiny.

80-watt HamsterDisagree. Alchemist isn't terrible, but it's just barely competitive. Efficiency is it's Achilles' heel: Alchemist cards require an outsize amount of silicon and power to reach performance parity with the other manufacturers. The sooner they can launch something that closes that gap, the better, not least for Intel. Alchemist margins have got to be tiny.

That's not what I meant. By "proper launch", I meant, it has to be available. In the UK, there is still only one store that sells Arc GPUs, and the only AIB we have is Asrock.

AusWolfThat's not what I meant. By "proper launch", I meant, it has to be available. In the UK, there is still only one store that sells Arc GPUs, and the only AIB we have is Asrock.

Would they have much to gain, though? They'd be pushing a mediocre product with poor margins. Were I Intel, delaying a broader launch until the more competitive offering is ready would seem like the better play.

80-watt HamsterWould they have much to gain, though? They'd be pushing a mediocre product with poor margins. Were I Intel, delaying a broader launch until the more competitive offering is ready would seem like the better play.

I see your point. Although, a bigger mind share wouldn't hurt, in my opinion.

Hi,
Can't be worse than the first release can it :fear:

MrMilliI think Intel will easily double its current performance. N5 alone will be enough to double the Xe cores to 64.
Hopefully Battlemage won't be just a simple die shrink.
If they fix one of these issues then they will be competitive:
- horrible cache/memory latency/bandwidth
- core dependency on high parallelization

I'm also hoping for more than a die shrink, but the jump from N6 to N5 isn't as great as you think. Let's take a look at four AMD GPUs to estimate what a N6 to N5 transition could do. The four GPUs are:

• RX 7600: N6, 32 CUs, 128 bit DRAM bus, 32 MB LLC, 204 mm^2 die size
• RX 6600 XT: N7, 32 CUs, 128 bit DRAM bus, 32 MB LLC, 237 mm^2 die size
• RX 6700 XT: N7, 40 CUs, 192 bit DRAM bus, 96 MB LLC, 335 mm^2 die size
• RX 7800 XT GCD: N5, 60 CUs, 200 mm^2 die size

Estimates for the die area of various blocks in the 6700 XT are: 58.46 mm^2 for the 96 MB LLC, 9.53 mm^2 for the 3 MB L2, 29.09 mm^2 for the 192-bit GDDR6 PHY, and 4.6 mm^2 for 16 PCIe 4 lanes. Extrapolating this to the 6600 XT and the 7600 yields 19.49 mm^2 for the 32 MB LLC, 6.35 mm^2 for the 2 MB L2, 19.39 mm^2 for the 128-bit GDDR6 PHY, and 2.3 mm^2 for 8 PCIe 4 lanes. This leaves 156.5 mm^2 for the 7600 and 189.5 mm^2 for the 6600 XT sans their L2, LLC and off chip IO PHYs. This yields a 21% increase in density going from N7 to N6. Now, it's unlikely to be the process alone; design probably plays a part too. Note that we haven't accounted for the larger media engine of the 7600.

Using the same process to estimate the die area of a 60 CU 6700 XT sans the same blocks gives us 362.66 mm^2 for N7 which is 299.5 mm^2 for N6. We haven't accounted for the larger register file in RDNA3 which would make the scaling better, but this works as a starting point. This suggests that Intel is unlikely to manage anything more than 48 EUs for Battlemage in a die of the same size as the one used for the A770.

AnotherReaderI'm also hoping for more than a die shrink, but the jump from N6 to N5 isn't as great as you think. Let's take a look at four AMD GPUs to estimate what a N6 to N5 transition could do. The four GPUs are:

• RX 7600: N6, 32 CUs, 128 bit DRAM bus, 32 MB LLC, 204 mm^2 die size
• RX 6600 XT: N7, 32 CUs, 128 bit DRAM bus, 32 MB LLC, 237 mm^2 die size
• RX 6700 XT: N7, 40 CUs, 192 bit DRAM bus, 96 MB LLC, 335 mm^2 die size
• RX 7800 XT GCD: N5, 60 CUs, 200 mm^2 die size

Estimates for the die area of various blocks in the 6700 XT are: 58.46 mm^2 for the 96 MB LLC, 9.53 mm^2 for the 3 MB L2, 29.09 mm^2 for the 192-bit GDDR6 PHY, and 4.6 mm^2 for 16 PCIe 4 lanes. Extrapolating this to the 6600 XT and the 7600 yields 19.49 mm^2 for the 32 MB LLC, 6.35 mm^2 for the 2 MB L2, 19.39 mm^2 for the 128-bit GDDR6 PHY, and 2.3 mm^2 for 8 PCIe 4 lanes. This leaves 156.5 mm^2 for the 7600 and 189.5 mm^2 for the 6600 XT sans their L2, LLC and off chip IO PHYs. This yields a 21% increase in density going from N7 to N6. Now, it's unlikely to be the process alone; design probably plays a part too. Note that we haven't accounted for the larger media engine of the 7600.

Using the same process to estimate the die area of a 60 CU 6700 XT sans the same blocks gives us 362.66 mm^2 for N7 which is 299.5 mm^2 for N6. We haven't accounted for the larger register file in RDNA3 which would make the scaling better, but this works as a starting point. This suggests that Intel is unlikely to manage anything more than 48 EUs for Battlemage in a die of the same size as the one used for the A770.

These things are really hard to estimate but if I may add:
AMD and it's infinity cache might not be the best reference point considering Intel might have less cache on board (and cache doesn't scale as well) but if we look at the 6700XT and 6800XT, then we see that they have around 51M transistors per mm². I don't think we can use any N5 RDNA3 product considering the memory controller (doesn't scale well) isn't on die (resulting in 150.2M / mm²).
If we look at the 4080, then we see that it has a realistic 121M transistors per mm².
Alchemist has 54M / mm² which is slightly better than AMD's N7 (Intel probably didn't do the best job, still.. ) but if we compare this to the 4080, then we see that it has more than double the density. The 4080 has comparable cache and memory configuration. So yeah, Alchemist's current 406 mm² could potentially be less than 200 on N5. That's why I said that it will easily have 64 Xe cores.

MrMilliThese things are really hard to estimate but if I may add:
AMD and it's infinity cache might not be the best reference point considering Intel might have less cache on board (and cache doesn't scale as well) but if we look at the 6700XT and 6800XT, then we see that they have around 51M transistors per mm². I don't think we can use any N5 RDNA3 product considering the memory controller (doesn't scale well) isn't on die (resulting in 150.2M / mm²).
If we look at the 4080, then we see that it has a realistic 121M transistors per mm².
Alchemist has 54M / mm² which is slightly better than AMD's N7 (Intel probably didn't do the best job, still.. ) but if we compare this to the 4080, then we see that it has more than double the density. The 4080 has comparable cache and memory configuration. So yeah, Alchemist's current 406 mm² could potentially be less than 200 on N5. That's why I said that it will easily have 64 Xe cores.

I agree that these things are really hard to estimate and that RDNA2 is very cache rich compared to most GPUs. That's why I removed the off chip IO and Infinity cache before doing the estimation. We can try this another way. We know that logic density increases by about 18% from N7 to N6 and about 80% from N7 to N5. This means that logic density increases by about 55% from N6 to N5 and cache density increase should be the same as N7 to N5, i.e. 35%. It's really hard to see a path to 64 EUs for Battlemage unless Intel screwed up with the layout of Alchemist. Intel is very good at physical design so it's very unlikely that the A770 was deficient in that metric; even clock speeds matched what AMD could do with RDNA2 on N7.

Don't worry; Moore's Law is Dead will put out several unsubstantiated claims to the contrary.

REDSLATEDon't worry; Moore's Law is Dead will put out several unsubstantiated claims to the contrary.

He's recommending them now, due to how much the drivers have improved.

I've actually watched his videos on Arc - his videos were talking about how Intel was looking to axe the entire division. At the time, he stated that it was up in the air. People on the internet extrapolated what he actually said into something else.

I'd also point out that a lot of his coverage turned out to be accurate. The Alchemist refresh never happened, and the leaked slides aren't showing a full release stack, just 2 discrete gpus. OTOH, if the top die gives 4070ti gaming performance - that covers 99% of the market.

SSGBryanHe's recommending them now, due to how much the drivers have improved.

I've actually watched his videos on Arc - his videos were talking about how Intel was looking to axe the entire division. At the time, he stated that it was up in the air. People on the internet extrapolated what he actually said into something else.

I'd also point out that a lot of his coverage turned out to be accurate. The Alchemist refresh never happened, and the leaked slides aren't showing a full release stack, just 2 discrete gpus. OTOH, if the top die gives 4070ti gaming performance - that covers 99% of the market.

The ARC division was never on the 'chopping block;' they're locked-in through Celestial. This was known from the beginning. MLiD's claims against this were baseless. There is no misunderstanding; it was slander.

The refreshed ARC Pros (A40, A50, A60) launched. Also, 'just two GPU skews' is a release. Hell, one GPU skew is a release.

SSGBryanHe's recommending them now, due to how much the drivers have improved.

I've just installed the latest xxx.5122 driver on my i7-11700, but Arc Control and GPU-Z still report it as xxx.5081. I even did a clean install with an installer downloaded from here on TPU, same results.

SSGBryanHe's recommending them now, due to how much the drivers have improved.

I've actually watched his videos on Arc - his videos were talking about how Intel was looking to axe the entire division. At the time, he stated that it was up in the air. People on the internet extrapolated what he actually said into something else.

I'd also point out that a lot of his coverage turned out to be accurate. The Alchemist refresh never happened, and the leaked slides aren't showing a full release stack, just 2 discrete gpus. OTOH, if the top die gives 4070ti gaming performance - that covers 99% of the market.

Watch his videos on nvidia 4000 series one day before official announcements.

It had absolutely nothing of value, with price and performance ranges in the +-100%. He vomited it out very confidently, though.

AusWolfI've just installed the latest xxx.5122 driver on my i7-11700, but Arc Control and GPU-Z still report it as xxx.5081. I even did a clean install with an installer downloaded from here on TPU, same results.

Yeah, that is correct and intended.

Currently there are two drivers packed together, should be temporary, then back to one unified package and version in the future.
Techpowerup maybe should include a note or something that is more obvious or easier for people to understand.

5081 for Arc discrete and Integrated Graphics in Processors like Core i3, Core i5, Core i7 and Core i9 (UHD, Iris, etc).
5122 for Core Ultra only.

From the release notes: "Intel® Graphics Driver package size has temporarily increased because of the inclusion of Intel® Core™ Ultra with Intel® Arc™ Graphics and Intel® Graphics driver."

There is a newer driver than TPU hosts, its on the Intel website, its a beta or non-WHQL though: 5084 Non-WHQL + 5122 WHQL... TPU has 5081 WHQL + 5122 WHQL
Whats the difference between 5081 and 5084 you ask? Absolutely no clue... Besides the number increasing by 3.

fiftofarYeah, that is correct and intended.

Currently there are two drivers packed together, should be temporary, then back to one unified package and version in the future.
Techpowerup maybe should include a note or something that is more obvious or easier for people to understand.

5081 for Arc discrete and Integrated Graphics in Processors like Core i3, Core i5, Core i7 and Core i9 (UHD, Iris, etc).
5122 for Core Ultra only.

From the release notes: "Intel® Graphics Driver package size has temporarily increased because of the inclusion of Intel® Core™ Ultra with Intel® Arc™ Graphics and Intel® Graphics driver."

There is a newer driver than TPU hosts, its on the Intel website, its a beta or non-WHQL though: 5084 Non-WHQL + 5122 WHQL... TPU has 5081 WHQL + 5122 WHQL
Whats the difference between 5081 and 5084 you ask? Absolutely no clue... Besides the number increasing by 3.

Wow, that is weird! I hope they'll change it back soon, as I don't want to download 1+ GB with driver updates on my crappy DSL connection.