![AMD Zen 5 Technical Deep Dive (1) AMD Zen 5 Technical Deep Dive (1)](https://i0.wp.com/www.techpowerup.com/review/amd-zen-5-technical-deep-dive/images/small@1x.png)
Ryzen AI 300 Series
AMD Ryzen AI 300 series processors will be the most consequential client segment processor for AMD in years. This is because it's the first mobile processor series with an NPU that meets Microsoft Copilot+ requirements, and spans the entire breadth of mobile platforms from thin-and-lights to mainstream notebooks, unlike the Intel Core Ultra 200V series Lunar Lake, which caters to a specific market segment—thin and lights and certain mainstream segments, but not value or enthusiast. Intel is banking on the mobile variants of its Core Ultra Arrow Lake processors for that role, but we're hearing that the first revision of Arrow Lake only has a 16 TOPS-class NPU and doesn't meet Copilot+ requirements—a big oversight on Intel's part—and the company plans to fix this with a future revision of Arrow Lake, as it requires a physical redesign of its SoC tile.
Such is AMD's footing with the Ryzen AI 300 series Strix Point, that the first notebooks powered by them should begin announcements this month. Intel created the NPU as a discrete component on the processor, with its Core Ultra Meteor Lake, AMD only played catch-up, and it now seems like AMD has a vital lead in go-to-market, if it can capitalize on it.
The Strix Point processor is a monolithic silicon, unlike Meteor Lake, Arrow Lake, and Lunar Lake. It is built entirely on the TSMC N4X node. It features a large CPU core complex (CCX) with two kinds of CPU cores—four Zen 5, and eight Zen 5c, which share a 24 MB L3 cache. The Zen 5 cores hold on to the highest boost frequencies. AMD has worked to improve the boost frequency capabilities of its compacted Zen 5c cores (when compared to Zen 4c), so these should boost better, although still not in the league of the Zen 5 cores. Both core types share an identical IPC and ISA, and support for SMT, so this really is a 12-core/24-thread processor, with basic UEFI CPPC preferred core flagging getting the OS to prioritize work to the Zen 5 cores.
Ryzen AI 300 Performance
Here are AMD's first-party performance claims for the Ryzen AI HX 370 processor compared to the Core Ultra 9 185H and Snapdragon X Elite.
And here's how it fares in gaming workloads. These numbers compare the iGPUs of the three processors, and we see AMD now has a definite iGPU gaming performance leadership.
AMD RDNA 3.5
The maxed out Ryzen AI HX 370 processor comes with the AMD Radeon 890M graphics model, the maxed out variant of the iGPU on Strix Point that's powered by the new RDNA 3.5 graphics architecture.
AMD is choosing to call this RDNA 3.5 as opposed to something like "RDNA 4" because the physical SIMD machinery is largely unchanged from RDNA 3, but AMD has made several updates up and down the graphics pipeline to improve performance/Watt, which it can trade in for pure performance.
AMD has doubled the number of TMUs in the design. This has a nearly universal impact on game performance, because textures drive today's games. The company has also doubled the rate of interpolation and compression ops. Lastly, AMD has worked on the memory management of the iGPU to make the most of available memory bandwidth, particularly in bandwidth-starved scenarios such as when paired with LPDDR5. This is using a combination of batch processing, reduction of memory roundtrips, and better compression techniques.
Pay attention to this graph—AMD claims performance gains of +32% in Time Spy, and +19% in Night Raid, when compared to the Radeon 780M iGPU of the Ryzen 8040 Hawk Point processor. It's important to note here that AMD has increased the compute units on Strix Point to 16, from just 12 on Hawk Point—a 33% numerical increase, which reflects in a linear increase in Time Spy performance. However, AMD is able to do this at the same 15 W power footprint. So the linear performance gain isn't coming at a proportionate power increase. That's where RDNA 3.5 is working its magic.
').appendTo('body');$(window).scroll(nextPageScrollHandler);}if (evt.touches.length==0)$('#keep-scrolling-buffer').removeClass('show');else {if ($('#keep-scrolling').hasClass('show'))$('#keep-scrolling-buffer').addClass('show');}});});
');$('.tpu-fancybox-wrap').css('maxWidth', maxWidth);*/instance.$refs.stage.on('transitionend', function() {updateButtonPos(instance);});},onUpdate: updateButtonPos,afterShow: function(instance, slide) {updateButtonPos(instance);instance.$refs.inner.find('.fancybox-tpu-nav').show();},beforeClose: function(instance, slide) {instance.$refs.inner.find('.fancybox-tpu-nav').hide();},afterClose: function(instance, slide) {$('.tpu-fancybox-wrap').contents().unwrap();$('body').removeClass('tpu-fancybox-body-wrap')},baseTpl: '
',});});}loadjs.ready(['jquery', 'fancybox'], function() {attachLightbox(function() {return $('.text a:has(>img)').filter(function(i, el) {return $(this).attr('href').match(/\.(jpg|gif|png)(?=$|[?#])/) != null;});});});loadjs.ready(['jquery'], function() {$('.review').on('click', '.spoiler > a', function(e) {e.preventDefault();$(this).next('div').toggle();});$('.review').on('click', '.ispoiler', function(e) {e.preventDefault();$(this).find('div').css('filter', '');$(this).removeClass('ispoiler');});});loadjs.ready(['jquery'], function() {$('table.tputbl.sort-header').find('th[scope=col]').css('cursor', 'pointer').on('click', function(e) {var header=$(this)header.parents('table').find('th[scope=col]').removeClass('sort-active');header.addClass('sort-active');var idx=header.parents('tr').find('th[scope=col]').index(this);header.parents('table').find('td, th[scope=row]').filter(function() {return $(this).index() == idx;}).sortElementsTPU(function(a, b) {a=$(a).text(); b=$(b).text();if (header.hasClass('sort-force-number')) {a=parseInt(a, 10);b=parseInt(b, 10);}if (header.hasClass('sort-nan-is-zero')) {if (isNaN(a))a=0;if (isNaN(b))b=0;}try {var result=a.localeCompare(b, undefined, { numeric: true, sensitivity: 'base' });return result;} catch(e) {if (a>b) return 1;if (a